File Structures

An Object-Oriented
Approach with C++
M
CO
P.
O
LO

Michael J. Folk
TU

University ofIllinois
.V
W

Bill Zoellick
W

CAP Ventures
W

Greg Riccardi
Florida State University

A ADDISON-WESLEY
yy
Addison-Wesley is an imprint of Addison Wesley Longman,Inc..

Reading, Massachusetts * Harlow, England Menlo Park, California

Berkeley, California * Don Mills, Ontario + Sydney
Bonn + Amsterdam « Tokyo * Mexico City

https://hemanthrajhemu.github.io
 Contents xix

9.11 Worst-Case Search Depth 401

9.12 Deletion, Merging, and Redistribution 403
9.12.1 Redistribution 406
9.13 Redistribution During Insertion: A Way to Improve Storage
Utilization 407
9.14 B* Trees 408
9.15 Buffering of Pages: Virtual B-Trees 409
9.15.1 LRU Replacement 410
9.15.2 Replacement Based on Page Height 411
9.15.3 Importance of Virtual B-Trees 412
9.16 Variable-Length Records and Keys 413
Summary 414 KeyTerms 416 FurtherReadings 417 Exercises 419
Programming Exercises 421 Programming Project 422

Chapter 10 Indexed Sequential File Access and Prefix B* Trees 423

M
CO
10.1 Indexed Sequential Access 424
10.2 Maintaining a Sequence Set 425
P.

10.2.1 The Use of Blocks 425

O

10.2.2 Thoice of Block Size 428

LO

10.3 Adding a Simple Index to the Sequence Set 430

10.4 The Content of the Index;Separators Instead of Keys 432
TU

10.5 The Simple Prefix B+ Tree 434 .

.V

10.6 Simple Prefix B+ Tree Maintenance 435.

W

10.6.1 Changes Localized to Single Blocks in the Sequence Set 435

10.6.2 Changes Involving Multiple Blocks in the Sequence Set 436
W

10.7 Index Set Block Size 439

W

10.8 Internal Structure of Index Set Blocks: A Variable-Order

B-Tree 440
10.9 Loading a Simple Prefix Bt Tree 443
10.10 BtTrees 447
10.11 B-Trees, B+ Trees, andSimple Prefix B+ Trees in Perspective 449
Summary 452 KeyTerms 455 FurtherReadings 456 Exercises 457
Programming Exercises 460 Programming Project 461

Chapter 11 Hashing «463

11.1 Introduction 464

11.1.1 What Is Hashing? 465
11.1.2 Collisions 466
11.2 A Simple Hashing Algorithm 468

https://hemanthrajhemu.github.io
 Contents

11.3 Hashing Functions and Record Distributions 472

11.3.1 Distributing Records among Addresses 472
11.3.2 Some Other Hashing Methods 473
11.3.3 Predicting the Distribution of Records 475
11.3.4 Predicting Collisions for a Full File 479
11.4 How Much Extra Memory Should Be Used? 480
11.4.1 Packing Density 481
11.4.2 Predicting Collisions for Different Packing Densities 481
11.5 Collision Resolution by Progressive Overflow 485
11.5.1 HowProgressive Overflow Works 485
11.5.2 Search Length 487
11.6 Storing More Than One Record per Address:Buckets 490
11.6.1 Effects of Buckets on Performance 491
11.6.2 Implementation Issues 496
11.7 Making Deletions 498
11.7.1 Tombstones for Handling Deletions 499

M
11.7.2 Implications of Tombstones for Insertions 500
CO
11.7.3 Effects of Deletions and Additions on Performance 501
11.8 Other Collision Resolution Techniques 502
P.

11.8.1 Double Hashing 502

O

11.8.2 Chained Progressive Overflow 502

LO

11.8.3 Chaining with a Separate Overflow Area 505

TU

11.8.4 Scatter Tables: Indexing Revisited 506

11.9 Patterns of Record Access 507
.V

Summary 508 KeyTerms 512 FurtherReadings 514 Exercises 515

W

Programming Exercises 520 .

W
W

Chapter 12. Extendible Hashing 523

12,1 Introduction 524

12.2. How Extendible Hashing Works 525
12.2.1 Tries 525
12.2.2 Turning the Trie into a Directory 526
12.2.3 Splitting to Handle Overflow 528
12.3. Implementation 530 .
12.3.1 Creating the Addresses 530
12.3.2 Classes for Representing Bucket and Directory Objects 533
12.3.3 Bucket and Directory Operations 536
12.3.4|mplementation Summary 542
12.4 Deletion 543°
12.4.1 Overview.of the Deletion Process - 543

https://hemanthrajhemu.github.io
 CHAPTER

M
CO
P.
O
LO
TU

CHAPTER OBJECTIVES
.V
W

Introduce the conceptof hashing.

~~ >,

Examine the problem of choosing a good hashing algorithm,

4.

present a reasonable one in detail, and describe some others.

Explore three approaches for reducing collisions: randomization of
addresses, use of extra memory, and storage of several records per
adaress.

Develop and use mathematical tools for analyzing performance

differences resulting from the use of different hashing techniques.
Examine problems associated with file deterioration and discuss
some solutions.
Examine effects of patterns of record access on performance.

463
https://hemanthrajhemu.github.io
464 Chapter 11 Hashing

CHAPTER OUTLINE
11.1 Introduction
11.1.1 What Is Hashing?
11.1.2 Collisions
11.2 A Simple Hashing Algorithm
11.3 Hashing Functions and Record Distributions
11.3.1 Distributing Records among Addresses
11.3.2 Some Other Hashing Methods
11.3.3 Predicting the Distribution of Records
11.3.4 Predicting Collisions for a Full File
11.4 How Much Extra Memory Should Be Used?
11.4.1 Packing Density
11.4.2 Predicting Collisions for Different Packing Densities
11.5 Collision Resolution by Progressive Overflow
11.5.1 How Progressive Overflow Works

M
11.5.2 Search Length CO
11.6 Storing More Than One Record per Address: Buckets
11.6.1 Effects of Buckets on Performance
P.

11.6.2 Implementation Issues ,

O

11.7 Making Deletions

LO

11.7.1 Tombstones for Handling Deletions

TU

17.7.2 mplications of Tombstonesfor Insertions

11.7.3 Effects of Deletions and Additions on Performance
.V

11.8 Other Collision Resolution Techniques

W

11.8.1 Double Hashing

W

11.8.2 Chained Progressive Overflow

W

11.8.3 Chaining with a Separate Overflow Area

11.8.4 Scatter Tables: Indexing Revisited
11.9 Patterns of Record Access

11.1 ‘Introduction

O(1) access to files means that no matter how bigthefile grows, access to.
a record always takes the same, small number of seeks. By contrast,
sequential searching gives us O(N) access, wherein the numberofseeks
grows in proportion to the size of the file. As we saw in the preceding
chapters, B-trees improve on this greatly, providing O(log, N) access; the
numberof seeks increases as the logarithm to the base k of the numberof
records, wherek is a measureofthe leaf size. O(log, N) access can provide

https://hemanthrajhemu.github.io
Introduction 465

verygoodretrieval performance,even for very largefiles, but it is still not

O(1) access.
In a sense, O(1) access has been the Holy Grail offile structure design.
Everyone agrees that O(1) access is what we want to achieve, but until
about ten years ago,it was not clear if one could develop a generalclass of
O(1) access strategies that would work on dynamicfiles that change great-
ly 1n size.
In this chapter we begin with a descriptionof static hashing tech-
niques. They provide us with O(1) access but are not extensibleas the file
increases in size. Static hashing was the'state of the art until about 1980. In
the following chapterwe show howresearch anddesign work during the
1980s found ways to extend hashing, and O01 ) access, to files that are
dynamic and increase in size over time.

11.1.1 What Is Hashing?

M
CO
A hash functionis like a black box that produces an addressevery time you
drop in a key. More formally,it is a function h(K) that transforms a key K
P.

into an address. The resulting addressis used as the basis for storing and
O
LO

retrieving records. In Fig. 11.1, the key LOWELLis transformed by the

hash function to the address 4. That is, k}( LOWELL) = 4. Address is said
TU

to be the home address of LOWELL.

.V

Hashing is like indexingin that it involves associating a keywith a

W

relative record address. Hashing |differs from indexingin two important

W

ways:
W

m With hashing,the addresses generated appear to be random—thereis

noimmediately obvious connectionbetween the key andthelocation
of the corresponding record, even though the key is used to determine
the location of the record. For this reason, hashing is sometimes
referred to as randomizing.
m With hashing, two different keys may be transformed to the same
address so two records maybe sent to thesameplacein thefile. When
this occurs, it is called a collision and some means must be found to
deal withit.
Consider the following simple example. Suppose you wantto store
seventy-five records in a file in which the key to each record is a person’s
name. Supposealso that you set aside space for one thousand records. The
key can be hashed by taking two numbers from the ASCII representations
of the first two characters of the name, multiplying these together, then

https://hemanthrajhemu.github.io
466 Chapter 11 Hashing

Address Record

2
—Key 3
Ke= Lowst1|
wu 4 | LOWELL... ++#—-LOWELL’s
ueTe home address
: A(K)” Address
5
4
6

M
Figure 11.1 Hashing the key LOWELL to address 4.
CO
P.

using the rightmost three digits of the result for the address. Table 11.1
O
LO

shows how three names would produce three addresses. Note that even
though the namesare listed in alphabetical order, there is no apparent
TU

order to the addresses. They appear to be in randomorder. |

.V
W

11.1.2 Collisions
W
W

Now suppose there is a key in the sample file with the name OLIVIER.
Because the nameOLIVIERstarts with the same twoletters as the name
LOWELL, they produce the same address (004). There is a collision
between the record for OLIVIER and the record for LOWELL. Wereferto
keys that hash to the same address as synonyms.
Collisions cause problems. We cannot put two records in the same
space, so we mustresolve collisions. We do this in two ways: by choosing
hashing algorithms partly on the basis of how fewcollisions they are like-
ly to produce and by playing sometricks with the way westore records.
The ideal solution to collisionsis to find a transformation algorithm
that avoidscollisions altogether. Such an algorithm is called a perfect hash-
ing algorithm,It turns out to be much moredifficult to find a perfect hash-
ing algorithm than one might expect. Suppose, for example, that you want
to store 4000 records among 5000 available addresses. It can be shown
(Hanson,1982) that of the huge numberof possible hashing algorithms

https://hemanthrajhemu.github.io
Introduction | - ' 467

Table 11.1 A simple hashing scheme

ASCII code
for first two Home
Name letters Product address

BALL 66 65 66 X 65 = 4290 290

LOWELL 76 79 76X79 = 6004 004
TREE 84 82 84 x 82 = 6888 88&§

for doing this, only one outof 10120000 avoidscollisions altogether. Hence,
it is usually not worth trying.!
A more practical solution is to reduce the numberofcollisions to an
acceptable number. For example, if only one out of ten searches for a

M
record results in a collision, then the average number of disk accesses
CO
required to retrieve a record remains quite low. Thereare several different
ways to reduce the numberofcollisions, including the following:
P.
O

m Spread out the records. Collisions occur when two or more records
LO

' ‘compete for the same address. If we could find a hashing algorithm
that distributes the records fairly randomly among theavailable
TU

addresses, then we would not have large numbersofrecords clustering

.V

around certain addresses. Our sample hash algorithm, which uses only
W

two letters from the key, is not good on this account because certain
W

combinations of two letters are quite common in starting names,

W

while others are uncommon(e.g., compare the numberof namesthat

start with “JO” with the numberthatstart with “XZ”). We needto find
a hashing algorithm that distributes records more randomly. |
m Use extra memory. It is easier to find a hash algorithm that avoidscolli-
sions if we have only a few recordsto distribute among manyaddress-
es than if we have about the same numberof records as addresses. Our
sample hashing algorithm is very good on this accountsince there are
one thousand possible addresses, and only seventy-five addresses
(corresponding to the seventy-five records) will be generated. The

1, It is not unreasonable to.try to generate perfect hashing functionsfor small (less than 500), stable
sets of keys, such as mightbe used to look up reserved words in a programming language. Butfiles
generally contain more than a few hundredkeys, or they contain sets of keys that change frequent-
ly, so they are not normally considered candidates for perfect hashing functions. See Knuth (1998),
Sager (1985), Chang (1984), and Chichelli (198% for more on perfect hashing functions.

https://hemanthrajhemu.github.io
468 Chapter 11 Hashing

obvious disadvantage to spreading out the recordsis that storage space

is wasted. (In the example, 7.5 percent of the available record space is
used, and the remaining 92.5 percent is wasted.) There is no simple
answer to the question of how much empty space shouldbe tolerated
to get the best hashing performance, but some techniquesare provid-
ed later in this chapter for measuring the relative gains in performance
for different amounts offree space.
m Put more than one record at a single address. Up to now we have
assumedtacitly that each physical recordlocation in a file could hold
exactly one record, but there is usually no reason we cannotcreate our:
file in such a way that every file address is big enough to hold several
records. If, for example, each record is 80 bytes long and wecreate a
file with 512-byte physical records, we can store up to six records at
each file address. Each address is able to tolerate five synonyms.
Addresses that can hold several records in this way are sometimes

M
called buckets. CO
In the following sections we elaborate on these collision-reducing
P.

methods, and as we do so we present some programsfor managing h| ashed

O

files. '
LO
TU

A Simple Hashing Algorithm

.V

11.2
W

as
W

One goal in choosing any hashing algorithm should be to spread out

W

records as uniformly as possible over the range of addressesavailable. The

use of the term hash for this technique suggests what is done to achieve
this. Our dictionary reminds us that the verb to hash means “to chop into
small pieces ... muddle or confuse.” The algorithm usedpreviously chops
off the first two letters and then uses the resulting ASCII codes to produce
a number that is in turn chopped to produce the address. It is not very
good at avoiding clusters of synonymsbecause so many names begin with
the same twoletters. :
One problem with the algorithm is that it does not do very much
hashing. It uses only twoletters of the key and doeslittle with those two.
letters. Nowlet us look at a hash function that does much more random-
izing, primarily becauseit uses moreofthe key.It is a reasonably good
basic algorithm and is likely to give good results no matter what kindsof
keys are used.It is also an algorithm thatis not too difficultto alter in case--
‘aspecificinstanceofthealgorithm does not workwell.

https://hemanthrajhemu.github.io
A Simple Hashing Algorithm 469

This algorithm hasthree steps:

. Represent the key in numerical form.
2. Fold and add.
3. - Divide by a prime number and use the remainderas the address.

Step 1. Represent the Key in Numerical Form

If the key is already a number, then this step is already accomplished.If it
is a string of characters, we take the ASCII code of each character and use
it to form a number. For example,
76 79 87 69 76 76 32 32 32 32 32 32
LOWELL =
L O W E L LBL fle Blanks = |

In this algorithm we use the entire key rather than just the first two

M
letters. By using moreparts of a key, we increase the likelihood that differ-
CO
ences among the keys cause differences in addresses produced. The extra
processing time required to do this is usually insignificant when compared
P.
O

with the potential improvement in performance.

LO
TU

Step 2. Fold and Add

.V

Folding and adding means choppingoff pieces of the number and adding
W

them together. In our algorithm we chop off pieces with two ASCII
W

numbers each:
W

76 79 | 87 69 | 76 76 | 32 32 | 32 321 32 32
These numberpairs can be thought of as integer variables (rather than
character variables, which is how they started out) so we can do arithmetic
on them. If we can treat them as integer variables, then we can add them.
This is easy to do in C because C allows us to do.arithmetic on characters.
In Pascal, we can use the ord() function to obtain the integer position of a
character within the computer’s characterset.
Before we add the numbers, we have to mention a problem caused by
the fact that in most cases the sizes of numbers we can add together are
limited. On some microcomputers, for example, integer values that exceed
32 767 (15 bits) cause overflow errors or become negative. For example,
adding thefirst five of the foregoing numbers gives
7679 + 8769 + 7676 + 3232 + 3232 = 30 588

https://hemanthrajhemu.github.io
470 Chapter 11 Hashing

Addingin the last 3232 would, unfortunately, push the result over the
maximum 32 767 (30 588 + 3232 = 33 820), causing an overflow error.
Consequently, we need to make sure that each successive sum is less than
32 767. We can dothis byfirst identifying the largest single value wewill
ever add in our summation and then making sure after each step that our
‘intermediate result differs from 32 767 by that amount.
In ourcase, let us assume thatkeys consist only of blanks and upper-
case alphabetic characters, so the largest addendis 9090, corresponding to
ZZ. Suppose we choose 19 937 as ourlargest allowable intermediate result.
This differs from 32 767 by much more than 9090, so we can be confident
(in this example) that no new addition will cause overflow. We can ensure
in our algorithm that no intermediate sum exceeds 19 937 by using the
mod operator, which returns the remainder when one integeris divided by
another:
7679 + 8769 — 16 448 — 16448 mod 19937 — 16448

M
16448 +7676 —+24124 — 24 124 mod 19 937
CO — 4187
4187 + 3232 — 7419 — mod 19 937 — 7419
7419 + 3232 > 10651 — mod 19 937 > 10651
P.
O

10651 +3232 —13 883 ~» 13883 mod 19937, 4 13 883

LO

The number 13 883 is the result of the fold-and-add operation.

TU

Whydid we use 19 937 as our upper bound rather than, say, 20 000?
Because the division and subtraction operations associated with the mod
.V

operator are more than just a way of keeping the number small; theyare
W

part of the transformation work of the hash function. As we see in the

W

discussion for the next step, division by a prime numberusually produces

W

a more randomdistribution than doestransformation by a nonprime. The

number 19 937 is prime. |

Step 3. Divide by the Size of the Address Space

The purpose of this step is to cut down to size the numberproduced in

step 2 so it falls within therange of addressesof recordsin the file. This can
be done by dividing that number by a number-thatis the address size of
the file, then taking the remainder. The remainder will be the home
address of the record.
Wecan represent this operation symbolically as follows:if s represents
the sum producedin step 2 (13 883:in the example), n represents the divi-.
sor (the number of addresses in the file), and a represents, the address we
are trying to produce, we apply the formula
a=smodn

https://hemanthrajhemu.github.io
A Simple Hashing Algorithm 471

The remainder produced by the mod operator will be a number

between Oandn-1.
"Suppose,for example, that we decide to use the 100 addresses 0--99 for
ourfile. In terms of the preceding formula,
= 13 883 mod 100
= 83
Since the numberofaddresses allocated for thefile does nothave to be
any specific size (as long as it is big enough to hold all of the records to be
stored in the file), we have a great deal of freedom in choosing the divisor
n, It is a good thing that we do, because the choice of n can have a major
effect on how well the records are spread out.
A prime numberis usually used for the divisor because primes tend to
distribute remainders much more uniformly than do nonprimes. A
nonprime can work well in many cases, however, especially if it has no
primedivisors less than 20 (Hanson, 1982). Since the remainderis going

M
to be the address of a record, we choose a numberasclose as possible to
CO
the desired size of the address space.This number determinesthe sizeof
P.

the address space. For a file with 75 records, a good choice might be 101,
O

which wouldleave thefile 74.3 percent full (74/101 = 0.743).

LO

If 101 is the sizé of the address space, the home address of the record
in the example becomes
TU

a= 13 883 mod 101

.V

=.46
W

Hence, the record whose key is LOWELLisassigned to record number 46

W

in thefile..
W

This procedure can be carried out with the function Hash in Fig. 11.2.
Function Hash takes two inputs: key, which must be an array of ASCII
codes for at least twelve characters, and maxAddress, which has the maxi-
mum address value. The value returned by Hashis the address.

int Hash (char key(12], int maxAddress)

{
int sum = 0;
for -(int j = 0; J < 12; j += 2)
sum = (sum * 100 * key(3] * key[j+1]) % 19937;
return sum % maxAddress;
)

Figure 11.2 Function Hash uses folding and prime number division to
compute a hash addressfor a twelve-character string,

https://hemanthrajhemu.github.io
472 Chapter 11 Hashing

11.3. Hashing Functions and Record Distributions

Of the two hash functions we have so far examined, one spreads out
records pretty well, and one does not spread them out well at all. In this
section we look at ways to describe distributions of records in files.
Understanding distributions makes it easier to discuss other hashing
methods.

11.3.1 Distributing Records among Addresses

Figure 11.3 illustrates three different distributions of seven records among
ten addresses. Ideally, a hash function should distribute records in file so
there are no collisions, as illustrated by distribution (a). Sucha distribu-
tion is called uniform because the records are spread out uniformly among

M
the addresses. We pointed out earlier that completely uniform distribu-
CO
tions are so hardto findit is generally not considered worth tryingto find
them.
P.

Distribution (b) illustrates the worstpossible kind of distribution. All

O
LO

records share the same homeaddress,resulting in the maximum number

of collisions. The more a distribution lookslike this one, the more that
TU

collisions will be a problem.

.V

Distribution (c) illustrates a distribution in which the records are

W

somewhatspread out, but with a few collisions. This is the mostlikely case
W

if we have a function that distributes keys randomly. If a hash functionis

W

random, then for a given key every address has the samelikelihood of.
being chosen as every other address. The fact that acertain addressis
chosen for one key neither diminishes nor increasesthelikelihoodthat the
same address will be chosen for anotherkey.
It should be clear that if a random hash functionis used to generate a
large numberof addresses from a large numberof keys, thensimply by
chance someaddresses are going to be generated moreoften than others.If
you have, for example, a random hash function that generates addresses
between 0 and 99 and you give the function one hundred keys, you would
expect some of the one hundred addresses to be chosen more than once '
and some to be chosen notatall. |
Although a random distribution of records amongavailable addresses
is not ideal, it is an acceptable alternative given thatit is practically impos-
sible to find a function that allows a uniform distribution. Uniform distri-

https://hemanthrajhemu.github.io
Hashing Functions and Record Distributions 473

Best Worst Acceptable

Record Address _ Record Address Record Address
1 1 1

A 2 A 3 A 2
3 3 3
B B B
4 4 4
0
D
5 °
D
5 C
D
5
6 6 6
E E E
7 7 7
F 8 r 8 8
G 9 G 9 C 9
10 10 10

(a) (b) (c)

Figure 11.3 Different distributions. (a) No synonyms(uniform). {b) All

synonyms (worst case).(c) A few synonyms.

M
CO
P.

butions may be out of the question, but there are times when we can_find
O

distributions that are better than random in the sense that, while they do
LO

generate a fair number of synonyms, they spread out records among

TU

addresses more uniformly than does a random distribution.

.V
W

11.3.2 Some Other Hashing Methods

W

It would be nice if there were a hash function that guaranteed a better-

W

than-randomdistribution in all cases, but there is not. The distribution

generated by a hashing function dependsontheset of keys that are actu-
ally hashed. Therefore, the choice of a proper hashing function should
involve some intelligent consideration of the keys to be hashed, and
perhaps some experimentation. The approaches to choosing a reasonable
hashing function covered in this section are ones that have been found to
work well, given the right circumstances. Further details on these and
other methods can be found in Knuth (1998), Maurer (1975), Hanson
(1982), and Sorensonet al. (1978).
Here are some methodsthat are potentially better than random:
m Examine keys for a pattern. Sometimes keysfall in patterns that natu-
rally spread themselves out. This is morelikely to be true of numer-
ic keys than of alphabetic keys. For example, a set of employee

https://hemanthrajhemu.github.io
474 Chapter 17 Hashing

identification numbers might be ordered according to when the

employees entered an organization. This might even lead to no
synonyms. If some part of a key shows a usable underlying pattern, a
hash function that extracts that part of the key can also be used.
mH Foldparts ofthe key. Foldingis one stage in the method discussedearli-
er, It involves extracting digits from part of a key and adding the
extracted parts together. This method destroys the original key
patterns but in some circumstances may preserve the separation
between certain subsets of keys that naturally spread themselves out.
= Divide the key by a number. Division by the addresssize and use of the
remainderusually is involved somewhere in a hash function since the
purpose of the function is to produce an address within a certain
range. Division preserves consecutive key sequences, so you can take
advantage of sequencesthat effectively spread out keys. However,if
there are several consecutive key sequences, division by a numberthat

M
has many small factors can result in manycollisions. Research has
CO
shown that numbers with no divisors less than 19 generally avoid this
P.

problem. Division by a prime is even morelikely than division by a

O

nonprime to generate different results from different consecutive

LO

sequences.
TU

The preceding methods are designed to take advantage of natural

.V

orderings amongthe keys. The next two methods should be tried when,
W

for some reason,the better-than-random methodsdo not work.In these

W

cases, randomizationis the goal.

W

m Square the key and take the middle. This popular method (often
called the mid-square method) involves treating the key as a single
large number, squaring the number, and’ extracting whatever
numberof digits is needed from the middle ofthe result. For exam-
ple, suppose you wantto generate addresses between 0 and 99. If the
key is the number 453, its square is 205 209. Extracting the middle
two digits yields a number between 0 and 99, in this case 52. As long
as the keys do not contain many leading ortrailing zeros, this
method usually produces fairly random results. One unattractive
feature of this method is thatit often requires multiple precision
arithmetic.
m Radix transformation. This method involves converting the key to
some numberbase other than the one you are workingin, then taking
the result modulo the maximum address as the hash address. For

https://hemanthrajhemu.github.io
 Hashing Functions and Record Distributions 475

example, suppose you wantto generate addresses between 0 and 99.If

_the key is the decimal number 453, its base 11 equivalentis 382; 382
mod 99 = 85, so 85 is the hash address.
Radix transformation is generally more reliable than the mid-square
method for approachingtrue randomization, though mid-square has been
found to give good results when applied to somesets of keys.

11.3.3 Predicting the Distribution of Records

Given that it is nearly impossible to achieve a uniform distribution of
records amongthe available addressesin file,it is important to be able to
predict how recordsarelikely to be distributed. If we know, for example,
that a large number of addresses are likely to have far more records
assigned to them than they can hold, then we knowthat there are going to
be a lot of collisions.

M
Althoughthere are no nice mathematica]tools available for predicting
CO
collisions among distributions that are better than random, there are
P.

mathematical tools for understanding just this kind of behavior when

O

records are distributed randomly. If we assume a random distribution

LO

(knowing that verylikely it will be better than random), we can use these
TU

tools to obtain conservative estimates of how our hashing method is like-

ly to behave.
.V
W

The Poisson Distribution2

W
W

We want to predict the numberofcollisions that are likely to occur in file

that can hold only one record at an address. We begin by concentrating on
what happensto a single given address when a hash functionis applied to
a key. We would like to answer the following questions: Whenall of the
keys in a file are hashed, whatis thelikelihood that
m= Nonewill hash to the given address?
m Exactly one key will hash to the address?
m Exactly two keys will hash to the address (two synonyms)?

2. This section develops a formula for predicting the ways in which recordswill be distributed among
addresses in file if a random hashing function is used. The discussion assumes knowledge of
some elementary concepts of probability and combinatorics. You may want to skip the develop-
ment and go straight to the formula, whichis introduced inthe next section.

https://hemanthrajhemu.github.io
476 Chapter 71 Hashing

m Exactly three, four, and so on keys will hash to the address?

m All keys in the file will hash to the same givenaddress?
Which of these outcomes would you expect to befairly likely, and
which quite unlikely? Suppose there are N addressesiri a file. When a
single key is hashed, there are two possible outcomes with respect to the
given address:
A—The address is not chosen; or
B—The address is chosen.
How do we express the probabilities of the two outcomes? If welet
both p(A) and a stand for the probability that the addressis not chosen,
and p(B) and b standfor the probability that the address is chosen, then

P
(B)=b=—
1
N

M
since the address has one chancein N of being chosen, and
CO
N-1 1
P.

(A)=a= =|- —
O

P N N
LO

since the address has N— chances in N of not being chosen. If there are
TU

10 addresses (N= 10), the probability of our address being chosen is b=

1/10 = 0.1, and the probability of the address not being chosen is a = 1 —
.V

0.1 = 0.9.
W

Now suppose two keys are hashed. Whatis the probability that both
W

keys hash to our given address? Since the two applications of the hashing
W

function are independent of one another, the probability that both will
produce the given addressis a product: ;

p(BB)=bxb=tx- forn=10:b b=01x01=0.01

NN
Of course, other outcomes are possible when two keys are hashed. For
example, the second key could hash to an address other than the given
address. The probability ofthis is the product
]
p(BA) = bx aa ix[i-e for N= 10: bx a=0.1X0.9 =0.09
N N
In general, when we want to know the probability of a certain
sequence of outcomes, such as BABBA, we can replace each A and B by a
_and 8, respectively, and computethe indicated product:

https://hemanthrajhemu.github.io
Hashing Functions and Record Distributions A77

p(BABBA) = bx ax bx bx a= a2b3 = (0.9)? (0.1)3.

This example shows howto find the probability of three Bs and two
As, where the Bs and As occur in. the order shown. We want to know the
probability that there are a certain number of Bs and As, but without
regard to order. For example, suppose we are hashing four keys and we
want to know howlikely it is that exactly two of the keys hash to our
given address. This can occurin six ways, all six ways having the same
probability:
Outcome Probability ‘For N= 10

BBAA bbaa = ba? (0.1)2(0.9)2 = 0.0036

BABA baba = ba? (0.1)2(0.9)2 = 0.0036
BAAB baal = b2a2 (0.1)2(0.9)? = 0.0036
ABBA abba = ba? (0.1)2(0.9)2 = 0.0036

M
ABAB abab = b2a2 (0.1)2(0.9)2 = 0.0036
CO
AABB aabb = b?a2 (0.1)2(0.9)2 = 0.0036
P.
O

Since these six sequences are independentof one another,the proba-

LO

bility of two Bs and two Asis the sum ofthe probabilities of the individual
TU

outcomes:
.V

p(BBAA) + p(BABA) +... + p(AABB) = 6b2a2 = 6 x 0.0036 = 0.0216.

W

The 6 in the expression 6b7arepresents the number of ways two Bs and

W

two As can be distributed among fourplaces.

W

In general, the event “r trials result in r~x As and x Bs” can happen in
as many ways as r~— x letters A can be distributed among places. The
probability of each such way is
grex px

and the numberof such waysis given by the formula

_ r!
(r—x )! xl

This is the well-known formula for the numberof ways of selecting x

items out ofa set of r items. It follows that when r keys are hashed, the
probability that an address will be chosen x times and not chosen r—x
times can be expressed as
p(x) = Carxbx

https://hemanthrajhemu.github.io
478 Chapter 11 Hashing

Furthermore, if we know that there are N addresses available, we can be

precise about the individual probabilities of A and B, and the formula
becomes

=C
Ly*(1
] —-— —_
Pe) | y A
where Chasthe definition given previously.
What doesthis mean? It means that if, for example, x = 0, we can
computethe probability that a given address will have 0 recordsassigned
to it by the hashing function using the formula

plo) =C
ow)bs
If x= 1, this formula gives the probability that one record will be-assigned

omfl
to a given address:

M
l
COr-l l 1

p N N
P.
O

This expression has the disadvantage that it is awkward to compute.

LO

(Try it for 1000 addresses and 1000 records: N = r = 1000.) Fortunately,

for large values of Nand r, there is a function that is a very good approx-
TU

imation for p(x) and is much easier to compute.It is called the Poisson
.V

function.
W
W

The Poisson Function Applied to Hashing

W

The Poisson function, which wealso denote by p(x), is given by

(r/N}* e -/N)
P(x) =
x!

where N, 1, x, and p(x) have exactly the same meaning they havein the
previoussection. Thatis,if
N =the numberofavailable addresses;
r =the numberof recordsto be stored; and
x =the numberofrecords assigned to a given address,
then p(x) gives the probability that.a given address will have had x records
assignedto it after the hashing function has been appliedto all n records.
Suppose, for example, that there are 1000 addresses (N = 1000) and
1000 records whosekeys are to be hashed to the addresses (r =1000). Since

https://hemanthrajhemu.github.io
Hashing Functions and Record Distributions 479

r/N = 1, the probability that a given address will have no keys hashedtoit
(x= 0) becomes
10 e-]
p(0) = = 0.368
0!
The probabilities that a given address will have exactly one, two, or
three keys, respectively, hashed toit are

pl)= = 0:368

p(2) _
= Pet
a _
= 0.184

eet
p(3) = sr = 0.061
_ If we can use the Poisson function to estimate the probability that a

M
given address will have a certain number of records, we can also use it to
CO
predict the numberof addresses that will have a certain numberof records
assigned.
P.
O

For example, suppose there are 1000 addresses (N = 1000) and 1000
LO

records (r= 1000). Multiplying 1000 by the probability that a given address
will have x records assigned to it gives the expected total number of
TU

addresses with x records assigned to them. That is, 1000p(x} gives the
.V

number of addresses with xrecordsassigned to them.

W

In general, if there are N addresses, then the expected numberof

W

addresses with x records assigned to them is

W

Np(x)
This suggests another wayof thinking about p(x). Rather than think-
ing about p(x) as a measure of probability, we can think of p(x) as giving
the proportion of addresses having x logical records assigned by hashing. |
Now that we havea tool for predicting the expected proportion of
addresses that will have zero, one, two,etc. records assigned to them bya
random hashing function, wecan applythis tool to predicting numbersof
collisions.

11.3.4 Predicting Collisionsfora Full File

Suppose you have a hashingfunction that you believe will distribute records
randomly and you wantto store 10 090 records in 10 000 addresses. How
manyaddresses do you expect to have no records assigned to them?

https://hemanthrajhemu.github.io
480 Chapter 11 Hashing

Since r= 10 000 and N= 10 000, r/N = 1. Hence the proportion of

addresses with 0 records assigned should be
19 g-l
p(0) = = 0.3679

The number of addresses with no records assignedis

10 000 x p(0) = 3679
How many addresses should have one, two, and three records
assigned, respectively?
10 000 x p(1) = 0.3679 x 10 000 = 3679
10 000 x p(2) = 0.1839 x 10 000 = 1839
10 000 x p(3) = 0.0613 x 10 000 = 613
Since the 3679 addresses corresponding to x = 1 have exactly one
record assigned to them, their records have no synonyms. The 1839

M
addresses with two records apiece, however, represent potential trouble.If
CO
each such address has space only for one record, and two records are
P.

assigned to them,thereis a collision. This means that 1839 recordswillfit

O

into the addresses, but another 1839 will not fit. There will be 1839 over-
LO

flow records.
TU

Each of the 613 addresses with three records apiece has an even
bigger problem. If each address has space for only onerecord,therewill
.V

be two overflow records per address. Correspondingto these addresses

W

will be a total of 2 x 613 = 1226 overflow records. This is a bad situation.

W

We have thousandsof records that do notfit into the addresses assigned

W

by the hashing function. We need to develop a method for handling

these overflowrecords. But first, let’s try to reduce the numberof over-
flow records.

11.4 How Much Extra Memory Should Be Used?

We have seenn the importance of choosing a good hashing algorithm to

thereby decrease the average‘search length) is to use extra memory. The

tools developed in the previoussection can be used to help us determine
the effect of the use of extra memory on performance.

https://hemanthrajhemu.github.io
How Much Extra Memory Should Be Used? 481

11.4.1 Packing Density

The term packing density refers to the ratio of the numberofrecords to be
stored (r) to the numberof available spaces (N):3
\ .
Numberof records
+ = packing densit
Numberof spaces ‘N y
For example, if there are 75 records (n = 75) and 100 addresses (N = 100),
the packing density is
73
=0. 75 = 75%
“100

The packing density gives a measure of the amount of space in file

that is used, andit is the only such value neededto assess performancein
a hashing environment, assuming that the hash method used gives a
reasonably random distribution of records. The rawsize ofa file and its
address space do not matter; what is important is the relative sizes of the
M
CO
two, which are given by the packing density.
P.

Think of packing density in terms of tin cans lined up on a 10-foot

O

length of fence. If there are ten tin cans and you throw rock, there is a
LO

certain likelihood that you will hit a can. If there are twenty cans on the
TU

samelengthof fence, the fence has a higher packing density and your rock
is more likely to hit a can. Soit is with records in a file. The more records
.V

there are packedinto given file space, the morelikelyit is that a collision
W

will occur when a new recordis added.

W

"We needto decide how much space weare willing to waste to reduce
W

the numberofcollisions. The answer dependsiin large measure on partic-

ular circumstances. We want to haveas few collisionsas possible, but not,
for example, at the expenseof requiring thefile to use two disks instead
of one.

11.4.2 Predicting Collisions for Different Packing Densities

We need a quantitative description of the effects of changing the packing
density. In particular, we need tobeable to predict the numberofcolli-
sions that are likelyto occurr for a givenpacking density. Fortunately, the
Poissonfunction provides us with just the tool to dothis.

3. We assumehere that only one record can be stored at each address. In fact, that is not necessarily
the case, as wesee later.

https://hemanthrajhemu.github.io
482 Chapter 11 Hashing

You may have noted already that the formula for packing density (r/N)
occurs twice in the Poisson formula
p(x) = (r/N)* @ -/N)
x!
Indeed, the numbers of records (r) and addresses (N) always occur togeth-
er as the ratio r/N. They never occur independently. An obvious implica-
tion of this is that the way records are distributed dependspartly on the
ratio of the numberof records to the numberof available addresses, and
not on the absolute numbersof records or addresses. The same behavioris
exhibited by 500 records distributed among 1000 addresses as by 500 000
records distributed among 1 000 000 addresses.
Suppose that 1000 addresses are allocated to hold 500 records in a
randomly hashed file, and that each address can hold one record. The
packing density forthefile is

M
CO
P.

Let us answer the following questions about the distribution of

O
LO

records amongthe available addressesin the file:

TU

@ How many addresses should have no recordsassigned to them?

m How many addresses should have exactly one record assigned (no
.V

synonyms)? | ' .
W
W

m How many addresses should have one record plus one or more
W

synonyms? = -
m Assuming that only one record can be assigned to each homeaddress,
how manyoverflow records can be expected?
m= Whatpercentage of records should be overflow records?

1. How many addresses should have no records assigned to them? Since

p(O) gives the proportion of addresses with no records assigned, the
numberof such addressesis .
. 0 p-0.5
Np(0) = 1000 x e"°
= 1000 x 0.607
= 607
2. How many addresses should have exactly one record assigned (no
synonyms)?

https://hemanthrajhemu.github.io
How Much Extra Memory Should Be Used? 483

Np(1) = 1000 x (0.5)! e 0.5

1!
= 1000 x 0.303
= 303

3. How many addresses should have one record plus one or more
synonyms? The valuesof p(2), p(3), p(4), and so on give the propor-
tions of addresses with one, two, three, and so on synonyms
assigned to them. Hencethe sum

p(2) + p(3) + pl4) +...

gives the proportion of all addresses with at least one synonym.
This may appear to require a great deal of computation, butit
doesn’t since the values of p(x) grow quite small for x larger than
3. This should make intuitive sense. Since the file is only 50

M
percent loaded, one would not expect very many keys to hash to
CO
any one address. Therefore, the numberof addresses with more
than about three keys hashed to them should be quite small. We
P.

need only compute the results up to p(5) before they become

O

insignificantly small:
LO

p(2) + p(3) + p(4) + p(5) = 0.0758 + 0.0126 + 0.0016 + 0.0002:

TU

= 0.0902
.V
W

The number of addresses with oneor more synonymsisjust the

W

product of Nandthis result:

W

N{[p(2) + p(3) +... ] = 1000 x 0.0902

= 90

4, Assuming that only one record can be assigned to each home address,
how many overflow records could be expected? For each of the
addresses represented by p(2), one record can be stored at the
address and one must be an overflow record. For each address
represented by p(3), one record can be storedat the address,two are
overflow records, and so on. Hence, the expected numberof over-
flow recordsis given by
1x Nx p(2) +2 Nx p(3) +3 Nx p(4) +4 Nx p(5)
= Nx [1X p(2) +2 x p(3) +3 x p(4) +4 p(5)]
= 1000 x [1 x 0.0758 + 2 x 0.0126 + 3 x 0.0016 + 4 x 0.0002]
= 107

https://hemanthrajhemu.github.io
484 Chapter 11 Hashing

5. What percentageof records should be overflow records?If there are 107

overflow records and 500 records in all, then the proportion of
overflow recordsis

107 ~ 0.124 =21.4%

500
Conclusion: if the packing density is 50 percent and each address can
hold only one record, we can expect about 21 percentofall records to be
stored somewhereother than at their home addresses.
Table 11.2 shows the proportionof records that are not stored in their
home addresses for several different packing densities.The table shows
that if the packing densityis 10 percent, then about5 percentofthe time
we try to access a record,there is already anotherrecordthere.If the densi-
ty is 100 percent, then about 37 percentof all records collide with other
records at their home addresses. The 4.8 percentcollision rate that results

M
when the packing density is 10 percent looks very good until you realize
CO
that for every record in yourfile there will be nine unused spaces!
The 36.8 percentthat results from 100 percent usage looks good when
P.

viewed in terms of 0 percent unused space. Unfortunately, 36.8 percent

O

doesn’t tell the whole story.If 36.8 percent of the records are notat their
LO
TU
.V

Table 11.2 Effect of packing density on the proportion

of records not stored at their home addresses :
W
W

_ Packing Synonyms as
W

density (percent) percent of records

10 4.8
20 9.4
30 13.6
40 17.6

50 21.4
60 24.8
70 ' 28.1
80 31.2
90 34.1
100 36.8

https://hemanthrajhemu.github.io
 Collision Resolution by Progressive Overflow 485

home addresses, then they are somewhereelse, probably in manycases

using addresses that are home addresses for other records. The more home-
less records there are, the more contention there is for space with other
homeless records. After a while, clusters of overflow records can form,lead-
ing in some cases to extremely long searches for some of the records.
Clearly, the placementof records that collide is an important matter. Let us
now look at one simple approach to placing overflow records.

11.5 Collision Resolution by Progressive Overflow

Even if a hashing algorithm is very good,it is likely that collisions will

occur, Therefore, any hashing program must incorporate some method for
dealing with records that cannot fit into their home addresses. There are a

M
numberof techniques for handling overflow records, and the search for
CO
ever better techniques continues to be lively area of research. We exam-
ine several approaches, but we concentrate on a very simple one thatoften
P.

works well. The technique has various names, including progressive over-
O

flow and linear probing.

LO
TU

11.5.1 How Progressive Overflow Works

.V

An exampleof a situation in which a collision occurs is shown in Fig. 11.4.

W

In the example, we wantto store the record whose keyis York in the file.
W

Unfortunately, the name York hashes to the same address as the name
W

Rosen, whoserecordis already stored there. Since York cannotfit in its

homeaddress, it is an overflow record. If progressive overflow is used,the
next several addresses are searched in sequence until an empty one is
found. Thefirst free address becomes the address of the record. In the
example, address 9 is the first record found empty,so the record pertain-
ing to York is stored in address 9. |
Eventually we need to find York’s record in thefile. Since Yorkstill
hashesto 6, the search forthe record begins at address 6. It does not find
York’s record there, so it proceedsto look at successive records until it gets
to address 9, where it finds York.
An interesting problem occurs when there is a search for an open
space or for a record at the end ofthefile. Thisis illustrated in Fig. 11.5, in
which it is assumedthat the file can hold 100 records in addresses 0-99.
Blue is hashed to record number99, whichis already occupied byJello.

https://hemanthrajhemu.github.io
486 Chapter 11 Hashing

Novak... '

Rosen... York's home

address (busy)
Jasper... 2ndtry (busy)

Moreley... 3rd try (busy)

4th try (open)

York's actual
address

M
CO
Figure 11.4 Collision resolution with progressive overflow.
P.
O
LO

Since thefile holds only 100 records, it is not possible to use 100 as the next
TU

address, The waythis is handled in progressive overflow is to wrap around.

the address space of the file by choosing address 0 as the next address.
.V

Since address 0 is not occupiedin this case, Blue gets stored in address0.
W

_ What happens if thereis a search for a record but the record was never
W

placed in the file? The search begins, as before, at the record’s home
W

address, then proceedsto look forit in successive locations. Two things can
happen: .
m@ [fan open addressis encountered,the searching routine might assume
this meansthatthe recordis notin thefile; or
m@ If the file is full, the search comés back to whereit began. Onlythenis
it clear that the record is not in the file. When this occurs, or even
when we approachfilling ourfile, searching can becomeintolerably
slow, whetheror not the record being sought is in thefile.
The greatest strength of progressive overflowis its simplicity. In many
cases, it is a perfectly adequate method. There are, however,collision-
handling techniquesthat perform better than progressive overflow, and we
examine someof themlater in this chapter. Now let us look at the effect of
progressive overflow on performance.

https://hemanthrajhemu.github.io
 Collision Resolution by Progressive Overflow 487

—— Key
Blue +

98
Hash: ; ~ Address
.routine, L 99 Jello.

Wrapping around

M
CO
Figure 11.5 Searching for an address beyond the end ofa file.
P.
O

11.5.2 Search Length

LO

Thereason to avoid overflow is, of course, that extra searches (hence, extra
TU

disk accesses) have to occur when a record is not found in its home
.V

address.If there are a lot of collisions, there are going to be a lot of over-
W

flow records taking up spaces where they ought not to be. Clusters of
W

records can form, resulting in the placement ofrecords a long way from
W

home, so many disk accesses are required to retrieve them.

Consider the following set of keys and the corresponding addresses
produced by some hash function.

Key _ Home Address

Adams 20
Bates 21
Cole 21
Dean 22
‘Evans 20

If these recordsare ioaded into an emptyfile and progressive overflow

is used to resolve collisions, only two of the records will be at their home
addresses. All the others require extra accesses to retrieve. Figure 11.6

https://hemanthrajhemu.github.io
488 Chapter. 11 Hashing

Numberof.
‘Actual Home accesses needed.
address address to retrieve

20 Adams... | 20 1.

21 Bates; .. 21 1

22 Cole... 21 2
23 Dean... 22 Z

24 Evans... 20 5

M
25 CO
P.
O

Figure 11.6 Illustration of the effects of clustering of records. As keysare clus-

LO

tered, the numberof accesses required to access later keys can become large.
TU
.V
W

shows where each keyis stored, together with information on how many
W

- accesses are requiredto retrieveit. .

W

The term search length refers to the numberof accesses requiredto

retrieve a record from secondary memory. Inthe context of hashing, the
search length for a record increases every time thereis a collision. If a
recordis a long wayfrom its horne address, the search length may berunac-

search length. The average search lengthis the average numberof times
you can expectto haveto access the disk to retrieve a record. A rough esti-
mate of average search length may be computed byfindingthetotal search
length (the sum ofthe search lengthsof the individual records) and divid-
ing this by the numberofrecords:

Averagesearch length =
panes —t0tal
searc h lengtords
h.
total numbe rofrec

In the example, the average search length for thefive recordsis

https://hemanthrajhemu.github.io
Collision Resolution by Progressive Overflow 489

L+14+24+24+5_ 2.2
5
With nocollisionsatall, the average search lengthis ], since only one
access is needed to retrieve any record, (We indicated earlier that an algo-
rithm that distributes records so evenly nocollisions occuris appropriate-
ly called a perfect hashing algorithm, and we mentioned that,
unfortunately, such an algorithm is almost impossible to construct.) On
the other hand,if a large numberof the records in file results in colli-
sions, the average search length becomesquite long. There are waysto esti-
mate the expectedaverage search length,given variousfile specifications,
and wediscuss them in later section.
It turns out that, using progressive overflow, the average search length
increases very rapidly as the packing density increases. The curve in Fig.
11.7, adapted from Peterson (1957), illustrates the problem.If the packing
density is kept as lowas 60 percent, the average record takes fewer than two

M
tries to access, but for a much moredesirable packing density of 80 percent
CO
or more,it increases very rapidly.
P.

Average search lengths of greater than 2.0 are generally considered

O

unacceptable, so it appears that it is usually necessary to use less than 40

LO

percentof your storage spaceto get tolerable performance. Fortunately, we

TU

can improve onthis situation substantially by making one small change to

.V
W

5
W
W

Average 3 oo ‘ hy
search
length

20 40 60 80. 100

Figure 11.7 Average search length versus packing density in a hashedfile in

which one record can be stored per address, progressive overflow is used to
resolve collisions, and the file has just been loaded.

https://hemanthrajhemu.github.io
490 Chapter 11 Hashing

our hashing program. The change involves putting more than one record
at a single address.

11.6 Storing More Than One Record per

Address: Buckets

Recal] that when a computerreceives information from disk,it is just

- about as easy for the I/O system to transfer several recordsasit is to trans-
fer a single record. Recall too that sometimesit might be advantageousto
think of records as being grouped together in blocks rather than stored
individually. Therefore, why not extend the idea of a record address in a
file to an addressof a group of records? The word bucket is sometimes used
to describe a block of records that is retrieved in one disk access, especial-

M
lywhen those records are seen as sharing the same address. On sector-
CO
addressingdisks, a bucket typically consists of one or more sectors; on
block-addressing disks, a bucket might be a block. .
P.

Consider the following set of keys, which is to be loaded into a

O
LO

hashfile.
TU

Key Home Address

.V

Gre en 30
W

Hall 30
W

Jenks 32
W

King 33
Land 33
Marx 33.
Nutt 33

Figure 11.8 illustrates part of a file into which therecords with these keys
are loaded. Each addressin the file identifies a bucket capable of holding
the records corresponding to three synonyms. Only the record corre-
sponding to Nutt cannot be accommodated in a home address.
Whena recordis to be stored orretrieved, its home bucket address is
determined by hashing. The entire bucket is loaded into primary memory.
An in-memorysearch through successive records in the bucket can then
be usedto find the desired record. When a bucketis filled, we still have to

https://hemanthrajhemu.github.io
Storing More Than One Record per Address: Buckets 491

Bucket contents

Green... Hall...

Jenks... .
(Nutt. . . is
King... Land... Marks... an overflow
record)

Figure 11.8 An illustration of buckets. Each bucket can hold up to three

records. Only one synonym (Nutt) results in overflow.

M
CO
WOITy about the record overflow problem (as in the case of Nutt), but this
occurs muchless often when buckets are used than when each address can
P.

hold only one record.

O
LO

11.6.1 Effects of Buckets on Performance

TU

Whenbuckets are used, the formula used to compute packing densityis

.V

changedslightly since each bucket address can hold more than one record.
W

To compute how densely packeda file is, we need to consider both the
W

number of addresses (buckets) and the numberof records we can put at

W

each address (bucket size). If Nis the number of addresses and b is the
numberof records thatfit in a bucket, then bN is the numberofavailable
locations for records.If ris still the numberof records in thefile, then

Packing g density
ty = =a
ON

Suppose we havea file in which 750 recordsare to be stored. Consider

the following two ways we might organizethefile.
m Wecanstore the 750 data records among 1000 locations, where each
Jocation can hold one record. The packing density in this case is

229 _ 75%
1000

https://hemanthrajhemu.github.io
492 Chapter 11 Hashing

m Wecanstore the 750 records among 500 locations, where each loca-
tion has a bucketsize of 2. There arestill 1000 places (2 x 500) to store
the 750 records, so the packing densityisstill

[= 0.75 = 75%

Since the packing density is not changed, we might at first not expect
the use of buckets in this way to improve performance, but in fact it does
improve performance dramatically. The key to the improvement is that,
although there are fewer addresses, each individual address has more room
for variation in the numberof records assigned to it. _
Let’s calculate the difference in performance for these two ways of
storing the same number of records in the same amount of space. The
starting pointfor ourcalculations is the fundamental description of each
file structure.

M
File without buckets File with buckets
CO
Numberof records r= 750 r= 750
P.
O

Numberof addresses N= 1000 N= 500

LO

Bucket size b=] b=2

TU

Packing density 0.75 0.75

.V

Ratio of records to addresses r/N = 0.75 r/N=1.5

W

To determine the numberof overflow records that are expected in the

W

case of eachfile, recall that when a random hashing functionis used, the
W

Poisson function
(r/N}* e (ND
p(x) =
x!

gives the expected proportion of addresses assigned x records. Evaluating

the function for the two different file organizations, we find that records
are assigned to addresses accordingto the distributions that are shown in
Table 11.3.
Wesee from the table that when buckets are not used, 47.2 percentof:
the addresses have norecords assigned, whereas when two-record buckets:
are used, only 22.3 percent of the addresses have no records assigned. This
should make intuitive sense—since in the two-record case there are only
half as many addresses to choose from,it stands to reason that a greater
proportion of the addresses are chosen to containat least one record.

https://hemanthrajhemu.github.io
Storing More Than One Record per Address: Buckets 493

Table 11.3 Poisson distributions for two different

file organizations.

File without ‘File with

buckets ‘buckets
p(x) (r/N = 0.75) (r/N = 1.5)
p(0) 0.472 0.223
pd) 0.354 0.335
p(2) - 0.133 0.251

p(3) 0.033 0.126

p(4) 0.006 0.047
p(5) 0.001 0.014
p(6) — 0.004
2(7) _ 0.001
M
CO
P.

Note that the bucket columnin Table 11.3 is longer than the nonbuck-
O
LO

et column. Does this mean that there are more synonymsin the bucket
case than in the nonbucketcase? Indeed it does, but half of those syno-
TU

nymsdo notresult in overflow records because each bucket can hold two
.V

records. Let us examine this further by computing the exact number of

W

overflow records likely to occur in the two cases.

In the case of the file with bucket size 1, any address thatis assigned
W

exactly one record doesnot have any overflow. Any address with more
W

than one record does have overflow. Recall that the expected number of
overflow records is given by
Nx [1 x p(2) +2 x p(3) +3 x p(4) +4 p(5)+...]
which, for r/N = 0.75 and N= 1000, is approximately

1000 x [1 x 0.1328 + 2 x 0.0332 + 3 x 0.0062 + 4 x 0.0009 + 5 x 0.0001)

= 222
The 222 overflow records represent 29.6 percent overflow.
In the case of the bucket file, any address thatis assigned either one or
two records does not have overflow. The value of p(1) (with r/N =1.5)
gives the proportion of addresses assigned exactly one record, and p(2)
(with r/N = 1.5) gives the proportion of addresses assigned exactly two
records.It is not until we get to p(3) that we encounter addresses for which

https://hemanthrajhemu.github.io
A494 Chapter 11 Hashing

there are overflow records. For each address represented by p(3), two
records can be stored at the address, and one must be an overflow record.
Similarly, for each address represented by p(4), there are two overflow
records, and so forth. Hence, the expected numberofoverflow records in
the bucketfile is

Nx {1 x p(3) + 2 x p(4) +3 x p(5) + 4x p(6) +...]

whichfor'r/N= 1.5 and N= 500 is approximately
500 x (1 x 0.1255 + 2 x 0.0471 + 3 x 0.0141 + 4 x 0.0035 + 5 x 0.0008]
= 140

The 140 overflow records represent 18.7 percent overflow.

We have shown that with one record per address and a packing densi-
ty of 75 percent, the expected numberof overflow records is 29.6 percent.
When 500 buckets are used, each capable of holding tworecords, the pack-

M
ing density remains 75 percent, but the expected number of overflow
CO
records drops to 18.7 percent. That is about a 37 percent decrease in the
numberof times the program has to look elsewhere for a record. As the
P.

bucket size gets larger, performance continues to improve.-

O

Table 11.4 showsthe proportions ofcollisions that occurfor different

LO

packing densities and for different bucketsizes. We see from thetable, for
TU

example, that if we keep the packing density at 75 pércent and increase

the bucket size to 10, record accesses result in overflow only 4 percent of
.V

the time. |
W

It should beclear that the use of buckets can improve hashing perfor-
W

mancesubstantially. One mightask, “How big should buckets be?” Unfor-.

W

tunately, there is no-simple answer to this question because it dependsa

great deal on a numberofdifferent characteristics of the system, including
the sizes of buffers the operating system can manage, sector and track
capacities on disks, and access times of the hardware(seek, rotation, and
data transfer times). _
As a rule, it is probably not a good idea to use buckets larger than a
track (unless recordsare very large). Even a track, however, can sometimes
be too large when one considers the amountof timeit takes to transmit an
entire track, as compared with the amount of time it takes to transmit a
few sectors. Since hashing almost always involves retrieving only one
record per search, any extra transmission time resulting from the use of
extra-large buckets is essentially wasted.
In manycases a single cluster is the best bucket size. For example,
suppose that a file with 200-byterecordsis to be stored on a disk system
that uses 1024-byte clusters. One could consider each cluster as a bucket,

https://hemanthrajhemu.github.io
Storing More Than One Record per Address: Buckets 495

Table 11.4 Synonyms causing collisions as a percent of recordsfor different

packing densities and different bucket sizes

Packing " Bucketsize

density
(%) 1 2 5 10 100
10 4.8 0.6 0.0 0.0 0.0
20 9.4 2.2 0.1 0.0 0.0
30 13.6 4.5 04 00 0.0
40 17.6 7.3 11 0.1 0.0

50 21.3 11.4 2.5 0.4 0.0

60 24.8 13.7 4.5 1.3 0.0
70 28.1 17.0 7.1 2.9 0.0

75 29.6 18.7 8.6 4.0 0.0

M
CO
80 31.2 20.4 11.3 5.2 0.1
90 34.1 23.8 13.8 8.6 0.8
P.
O

100 36.8 27.1 17.6 12.5 4.0

LO
TU

store five records per cluster, and let the remaining 24 bytes go unused.
.V

Sinceit is no more expensive,in terms of seek time, to access a five-record

W

cluster than it is to access a single record, the only losses from the use of
W

buckets are the extra transmission time and the 24 unused bytes.
W

The obvious question nowis, “How do improvements in the number

of collisions affect the average search time?” The answer depends in large
measure on characteristics of the drive on which thefile is loaded.If there
are a large number oftracks in each cylinder, there will be verylittle seek
time because overflow records will be unlikely to spill over from one cylin-
der to another. If, on the other hand, there is only one track percylinder,
seek time could be a major consumerofsearch time.
A less exact measure of the amount of time required to retrieve a
record is average search length, which weintroducedearlier. In the case of
buckets, average search length represents the average numberof buckets
that must be accessedto retrieve a record. Table 11.5 shows the expected
average searchlengthsforfiles with different packing densities and bucket
sizes, given that progressive overflow is used to handlecollisions. Clearly,
the use of buckets seems to help a great deal in decreasing the average
search length. The bigger the bucket, the’shorter the search length.

https://hemanthrajhemu.github.io
496 Chapter 11 Hashing

Table 11.5 Average numberof accesses required irin a successful search by

progressive overflow,

Packing Bucketsize
density
(%) 1 2 3 10 100 —

10 1.06 1.01 1.00 1.00 1.00

30 1.21 1.06 1.00 1.00 1.00
. 40 1.33 1.10 1.01 1.00 1,00
50 1.50 1.18 1.03 1.00 1.00
60 1.75 1.29 1.07 1.01 1.00
70 2.17 1.49 1.14 1.04 1.00
380 3.00 1.90 1.29 Li 1.01
90 5.50 3.15 1.78 1.35 1.04

M
95 10.50 5.60 2.70
CO 1.80 1.10

Adapted from Donald Knuth, The Art of Computer Programming, Vol. 3, ©1973, Addison-Wesley,
P.

Reading, Mass. Page 536. Reprinted with permission..

O
LO

11.6.2 Implementation Issues

TU

In the early chapters ofthis text, we paid quite a bit of attention to issues
.V

involved in producing, using, and maintaining random-accessfiles with

W

fixed-length records that are accessed byrelative record number (RRN).

W

Since a hashed file is a fixed-length record file whose records areaccessed

W

by RRN,you should already know much about implementinghashedfiles.

Hashedfiles differ from the files we discussed earlier in two important
respects, however:
1. Since a hash function depends on there being a fixed numberofavail-
able addresses, the logical size of a hashed file mustbe fixed before the
file can be populated with records, and it must remain fixed as long as
the same hash function is used. (We use the phrase logical size to leave
open the possibility that physical space be allocated as needed.)
Since the home RRN ofa record in a hashedfile is uniquelyrelated to
its key, any procedures that add,delete, or change a record must do so
without breaking the bond between a recordand its home address.If
this bondis broken, the record is no longer accessible by hashing.
We must keep these.special needs in mind when wewrite programsto
work with hashed files.

https://hemanthrajhemu.github.io
 Storing More Than One Record per Address: Buckets 497

Bucket Structure

The only difference between a file with buckets and one in which each
address can hold only one keyis that with a bucketfile each address has
enough space to hold more than one logical record. All records that are
housed in the same bucket share the same address. Suppose, for example,
that we wantto store as many as five names in one bucket. Here are three
such buckets with different numbersof records.

Anempty bucket: JO] // /// fide “hed dd Afddd Piddf¢

Two entries: 2| JONES ARNSWORTH| / / / 7 / Piddd |

M
A full bucket: 5| JONES ARNSWORTH| STOCKTON| BRICE THROOP
CO
P.

Each bucket contains a counter that keeps track of how manyrecords

O
LO

it has stored in it. Collisions can occur only whenthe addition of a new
record causes the counter to exceed the numberof records a bucket can
TU

hold.
.V

' The countertells us how many datarecords are stored in a bucket, but
W

it does not tell us which slots are used and which are not. We need a way to
W

tell whether a recordslot is empty. One simple wayto dothis is to use a

W

‘special marker to indicate an emptyrecord,just as we did with deleted

‘records earlier. We use the key value ///// to mark empty records in the
preceding illustration.

Initializing a File for Hashing

Since the logicalsize of a hashed file must remain fixed, it makes sense in
most cases to allocate physical space for the file before we begin storing
data recordsin it. This is generally done by creating file of empty spaces
for all records, then filling the slots as they are needed with the data
records, (It is not necessary to construct a file of empty records before
putting data in it, but doing soincreasesthelikelihood that recordswill be
stored close to one another on thedisk, avoids the error that occurs when
an attempt is made to read a missing record, and makesit easy to process
the file sequentially, without having to treat the empty records in any
special way.)

https://hemanthrajhemu.github.io
498 Chapter 11 Hashing

Loading a HashFile
A program that loads a hash. file is similar in many ways to earlier
programsweused for populatingfixed-length record files, with two differ-
ences. First, the program uses the function hash to produce a home
address for each key. Second, the program looksfor a free space for the
record by starting with the bucket stored at its home address and then,if
the home bucketis full, continuing to look at successive buckets until one
is found that is not full. The newrecord is inserted in this bucket, and the
bucketis rewritten to thefile at the location from whichit was loaded.
If, as it searches for an empty bucket, a loading program passes the
maximum allowable address, it must wrap around to the beginning
address. A potential problem occurs in loading a hash file when so many
records have been loaded into the file that there are no empty spacesleft.
A naive search for an open slot can easily result in an infinite loop.
Obviously, we want to prevent this from occurring by having the program

M
make sure that there is space available for each new record somewherein
CO
thefile.
P.

- Anotherproblem that often arises when addingrecordstofiles occurs

O

when an attempt is made to add a record thatis already stored in thefile.

LO

If there is a danger of duplicate keys occurring, and duplicate keys are not
TU

allowed in the file, some mechanism must be found for dealing with this.
problem.
.V
W
W

11.7 Making Deletions

W

Deleting a record from a hashedfile is more complicated than adding a

record for two reasons:
BW The slot freed by the deletion must not be allowed. to hinder later
searches; and
m It should be possible to reuse the freed slot for later additions.
Whenprogressive overflowis used, a search for a record terminatesif
an open address is encountered. Because of this, we do not want to leave
open addresses that break overflow searches improperly. The following
example illustrates the problem. . |
Adams, Jones, Morris, and Smith are stored in a hash-file in which
each address can hold one record. Adams and Smith both are hashed to
address 5, and Jones and Morris are hashedto address6. If they are loaded

https://hemanthrajhemu.github.io
' Making Deletions’ =< > > -- 499

Home Actual
Record address address

4
Adams 5 5

Jones 6 6 5 Adams...

Morris 6 7 6 Jones...

Smith 5 8 4 Morris...

8 Smith...

Figure 11.9 File organization before deletions.

M
in alphabetical order using progressive overflow for collisions, they are
CO
storedin the locations shownin Fig. 11.9.
P.

A search for Smith starts at address 5 (Smith’s home address), succes-

O

‘sively looks for Smith at addresses 6, 7, and 8, then finds Smith at 8. Now
LO

suppose Morris is deleted, leaving an empty space,as illustrated in Fig.

11.10. A search for Smith again starts at address 5, then looks at addresses
TU

6 and 7. Since address 7 is now empty,it is reasonable for the program to

.V

conclude that Smith's recordis not in the file.

W
W

11.7.1 Tombstones for Handling Deletions

W

In Chapter 6 we discussed techniques for dealing with the deletion prob-

lem. One simple technique weuse for identifying deleted records involves
replacing the deleted record (orjust its key) with a marker indicating that
a record oncelived there but no longer does. Such a markeris sometimes
referred to as a tombstone (Wiederhold, 1983). The nice thing about the
use of tombstonesis that it solves both of the problems described previ-
ously:
m Thefreed space does not break a sequence of searches for a record;
and ,
m Thefreed space is obviously available and maybe reclaimedforlater
additions.
Figure 11.11 illustrates how the samplefile might look after the tomb-
stone ###### is inserted for the deleted record. Now a search for Smith

https://hemanthrajhemu.github.io
500 Chapter 11 Hashing

4 5 Adams. .

5 Adams... 6 Jones...

6 Jones... 7 HEEHRAF

7 8 Smith...

8 Smith...

Figure 11.11 The same file

as in Fig. 11.9 after the
Figure 11.10 The same insertion of a tombstone
organization asin Fig. 11.9, for Morris.

M
with Morris deleted.
CO
P.
O

does not halt at the empty record number7. Instead,it uses the ###### as
LO

an indication that it should continue the search.

It is not necessary to insert tombstones every time a deletion occurs.
TU

For example, supposein the preceding example that the record for Smith
.V

is to be deleted. Since the slot following the Smithrecordis empty, noth-

W

ing is lost by marking Smith’s slot as emptyrather than inserting a tomb-

W

stone. Indeed,it is unwise to insert a tombstone whereit is not needed.(If,

W

after putting an unnecessary tombstone in Smith’s slot, a new record is

added at address 9, how would a subsequent unsuccessful search for Smith
be affected?)

11.7.2 Implications of Tombstonesfor Insertions

With the introduction of the use of tombstones, the insertion of records
becomes slightly more difficult than our earlier discussions imply.
Whereas programsthat perform initial loading simply searchfor thefirst
occurrence of an empty recordslot (signified by the presenceofthe key
////1), it is now permissible to insert a record whereeither ///// or ######
occurs as the key.
This new feature, whichis desirable because it yields a shorter average
search length, brings with it a certain danger. Consider, for example, the
earlier example in which Morrisis deleted, giving the file organization

https://hemanthrajhemu.github.io
Making Deletions 504

shownin Fig. 11.11. Now suppose you want a programto insert Smith
into the file. If the program simply searches until it encounters a ####4#,
it never notices that Smith is already in the file. We almost certainly don’t
want to put a second Smith recordintothefile, since doing so means that
later searches would never find the older Smith record. To prevent this
from occurring, the program must examinethe entire cluster of contigu-
ous keys and tombstones to ensure that-no duplicate key exists, then go
back and insert the record in thefirst available tombstone, if there is one.

11.7.3 Effects of Deletions and Additions on Performance

The use of tombstones enables our search algorithms to work and helpsin
storage recovery, but one can still expect some deterioration in perfor-
mance after a numberof deletions and additions occur within file.
Consider, for example, our little four-record file of Adams, Jones,

M
Smith, and Morris. After deleting Morris, Smithis one slotfurther from its
CO
homeaddress than it needs to be. If the tombstone is never to be used to
store another record, every retrieval of Smith requires one more access
P.

than is necessary. More generally, after a large numberof additions and

O

deletions, one can expect to find many tombstones occupying places that
LO

could be occupied by records whose home records precede them butthat

TU

are stored after them. in effect, each tombstone represents an unexploited

.V

opportunity to reduce by one the numberof locations that must be

scanned while searching for these records.
W

.Some experimental studies show thatafter a 50 percent to 150 percent

W

turnover of records, a hashedfile reaches a point of equilibrium, so aver-

W

age search lengthis aslikely to get betteras it is to get worse (Bradley, 1982;
Peterson, 1957). By this time, however, search performancehas deteriorat-
ed to the point at which the average recordis three times as far (in terms
of accesses) from its home addressas it would be after initial loading. This
means, for example, that if after original loading the average search length
is 1.2, it will be about 1.6 after the point of equilibrium is reached.
There are three types of solutions to the problemof deteriorating
average search lengths. One involves doing a bit of local reorganizing every
time a deletion occurs. For example, the deletion algorithm might exam-
ine the records that follow a tombstone to see if the search length can be
shortened by moving the record backward toward its home address.
Another solution involves completely reorganizing the file after the aver-
age search length reaches an unacceptable value. A third type ofsolution
involves using an altogether differentcollision resolution algorithm.

https://hemanthrajhemu.github.io
“502 Chapter 11 -Hashing

11.8 OtherCollision Resolution Techniques

Despite its simplicity, randomized hashing using progressive overflow with

reasonably sized buckets generally performs well. If it does not perform
well enough, however, there are a numberof variations that may perform
even better. In this section we discuss some refinements that can often
improve hashing performance when using external storage.

11.8.1 Double Hashing

One of the problems with progressive overflow is that if manyrecords
hash to buckets in the same vicinity, clusters of records can form. As the
packing density approachesone,this clustering tends to lead to extremely
long searches for some records. One method for avoiding clustering1s to

M
store overflow records a long way from their home addresses by double
CO
hashing. With double hashing, when-a collision occurs, a second hash
P.

function is applied to the key to produce a number thatis relatively

O

prime to the number of addresses.4 The value c is added to the home

LO

address to produce the overflow address.If the overflow addressis already

occupied, cis addedto it to produce another overflow address. This proce-
TU

dure continues until a free overflow address is found.

.V

Double hashing doestend to spread outthe records in file, butit

W

suffers from a potential problem that is encountered in several improved

W

overflow methods: it violates locality by deliberately moving overflow

W

records some distance from their home addresses, increasing the likeli-
hoodthat the disk will need extra time to get to the new overflow address.
If the file covers more than one cylinder, this could require an expensive
extra head movement. Double hashing programscansolvethis problem if
they are able to generate overflow addressesin such a way that overflow
records are kept on the samecylinder as homerecords.

11.8.2 Chained Progressive Overflow

Chained progressive overflow is another technique designed to avoid the
problems caused by clustering. It works in the same manneras progressive
overflow, except that synonyms are linked together with pointers. Thatis,
each homeaddress contains a numberindicating the location of the next

4. If Nis the number of addresses, then cand Narerelatively prime if they have no common divisors.

https://hemanthrajhemu.github.io
OtherCollision Resolution Techniques 503

Home Actual Search

Key address address length
Adams 20: 20 ]
Bates 21 2) |
Cole 90—. 22 3
Dean - 7 2) 23 3
Evans 24 24 !
Flint 90. 25 6
Average search length = (1 + 1 +343 + 1 + 6/6 = 2.5

Figure 11.12 Hashing with progressive overflow.

record with the same home address. The next record in turn contains a
pointerto the following record with the sarme home address, and so forth.
The neteffect of this is that for each set of synonymsthereis a linked list

M
connecting their records, and it is this list that is searched when a recordis
CO
sought.
P.

The advantage of chained progressive overflow over simple progres-

O

sive overflow is that only records with keys that are synonymsneed to be
LO

accessed in any given search. Suppose, for example, that the set of keys
shownin Fig. 11.12 is to be loaded in the order showninto a hash file with
TU

bucket size 1, and progressive overflow is used. A search for Cole involves
.V

an access to Adams (a synonym) and Bates (not a synonym). Flint, the

W

worst case, requires six accesses, only two of which involve synonyms.
W

Since Adams, Cole, and Flint are synonyms, a chaining algorithm

W

forms a linked list connecting these three names, with Adamsat the head
of the list. Since Bates and Deanare also synonyms, they form a second
list. This arrangementis illustrated in Fig. 11.13. The average search length
decreases from 2.5 to : Se
1+1424+2+14+3
= 1.7
6
The use of chained progressive overflow requires that we attend to
some details that are not required for simple progressive overflow. First, a
link field must be added to each record, requiring the use ofa little more
storage. Second, a chaining algorithm must guaranteethatit is possible to
get to any synonym bystarting at its home address. This second require-
mentis not a trivial one, as the following example shows.
Suppose that in the example Dean’s homeaddressis 22 instead of 21.
Since, by the time Deanis loaded, address 22 is already occupied by Cole,

https://hemanthrajhemu.github.io
504 Chapter 11 Hashing

Home Actual Addressof Search

address address Data next synonym length

20 20 Adams... ‘ 22:0 1

21 21 Bates... 88 1
20 22 Cole... - 25° 2

21 23 ‘| Dean... Pai 2

24 24 Evans... = 1
20 25 Flint... | Soa 3

M
Figure 11.13 Hashing with chained progressive overflow. Adams, Cole, and
CO
Flint are synonyms; Bates and Dean are synonyms.
P.
O
LO

Deanstill ends up at address 23. Does this mean that Cole’s pointer should
point to 23 (Dean’s actual address) or to 25 (the address of Cole’s synonym
TU

Flint)? If the pointeris 25, the linked list joining Adams, Cole, and Flintis
.V

kept intact, but Deanislost. If the pointer is 23, Flintis lost.

W

The problem here is that a certain address (22) that should be‘occu-
W

pied by a homerecord (Dean)is occupied by a different record. One solu-

W

tion to the problem is to require that every address qualifying as a home

address for somerecordin thefile actually hold a homerecord. The prob-
lem can be handled easily when file is first loaded by using a technique
called two-pass loading.
Two-pass loading, as the name implies, involves loading a hash file in
two passes. On thefirst pass, only homerecordsare loaded. All records that
are not home records are kept in a separate file. This guarantees that no
potential home addresses are occupied by overflow records. On the second
pass, each overflow record is loaded andstored in oneof the free addresses
according to whatevercollision resolution techniqueis being used..
Two-pass loading guarantees that every potential home address actu-
ally is a home address, so it solves the problemin the example. It does not
guarantee that later deletions and additions will not re-create the same
problem, however. As lors as the file is used to store both home records
and overflow records, there remains the problem of overflow records

https://hemanthrajhemu.github.io
Other Collision Resolution Techniques 505

displacing newrecords that hash to an address occupied by an overflow

record.
The methodsused for handling these problemsafterinitial loading are
somewhat complicated and can, in a very volatile file, require many extra
disk accesses. (For more information on techniquesfor maintaining point-
ers, see Knuth, 1998 and Bradley, 1982.) It would be nice if we could some-
how altogether avoid this problem of overflow lists bumping into one
another, andthat is what the next method does. ae

11.8.3 Chaining with a Separate Overfiow Area

One way to keep overflow records from occupying home addresses where
they shouldnotbe is to move themall to a separate overflow area. Many
hashing schemesare variations of this basic approach. The set of home
addresses is called the prime data area, and the set of overflow addressesis

M
called the overflow area. The advantage ofthis approachis thatit keepsall
CO
unused but potential home addressesfree for later additions.
In terms of the file we examinedin the precedingsection, the records
P.

for Cole, Dean, and Flint could have been stored in a separate overflow
O
LO

area rather than in potential home addresses for later-arriving records

(Fig. 11.14). Now no problem occurs when a new recordis added. If its
TU

home address has room, it is stored there. If not, it is moved to the over-
.V

flow file, whereit is added to the linkedlist that starts at the home address.
W
W

Home Primary Overflow

W

address data area area

20 Adams... | 0 —-————p 0 Cole... 2

21 Bates... Lj 1 Dean... —-1

22 2 Flint... -l je

23 3 |

24 Evans. . - —I1

Figure 11.14 Chaining toa separate overflow area. Adams, Cale, andFlint are
synonyms; Bates and Dean are synonyms.

https://hemanthrajhemu.github.io
506 Chapter 11 Hashing

If the bucket size for the primaryfile is large enough to prevent exces-
sive numbers of overflow records, the overflowfile can be a simple entry-
sequencedfile with a bucketsize of 1. Space can be allocated for overflow
‘records only whenit is needed. |
The use of a separate overflow area simplifies processing somewhat
and would seem to improve performance, especially when many additions
and deletions occur. However, this is not always the case.If the separate
overflow area is on a different cylinder than is the home address, every
search for an overflow record will involve a very costly head movement.
Studies show that access timeis generally worse when overflow recordsare
stored in a separate overflow area than when they are stored in the prime
overflow area (Lum,1971).
One situation in which a separate overflow area is required occurs
whenthe packing density is greater than one—there are more records than
home addresses.If, for example, it is anticipated that a file will grow

M
beyondthe capacity oftheinitial set of home addresses and that rehashing
CO
the file with a larger addressspace is not reasonable, then a separate over-
flow area must be used.
P.
O
LO

11.8.4 Scatter Tables: Indexing Revisited

TU

Suppose you have a hashfile that contains no records, only pointers to

.V

‘records. The file is obviously just an index that is searched by hashing

W

rather than by some other method. The term scatter table (Severance,
W

1974) is often applied to this approachtofile organization. Figure 11.15

W

illustrates the organization ofa file using a scatter table.

fp} Adams. ..| tpl Cole... Ss Flint...1

Bates... ‘—Pl Dean... =

Evans. ..|—~1

Figure 11.15 Example ofa scatter table structure. Because the hashed partis an index,
the data file may be organized in any waythat is appropriate.

https://hemanthrajhemu.github.io
 Patterns of Record Access , "507

Thescatter table organization provides many of the same advantages

simple indexing generally provides, with the additional advantage thatthe
search of the indexitself requires only one access. (Of course, that one
-access is one more than other formsof hashing require, unless the scatter
table can be kept in primary memory.) The datafile can be implemented
in many different ways. For example,it can be a set of linked lists of
synonyms (as shownin Fig. 11.15), a sorted file, or an entry-sequenced
file. Also, scatter table organizations conveniently supportthe use of vari-
able-length records, For more information onscatter tables, see Severance
(1974).

11.9 Patterns of Record Access

M
Twenty percent of the fishermen catch 80 percent of the fish.
CO
Twentypercent of the burglars steal 80 percent of the loot.
P.

. L.M. Boyd
O
LO

The use of different collision resolution techniquesis not the only nor
necessarilythe best way to improve performance in a hashedfile. If we
TU

know something aboutthe patterns of record access, for example, then it

.V

is often possible to use simple progressive overflow techniquesand still

W

achieve very good performance.

W

Suppose you have a grocery store with 10 000 different categories of

W

grocery items and you have on your computera hashed inventoryfile with
a record for each of the 10 000 items that your company handles. Every
time an item 1s purchased,the record that correspondsto that item must
be accessed. Since the file is hashed, it is reasonable to assume that the
10 000 records are distributed randomly amongthe available addresses that
make up thefile. Is it equally reasonable to assume that the distribution of
accesses to the records in the inventory are randomlydistributed? Probably
not. Milk, for example, will be retrieved very frequently, brie seldom.
There is a principle used by economists called the Pareto Principle, or
The Conceptof the Vital Few and the Trivial Many, whichin file terms says
that a small percentage of the records in a file accountfor a large percent-
age of the accesses. A popular version of the Pareto Principle is the 80/20
Rule of Thumb: 80 percentof the accesses are performed on 20 percent of
the records. In our groceriesfile, milk would be among the 20 percent
high-activity items, brie among the rest.

https://hemanthrajhemu.github.io
508 Chapter 11 Hashing

Wecannot take advantage of the 80/20 principle in file structure

unless we know something about the probable distribution of record
accesses. Once wehave this information, we need to find a wayto place the
high-activity items where they can be found with as few accesses as possi-
ble. If, when items are loaded intoa file, they can be loaded in such a way
that the 20 percent (moreor less) that are most likely to be accessed are
loaded at or near their home addresses, then most of the transactions will
access records that have short search lengths, so the effective average search
length will be shorter than the nominal average search length that we
defined earlier.
For example, suppose our grocery store’s file handling program keeps
track of the numberof times each item is accessed during a one-month
period. It mightdo this by storing with each record a counter thatstarts at
zero and is incremented every time the item is accessed. At the end of the
month therecordsforall the items in the inventory are dumped ontoa file
that is sorted in descending order according to the numberoftimes they

M
have been accessed. When the sortedfile is rehashed and reloaded,thefirst
CO
records to be loaded are the onesthat, according to the previous month’s
P.

experience, are most likely to be accessed. Since they are the first’ ones
O

loaded, they are also the ones most likely to be loaded into their home
LO

addresses. If reasonably sized buckets are used, there will be very few, if
TU

any, high-activity items that are not in their home addresses and therefore
retrievable in one access.
.V
W
W
W

SUMMARY
There are three major modes for accessing files: sequentially, which
provides O(N) performance, through tree structures, which can produce
O(log,N) performance, and directly. Direct access provides O(1) perfor-
mance, which means that the numberof accesses required to retrieve a
record is constant and independentofthesize ofthefile. Hashingis the
primary form of organization used to provide direct access.
Hashing can provide faster access than most of the other organiza- .
tions we study, usually with verylittle storage overhead,andit is adapt- .
able to most types of primary keys. Ideally, hashing makesit possibleto
find any record with only one disk access, butthis idealis rarely achieved.
The primary disadvantage of hashingis that hashedfiles may not be sort-
ed by key.

https://hemanthrajhemu.github.io
Summary 509

Hashing involves the application of a hash function h(K) to a record

key K to produce an address. The addressis taken to be the home address
of the record whose key is K, and it forms the basis for searching for the
record. The addresses produced by hash functions generally appear to be
random.
When two or more keys hash to the same address, they are called
synonyms. If an address cannot accommodateall of its synonyms,collisions
result. When collisions occur, some of the synonyms cannotbe stored in
the home address and mustbe stored elsewhere.Since searches for records
begin with homeaddresses, searches for records that are not stored at their
home addresses generally involve extra disk accesses. The term average
search length is used to describe the average numberof disk accesses that
are requiredtoretrieve a record. An average search length of 1 is ideal.
Muchof the study of hashing deals with techniquesfor decreasingthe
numberandeffects of collisions. In this chapter we look at three general

M
approachesto reducing the numberofcollisions: CO
m Spreading out the records;
P.

m Using extra memory; and

O

m Using buckets.
LO

Spreading out the records involves choosing a hashing function that

TU

distributes the records at least randomlyover the address space. A uniform

.V

distribution spreads out records evenly, resulting in no collisions. A

W

random or nearly random distribution is much easier to achieve andis

W

usually considered acceptable.

W

In this chapter a simple hashing algorithm is developed to demon-

strate the kinds of operationsthat take place in a hashing algorithm. The
three steps in the algorithmare:
1. Represent the key in numerical form;
2. Fold and add; and
3. Divide by the size of the address space, producing a valid address.
When we examineseveral different types of hashing algorithms, we see
that sometimes algorithmscan be foundthat produce better-than-random
distributions. Failing this, we suggest some algorithms that generally
producedistributions that are approximately random.
The Poisson distribution provides a mathematical tool for examining
in detail the effects of a random distribution. Poisson functions can be
used to predict the numbers of addresseslikely to be assigned0, 1, 2, and

https://hemanthrajhemu.github.io
510 Chapter 11 Hashing

so on, records, given the numberofrecords to be hashed and the number

of available addresses. This allows us to predict the numberof collisions
likely to occur whena file is hashed, the numberof overflow recordslike-
ly to occur, and sometimesthe average search length.
_ Using extra memory is another wayto avoid collisions. Whena fixed
numberofkeys is hashed,the likelihood of synonyms occurring decreases
as the numberofpossible addresses increases. Hence; file organization
that allocates many more addresses than arelikely to be used has fewer
synonyms than one that allocates few extra addresses. The term packing
density describes the proportion of available address space that actually
holds records, The Poisson function is used to determine how differences
in packing density influence the percentage of recordsthat are likely to be
synonyms.
Using buckets is the third method for avoiding collisions. File address-
es can hold one or morerecords, depending on howthefile is organized by

M
the file designer. The number of records that can be stored at a given
CO
address, calléd bucket size, determines the pointat which records assigned
to the address will overflow. The Poisson function can be usedto explore
P.

the effects of variations in bucketsizes and packing densities. Large buck-

O
LO

ets, combined with a low packing density, can result in very small average
search lengths.
TU

Although we can reduce the numberof collisions, we need some.

.V

means to deal with collisions when they do occur. We examined one

W

simple collision resolution technique in detail—progressive overflow. If

W

an attempt to store a new record results in a collision, progressive over-

flow involves searching through the addresses that follow the record’s
W

home address in order until one is found to hold the new record. If a
record is sought and is not foundin its home address, successive address-
es are searched until either the record is found or an empty address is
encountered.
Progressive overflow is simple and sometimes works very well.
However, progressive overflow creates long search lengths when the pack-
ing density is high and the bucketsize is low. It also sometimes produces
clusters of records, creating very long search lengths for new records whose
homeaddressesare in theclusters.
Three problemsassociated with record deletion in hashedfiles are
1. The possibility that emptyslots created by deletions will hinderlater
searches for overflow records;

https://hemanthrajhemu.github.io
Summary- 511

2. The need to recover space made available when recordsare deleted;

and
3. The deterioration of average search lengths caused by empty spaces
keeping records further from homethan they need be.
The first two problems can be solved by using tombstones to mark
spaces that are empty (and can bereused for newrecords) but shouldnot
halt a search for a record, Solutions to the deterioration problem include
local reorganization, complete. file reorganization, and the choice of a
collision-resolving algorithm that does not cause deterioration to occur.
Because overflow records have a major influence on performance,
many different overflow handling techniques have been proposed. Four
such techniques that are appropriate for file applications are discussed
briefly:
1. Double hashing reduces local clustering but may place some overflow

M
records so far from home that they require extra seeks.
CO
2. Chained progressive overflow reduces search lengths by requiring that
P.

only synonyms be examined when record is being sought. For

O

chained overflow to work, every address that qualifies as a home

LO

record for some record in the file must hold a home record..
Mechanisms for making sure that this occurs are discussed.
TU

3. Chaining with a separate overflow area simplifies chaining substantial-

.V

ly and has the advantage that the overflow area may be organized in
W

ways more appropriate to handling overflow records. A danger of this

W

approachIs that it mightlose locality.

W

4, Scatter tables combine indexing with hashing. This approach provides

much moreflexibility in organizing the data file. A disadvantage of
using scatter tables is that, unless the index can be held in memory,it
requires one extra disk access for every search.
Since in many cases certain records are accessed morefrequently than
others (the 80/20 rule of thumb), it is often worthwhile to take access
patterns into account. If we can identify those records that are mostlikely
to be accessed, we can take measures to makesure they are stored closer to
home than less frequently accessed records, thus decreasing the effective
average search length. One such measureis to load the most frequently
accessed records before the others.

https://hemanthrajhemu.github.io
512 Chapter 11 Hashing

KEY TERMS _

Average search length, We define average search length as the surn of the
numberof accesses required for each record in the file divided by the
numberof records in the file. This definition does not take into account
the numberof accesses required for unsuccessful searches, nor doesit
account for the fact that somerecordsare likely to be accessed more
often than others. See 80/20 rule of thumb.
Better-than-random. This term is applied to distributions in which the
records are spread out more uniformly than they would be if the hash
function distributed them randomly. Normally, the distribution
produced by a hash functionis little bit better than random.
Bucket. An area of space on the file that is treated as a physical record for
storage and retrieval purposes butis capable of storing severallogical
records. By storing and retrieving logical records in buckets rather

M
than individually, access times can, in many cases, be improved
CO
substantially.
P.

Collision. Situation in which a record is hashed to an address that does

O

not havesufficient room to store the record. Whena collision occurs,

LO

some meanshas to be found to resolve the collision.

TU

Double hashing. A collision resolution scheme in which collisions are

.V

handled by applying a second hash function to the key to produce a

W

numberc, which is addedto the original address (modulo the number

W

of addresses) as many times as necessary until either the desired

W

record is located or an empty space is found. Double hashing helps

avoid some ofthe clustering that occurs with. progressive overflow.
The 80/20 rule of thumb. An assumption that a large percentage (e.g., 80
percent) of the accesses are performed on a small percentage(e.g., 20
percent)of the recordsin file. When the 80/20 ruleapplies,the effec-
tive average search lengthis determinedlargely by the search lengths of
the more active records, so attempts to make these search lengths short
can result in substantially improved performance.
@
Fold and add. A methodof hashing in which the encodingsoffixed-sized
parts of a key are extracted (e.g., every two bytes) and are added. The
resulting sum can be used to produce an address.
Hashing. A technique for generating a unique homeaddress for a given.
key. Hashingis used when rapid access to a key(or its corresponding
record) is required. In this chapter applications of hashing involve |

https://hemanthrajhemu.github.io
KeyTerms 513

" direct access to recordsin file, but hashing is also often used to access
items in arrays in memory.In indexing, for example, an index might
be organized for hashingrather than for binarysearchif extremely fast
searchingof the index is desired.
Home address. The address generated by a hash function for a givenkey.
If a recordis stored at its home address, then the search length for the
record is 1 because only oneaccessis required to retrieve the record. A
record not at its home address requires more than one access to
retrieve or store.
Indexed hash.Instead of using theresults of a hash to producethe address
of a record, the hash can be used to identify a location in an index that
in turn points to the address of the record. Although this approach
requires one extra access for every search,it makesit possible to orga-
nize the data records in a waythatfacilitates other types of processing,
such as sequential processing.

M
Mid-square method. A hashing method in which a representation of the
CO
key is squared and some digits from the middle ofthe result are used
P.

to producethe address.
O

Minimum hashing. Hashing scheme in which the number of addressesis

LO

exactly equal to the numberof records. No storage spaceis wasted.

TU

Open addressing. See progressive overflow.

.V

Overflow. Thesituation that occurs when a record cannot be stored in its

W

home address.
W

Packing density. The proportion ofallocated file space thatactually holds

W

records. (This is sometimes referred to as load factor.) If file is half

full, its packing density is 50 percent. The packing density and bucket
size are the two most important measures in determiningthelikeli-
hoodof a collision occurring when searching for a record ina file.
Perfect hashing function. A hashing function that distributes records
uniformly, minimizing the numberof collisions. Perfect hashing func-
tions are very desirable, but they are extremely difficult to find for
large sets of keys.
Poisson distribution. Distribution generated by the Poisson function,
which can be used to approximatethe distribution of records among
addresses if the distribution is random. A particular Poisson distribu-
tion depends on theratio of the numberofrecords to the number of
available addresses. A particular instance of the Poisson function, p(x),

https://hemanthrajhemu.github.io
314 Chapter 11 Hashing

gives the proportion of addresses that will have x keys assigned to

them. See better-than-random,
Prime division. Division of a number by a prime numberanduseofthe
remainder as an address. If the address size is taken to be a prime
numberp, a large number can be transformed into a valid address by
dividing it by p. In hashing, division by primes is often preferred to
division by nonprimesbecause primes tend to produce more random
remainders. .
‘Progressive overflow. An overflow handling techniquein whichcollisions
are resolvedby storing a record in the next available addressafterits
home address. Progressive overflowis not the mostefficient overflow
handling technique, butit is one of the simplest and is adequate for
many applications. |
Randomize. To.produce a number(e.g., by hashing) that appears to be
random.

M
CO
Synonyms. Two or moredifferent keys that hash to the same address.
Wheneachfile address can hold only one record, synonymsalways
P.

result in collisions. If buckets are used, several records whosekeysare

O

synonyms maybe stored withoutcollisions,

LO

Tombstone.A special marker placed in the key field of a record to mark it

TU

as no longer valid. The use of tombstones solves two problemsassoci-

.V

ated with the deletion of records: the freed space does not break a
W

sequential search for a record, and the freed spaceis easily recognized
W

as available and maybe reclaimedfor later additions.

W

Uniform. A distribution in which records are spread out evenly among

addresses. Algorithms that produce uniform distributions are better.
than randomizingalgorithms in that they tend to avoid the numbers
of collisions that would occur with a randomizingalgorithm.

FURTHER READINGS
There are a numberof goodsurveys of hashing andissues related to hash-
ing generally, including Knuth (1998), Severance (1974), Maurer (1975),
and Sorenson, Tremblay, and Deutscher (1978). Textbooks concerned with
: file design generally contain substantial amounts of material on hashing,
and they often provide extensive references for further study. Loomis
(1989) also covers hashing generally, with additional emphasis on pro-

https://hemanthrajhemu.github.io
‘Exercises 515

gramming for hashed files in COBOL. Cormen, Leiserson and Rivest

(1990), Standish (1995), Shaffer (1997), and Wiederhold (1983) will be
useful to practitioners interested in analyses of trade-offs among the basic
hashing methods.
_ Oneof the applications of hashing that has stimulated a great deal of
interest recently is the developmentofspelling checkers. Because of special
characteristics of spelling checkers, the types of hashing involved are quite
different from the approaches we describe in this text. Papers by Bentley
(1985) and Dodds (1982) provide entry into the literature on this topic.
(See also exercise 14.)

[EXERCISES
1. Use the function hash(KEY, MAXAD) described in the text to

M
CO
answerthe following questions.
a. Whatis the value of hash("Jacobs", 101)?
P.

b. Find two different words of more than four characters that are
O
LO

synonyms.
c. It is assumed in the text that the function hash does not need to
TU

generate an integer greater than 19 937. This could present a prob-

.V

lem if we have a file with addresses larger than 19 937. Suggest some
W

ways to get aroundthis problem.

W

2. In understanding hashing,it is important to understand the relation-

W

ships betweenthe size of the available memory, the numberof keys to

be hashed,the rangeof possible keys, and the nature of the keys. Let us
give namesto these quantities, as follows:
: Mz=the numberof memoryspaces available (each capable of hold-
ing one record);
r = the numberofrecordsto be stored in the memoryspaces; n =
the numberof unique home addresses produced by hashingthe r
record keys; and
+ K=a key, which may be any combination of exactly five uppercase
characters.
Suppose h(K) is a hash function that generates addresses between 0
and M-1.

https://hemanthrajhemu.github.io
516 Chapter 11 Hashing

a. How many unique keys are possible? (Hint: If K were one upper-
case letter rather than five, there would be 26 possible unique
keys.)
b. How are nand r related?
c. How are rand M related?
d. If the function h were a minimum perfect hashing function, how
would x, 7, and M be related?
The following table shows distributions of keys resulting from three
different hash functions on a file with 6000 records and 6000 ad-
dresses.
Function A Function B Function C
d(0) 0.71 0.25 ‘0.40
d(1) 0.05 0.50 0.36
d{2) 0.05 0.25 0.15

M
d(3) 0.05 0,00CO 0.05
d(4) 0.05 0.00 0.02
d(5) 0.04 0.00. 0.01
P.

d(6) 0.05 0.00 0.01

O
LO

d(7* 0,00 0.00 0.00

a. Which of the three functions(if any) generates a distribution of
TU

recordsthat is approximately random?

.V

b. Which generates a distribution that is nearest-to uniform?

W

c. Which (if any) generates a distribution that is worse than random?

W

d. Which function should be chosen?

W

Thereis a surprising mathematical result called the birthday paradox

that says that if there are more than 23 people in a room, thereis a
better than 50-50 chance that two of them have the same birthday.
Howis the birthday paradoxillustrative of a major problem associat-
ed with hashing?
Supposethat 10 000 addresses are allocated to hold 8000 records in a
randomly hashed file and that each address can hold onerecord.
Compute the following values:
a. The packing density forthefile;
b. The expected number of addresses with no records assigned to
them by the hash function;
c. The expected numberof addresses with one record assigned (no
synonyms);

https://hemanthrajhemu.github.io
Exercises 517

-d. The expected number of addresses with one record plus one or
more synonyms;
e. The expected number of overflow records; and
f. The expected percentage of overflow records.
6. Considerthe file described in the preceding exercise. What is the
expected numberof overflow recordsif the 10 000 locations are reor-
ganized as
a. 5000 two-record buckets; and
b. 1000 ten-record buckets?
7. Make table showing Poisson function values for r/N= 0.1, 0.5, 0.8, 1,
2, 5, and 11. Examine the table and discuss any features andpatterns
that provide useful information about hashing.
8. Thereis an overflow handling techniquecalled count-key progressive
overflow (Bradley, 1982) that works on block-addressable disks as

M
follows. Instead of generatinga relative record numberfrom key,the.
CO
hash function generates an address consisting of three values: a cylin-
der, a track, and a block number. The corresponding three numbers
P.
O

constitute the home address of the record. .

LO

Since block-organized drives (see Chapter 3) can often scan a

track to find a record with a givenkey,there is no need to load a block
TU

into memoryto find out whetherit contains a particular record. The

.V

I/O processorcan direct the disk drive to search a track for the desired
W

record. [t can even direct the disk to search for an empty record slot if
W

a record is not foundin its homeposition, effectively implementing:

W

progressive overflow.
a. Whatis it aboutthis technique that makes it superior to progressive
overflow techniques that might be implemented on sector-orga-
nized drives?
b. The main disadvantage of this techniqueis that it can be used only
with a bucketsize of 1. Whyis this the case, and whyis it a disad-
vantage?
9. In discussing implementation issues, we suggest initializing the data
file by creating real records that are marked empty before loading the
file with data. There are some good reasonsfor doing this. However,
there might be some reasons not to do it this way. For example,
suppose you want a hash file with a very low packing density and
cannotafford to have the unused space allocated. How might a file
managementsystem be designed to work with a very large Jogicalfile
but allocate space only for those blocks in the file that contain data?

https://hemanthrajhemu.github.io
518 Chapter 11 Hashing

10. This exercise (inspired by an example in Wiederhold, 1983, p. 136)

concerns the problem of deterioration. A numberof additions and
deletions are to be madeto file. Tombstonesare to be used where
necessary to preserve search paths to overflow records.
a. Show whatthe file lookslike after the following operations, and
compute the average search length.

Operation Home Address

Add Alan 0°
Add Bates 2
Add Cole 4
Add Dean 0°
Add Evans ]
Del Bates
Del Cole

M
Add Finch CO 0
Add Gates 2
P.

Del Alan
O

Add Hart 3
LO

- How has the use of tombstones caused the file to deteriorate?

TU

- What wouldbe theeffect of reloading the remaining itemsin the

file in the order Dean, Evans, Finch, Gates, Hart?
.V
W

b. What would be the effect of reloading the remaining items using

W

two-pass loading?
W

1}. Suppose you havea file in which 20 percent of the records account for
80 percent of the accesses and that you wantto store,thefile with a
packing density of 0 and a bucket size of 5. Whenthefile is loaded,
you load the active 20 percentof the recordsfirst. After the active 20
percent of the records are loaded and before the other records are
loaded, whatis the packing densityof the partially filled file? Using
this packing density, compute the percentage of the active 20 percent
that would be overflow records. Commentontheresults.
“12. In our computations of average search lengths, we consider only the
timesit takes for successful searches. If our hashedfile were to be used
in such a way that searches were often madefor items that are not in
the file, it would be useful to havestatistics on average search length
for an unsuccessful search. If a large percentage of searches to a hashed
file are unsuccessful, howdo you expectthis to affect overall perfor-
mance if overflow is handled by

https://hemanthrajhemu.github.io
Exercises 519

a. Progressive overflow; or
b. Chaining to a separate overflow area?
(See Knuth, 1973b, pp. 535-539 for a treatmentofthese differences.)
13. Although hashed files are not generally designed to support access to
records in any sorted order, there may be times when batches of
transactions need to be performed on a hashed datafile. If the data
file is sorted (rather than hashed), these transactions are normally
carried out by somesort of cosequential process, which means that
the transaction file also has to be sorted. If the datafile is hashed, the
transaction file might also be presorted, but on the basis of the home
addresses of its records rather than some more “natural”criterion.
Suppose you havea file whose recordsare usually accessed direct-
ly but is periodically updated from a transactionfile. List the factors
you would have to consider in deciding between using an indexed
sequential organization and hashing. (See Hanson, 1982, pp.

M
280-285, for a discussion of these issues.)
CO
14, Weassume throughoutthis chapter that a hashing program should
P.

be able to tell correctly whether a given key is located at a certain

O

address. If this were not so, there would be times when we would
LO

assume that a record exists when in factit does not, a seeminglydisas-

TU

trous result. But consider what Doug Mcllroy did in 1978 when he
was designing a spelling checker program. He found that by letting
.V

his program allow one outof every four thousand misspelled words
W

to sneakby as valid (and using a few othertricks), he could fit a 75

W

000-wordspelling dictionary into 64 kilobytes of memory, thereby

W

improving performance enormously.

Mcllroy was willing to tolerate one undetected misspelled word
out of every four thousand because he observed thatdrafts of papers
rarely contained more than twenty errors, so one couldexpect at
most one out of every two hundred runs of the program to fail to
detect a misspelled word. Can you think of some other cases in
which it might be reasonable to report that a key exists when in fact
it does not?
Jon Bentley (1985) provides an excellent account of McIlroy’s
program, plus several insights on the process of solving problems of
this nature. D, J. Dodds (1982) discusses this general approach to
hashing, called check-hashing. Read Bentley’s and Dodds’s articles and
report on them to your-class. Perhaps they will inspire you to write a
spelling checker.

https://hemanthrajhemu.github.io
520 Chapter 11 Hashing

PROGRAMMING
15. Implement and test a version of the. function hash.
16. Create a hashed file with one record for every city in California. The
key in each recordis to be the name of the correspondingcity. (For.
the purposes of this exercise, there need be no fields other than the
key field.) Begin by creating a sortedlist of the namesofall of the
cities and townsin California. (If time or spaceis limited, just make a
list of namesstarting with theletter S.)
a. Examine the sorted list. What patterns do you notice that might
affect your choice of a hash function?
Implement the function hash in such a way.that you canalter the
numberof characters that are folded. Assuming a packing density
of 1, hash the entire file several times, each timefolding a different

M
number of characters and producing the following statistics for
CO
each run:
* The numberofcollisions; and
P.
O

* The numberof addresses assigned 0, 1, 2,..., 10, and '10-or-more

LO

records.
Discussthe results of your experimentin termsoftheeffects of
TU

folding different numbers of characters and how they compare

.V

with the results you might expect from a random distribution.

W

Implement and test one or more of the other hashing methods

W

described in the text, or use a method of your own invention.

W

17, Using a set of keys, such as the names of California towns, do the
following:
a. Write and test a program for loading the keys into three different
hashfiles using bucketsizes of 1, 2, and 5, respectively, and a pack-
ing density of 0.8. Use progressive overflow for handlingcollisions.
. Have your program maintain statistics on the average search
length, the maximum search length, and the percentage of records
that are overflowrecords.
Assuminga Poisson distribution, compare yourresults with the
expected values for average search length and the percentage of
records that are overflow records.
18. Repeat exercise 17, but use double hashing to handle overflow.

https://hemanthrajhemu.github.io
Programming Exercises 521

19. Repeat exercise 17, but handle overflow using chained overflow into a
separate overflow area. Assume that the packing density is the ratio of
numberof keys to available home addresses.
20. Write a program that can perform insertions and deletionsin thefile
created in the previous problem using a bucket size of 5. Have the
program keep runningsiatistics on average search length. (You might
also implement a mechanismto indicate whensearch length has dete-
riorated to a point wherethefile should be reorganized.) Discuss in
detail the issues you have to confrontin deciding how to handle inser-
tions and deletions.

M
CO
P.
O
LO
TU
.V
W
W
W

https://hemanthrajhemu.github.io
 M
CO
P.
O
LO
TU
.V
W
W
W

https://hemanthrajhemu.github.io
 CHAPTER

Extendible
Hashing

M
CO
P.
O
LO
TU
.V
W

CHAPTER OBJECTIVES
W
W

Describethe problem solved by extendible hashing and related

approaches.
Explain how extendible hashing works; show how it combines tries
with conventional, static hashing.
Use the buffer, file, and index classes of previous chapters to
implement extendible hashing, including deletion,
Review studies of extendible hashing performance.
“*
+

Examine alternative approachesto the same problem, including

ate
+,

dynamic hashing, linear hashing, and hashing schemesthat contro!

splitting by allowing for overflow buckets.

523
https://hemanthrajhemu.github.io
524 Chapter 12 Extendible Hashing

CHAPTER OUTLINE
12.1 Introduction
12.2 How Extendible Hashing Works
12.2.1 Tries
12.2.2 Turning the Trie into a Directory
12.2.3 Splitting to Handle Overflow
12.3. Implementation
12.3.1 Creating the Addresses
12.3.2 Classes for Representing Bucket and Directory Objects
12.3.3 Bucket and Directory Operations
12.3.4 Implementation Summary
12.4 Deletion
12.4.1 Overview of the Deletion Process
12.4.2 A Procedure for Finding Buddy Buckets
12.4.3 Collapsing the Directory-

M
12.4.4 Implementing the Deletion Operations
CO
12.4.5 Summary of the Deletion Operation
12.5 Extendible Hashing Performance
P.

12.5.1 Space Utilization for Buckets

O

12.5.2 Space Utilization for the Directory

LO

12.6 Alternative Approaches

TU

12.6.1 Dynamic Hashing

12.6.2 Linear Hashing
.V

12.6.3 Approaches to Controlling Splitting

W
W
W

12.1 Introduction

In Chapter 9 we began with a historical review of the work that led up to

B-trees. B-trees are such an effective solution to the problemsthat stimu-
lated their development that it is easy to wonder if there is any more
important thinking to be done aboutfile structures. Work on extendible
hashing during the late 1970s and early 1980s showsthat the answer to
that question is yes. This chaptertells the story of that work and describes
someofthefile structures that emerge from it.
B-trees do for secondary storage what AVLtrees do for storage in
memory: they provide a way of using tree structures that works well with
dynamic data. By dynamic we mean that records are added anddeleted
from the data set. The key feature of both AVLtrees and B-treesis that
they are self-adjusting structures that include mechanisms to maintain

https://hemanthrajhemu.github.io
 How Extendible Hashing Works 525

themselves. As we addand delete records, the tree structures use limited,

local restructuring to ensure that the additions and deletions do_not
degrade performance beyond some predeterminedlevel,
Robust,self-adjusting data andfile structures arecritically important
to data storage and retrieval. Judging from thehistorical record, they are
also hard to develop. It was not until 1963 that Adel’son-Vel’skii and
Landis developeda self-adjusting structure for tree storage in memory,
and it took another decade of work before computer scientists found, in B-
trees, a dynamictree structure that works well on secondarystorage.
B-trees provide O(log,N) access to the keys in a file. Hashing, when
there is no overflow, provides access to a record with a single seek. But as a
file growslarger, looking for records that overflow their buckets degrades
performance. For dynamic files that undergo a lot of growth,the perfor-
mance of a static hashing system such as we described in Chapter 11 is
typically worse than the performance of a B-tree. So, by the late 1970s,

M
after the initial burst of research and design work revolving around B-trees
CO
was over, a number of researchers began to work on finding ways to modi-
fy hashing sothat it, too, could be self-adjusting as files grow andshrink.
P.

As often happens when a numberof groups are working on the same

O
LO

problem, several different, yet essentially similar, approaches emerged to

extend hashing to dynamic files. We begin our discussion of the problem
TU

by looking closely at the approachcalled “extendible hashing” described by

.V

Fagin, Nievergelt, Pippenger, and Strong (1979). Later in this chapter we

W

compare this approach with others that emerged morerecently,

W
W

12.2. How Extendible Hashing Works

12.2.1 Tries

The key idea behind extendible hashing is to combine conventional hash-

ing with another retrieval approach called the trie. (The word trie is
pronounced so that it rhymes with sky.) Tries are also sometimesreferred
to as radix searching because the branching factor of the searchtreeis
equal to the numberofalternative symbols (the radix of the alphabet) that
can occur in each position of the key. A few exampleswill illustrate how
this works.
Suppose we want to build trie that stores the keys able, abrahms,
adams, anderson, andrews, and baird. A schematic form of the trie Is
shownin Fig. 12.1. As you cansee, the searching proceedsletter by letter

https://hemanthrajhemu.github.io
526 Chapter 12 Extendible Hashing

able

abrahms

anderson

andrews

M
Figure 12.1 Radix 26 trie that indexes names according to the letters of the
CO
alphabet.
P.
O

throughthe key. Because there are twenty-six symbols in the alphabet,the

LO

potential branching factor at every node of the search is twenty-six. If we

TU

used the digits 0-9 as our search alphabet rather than the letters a—z, the
radix of the search would be reduced to 10. A search tree using digits
.V

might looklike the one shownin Fig. 12.2.

W

Notice that in searching a trie we sometimesuse only a portion ofthe

W

key. We use moreof thekey as we need more information to complete the

W

search. This use-more-as-we-need-more capability is fundamental-to the

structure of extendible hashing. |

12.2.2 Turning the Trie into a Directory

We use tries with a radix of 2 in our approach to extendible hashing:
search decisions are madeon a bit-by-bit basis. Furthermore,since we are
retrieving from secondarystorage, we will not work in terms of individual
keys but in terms of buckets containing keys, just as in conventional hash-
ing. Suppose we have bucket A containing keys that, when hashed, have
hash addresses that begin with the bits 0J. Bucket B contains keys with
hash addresses beginning with 10, and bucket C contains keys with ad-
dresses that start with 11. Figure 12.3 showsa trie that allowsusto retrieve
these buckets.

https://hemanthrajhemu.github.io
How Extendible Hashing Works 527

7263

7268

M
CO
Figure 12.2 Radix 10 trie that indexes numbers according to the digits they
contain. .
P.
O
LO

How should werepresent the trie? If we represent it as a tree struc-

TU

ture, we are forced to do a numberof comparisons as we descend thetree.

Even worse, if the trie becomesso large that it, too, is stored on disk, we
.V

are faced once again with all of the problemsassociated with storing trees
W

on disk. We might as well go back to B-trees and forget about extendible

W

hashing.
W

So, rather than representingthetrie as a tree, weflatten it into an array

of contiguous records, forming a directory of hash addresses and pointers
to the corresponding buckets. Thefirst step in turning a tree into an array
involves extendingit so it is a complete binary tree with all of its leaves at

Figure 12.3 Radix 2 trie

that provides an index to
buckets.

https://hemanthrajhemu.github.io
528 Chapter 12 Extendible Hashing

00
: :
01 ‘

°
n
(a) (b)

Figure 12.4 Thetrie from Fig. 12.3 transformed first into a complete binary tree, then
flattened into a directory to the buckets. .

M
the same level as’shownin Fig. 12.4(a). Even thoughthe initial 0 is enough
CO
to select bucket A, the new form ofthe tree also uses the secondaddressbit
so bothalternatives lead to the same bucket.Once we have extendedthe
P.

tree this way, we can collapseit into the directory structure shownin Fig.
O

12.4(b). Now we havea structure that provides the kind of direct access
LO

associated with hashing: given an address beginningwith the bits 10, the
TU

10,'h directory entry gives us a pointer to the associated bucket.

.V
W

12.2.3 Splitting to Handle Overflow

W

A keyissue in any hashing system is what happens when a bucket overflows.

W

The goal in an extendible hashing system is to find a way to increase the

address space in response to overflow rather than respond by creating long
sequencesof overflow records and buckets that have to be searchedlinearly.
Supposeweinsert records that cause bucket A in.Fig. 12.4(b) to over-
flow. In this case the solution is simple: since addresses beginning with 00
and OJ are mixed together in bucket A, we can split bucket A by puttingall
the 01 addresses in a new bucketD, while keeping only the 00 addresses in
A. Put another way, we already have 2 bits of address information but are
throwing 1 away as we access bucket A. So, now that bucket A is overflow-
ing, we must use the full 2 bits to divide the addresses between two buck-
ets. We do not need to extend the address space; we simply makefull use of
the address information that we already have. Figure 12.5 showsthe direc-
tory and buckets after the split.

https://hemanthrajhemu.github.io
How Extendible Hashing Works 529

Figure 12.5 The directory

from Fig. 12.4(b) after bucket
A overflows. .

Let’s consider a more complex case. Starting once again with the
directory and buckets in Fig. 12.4(b), suppose that bucket B overflows.
How do wesplit bucket B and where do we attach the new bucket after
the split? Unlike our previous example, wedo not have additional, unused.

M
bits of address space that we can press into duty as wesplit the bucket. We
CO
now need to use 3 bits of the hash address in order to divide up the
records that hash to bucket B. The trie iilustrated in Fig. 12.6(a) makes
P.

the distinctions required to completethe split. Figure 12.6(b) shows what

O
LO

this trie looks like once it is extended into a completely full binary tree
with all leaves at the same level, and Fig. 12.6(c) shows the collapsed,
TU

directory form of:thetrie.

.V

By building on thetrie’s ability to extend the amount of information

used in a search, we have doubled the size of our address space (and, there-
W

fore, of our directory), extendingit from 2? to 23 cells. This ability to grow

W

(or shrink) the address space gracefully is what extendible hashingis all
W

about. .
We have been concentrating on the contribution that tries make to
extendible hashing; one might well ask where the hashing comesinto play.
Whynot just use the tries on the bits in the key, splitting buckets and
extending the address space as necessary? The answer to this question
grows out of hashing’s most fundamental characteristic: a good hash func-
tion produces a nearly uniform distribution of keys across an address
space. Notice that the trie shown in Fig. 12.6 is poorly balanced, resulting
in a directory that is twice as big as it needs to be. If we had an uneven
distribution of addresses that placed even more records in buckets B and
D without using other parts of the addressspace, the situation would get
even worse. By using a good hashfunction tocreate addresses with a near-
ly uniform distribution, we avoid this problem,

https://hemanthrajhemu.github.io
530 Chapter 12 Extendible Hashing ~N

(a}

000

001 ‘a
sO)
oa)

M
010
CO /

011
P.
O
LO
TU

10}
C2)
.V

ne | cs :
W
W
W

My

(b) (c)

Figure 12.6 The results of an overflow of bucket B in Fig. 12.4(b), represented first as a
trie, then as a complete binary tree, and finally as a directory.

12.3. Implementation

12.3.1 Creating the Addresses

Nowthatwe have a high-level overviewof how extendible hashing works,
let’s look at an object-oriented implementation. Appendix J contains the

https://hemanthrajhemu.github.io
 Implementation 531

int Hash (char* key)

{
int sum = Q;
int len = strlen(key);
if (len % 2 == 1) len ++; // make len even
// for an odd length, use the trailing '\0' as part of key
for (int j = 0; 3 < len; j+=2)
sum = (sum + 100 * key[j] + key[j+1])) % 19937;
return sum;
}

Figure 12.7 Function Hash (key) returns an integer hash value for key for a 15-bit

full class definitions and method bodies for extendible hashing. The place
to start our discussion of the implementation is with the functions that

M
create the addresses,since the notion of an extendible address underliesall
CO
other extendible hashing operations.
P.

The function Hash given in Fig. 12.7, and file hash. cpp of Appen-
O

dix J, is a simple variation on the fold-and-add hashing algorithm weused

LO

in Chapter 11. The only difference is that we do not conclude the opera-
TU

tion by returning the remainder of the folded address divided by the

address space. We don’t need to do that, since in extendible hashing we
.V

don’t have a fixed address space, instead we use as much of the address as
W

we need.The division that we perform in this function, when wetakethe

W

sum of the folded character values modulo 19 937, is to make sure that the
W

character summation stays within the range of a signed 16-bit integer. For
machinesthat use 32-bit integers, we could divide by a larger number and
create an even larger initial address.
Because extendible hashing uses more bits of the hashed address as
they are needed to distinguish between buckets, we need a function
MakeAddress that extracts just a portion of the full hashed address. We
also use MakeAddress to reverse the order of the bits in the hashed
address, making the lowest-orderbit of the hash address the highest-order
bit of the value used in extendible hashing. To see whythis reversal ofbit
order is desirable, look at Fig. 12.8, which is a set of keys and binary hash
addresses produced by our hash function. Even a quick scan of these
addresses reveals that the distribution of the least significant bits of these
integer values tends to have more variation than the high-orderbits. This
isbecause many of the addresses do not make use of the upper reaches of
our address space; the high-order bits often turn out to be 0.

https://hemanthrajhemu.github.io
532 Chapter 12 Extendible Hashing

bill 0000 0011 0110 1100

lee 0000 0100 0010 1000.
pauline 0000 1111 0110 0101
alan 0100 1100 1010 0010
julie 0010 1110 G000 1001
mike 0000 0111 0100 1101
elizabeth 0010 1100 0110 1010
mark 0000 1010 0000 0111

Figure 12.8 Output from the hash function for a number of keys.

By reversing the bit order, working from rightto left, we take advan-
tage of the greater variability of low-order bit values. For example, given a
4-bit address space, we want to avoid having the addresses of bill, lee, and

M
pauline turn out to be 0000, 0000, and 0000. If we work from rightto left,
CO
Starting with the low-order bit in each address, we get 0011 for bill, 0001
for lee, and 1010 for pauline, which is a much moreusefulresult.
P.

Function MakeAddress, givenin Fig. 12.9 and filehash.cpp of

O
LO

Appendix J, accomplishes this bit extraction and reversal. The depth

argument tells the function the number of addressbits to return.
TU
.V
W

int MakeAddress (char * key, int depth)

W

{
W

int retval = 0;
int hashVal = Hash(key);
// reverse the bits
for (int j = 0; j < depth; j++)
{
retval = retval << i;
int lowbit = hashVal & 1;
retval = retval | lowbit;
hashVal = hashVal >> 1;
}
return retval;
}

Figure 12,9 Function MakeAddress (key, depth) gets a hashed address,

reverses the order of the bits, and returns an addressof depthbits.

https://hemanthrajhemu.github.io
Implementation 533

12.3.2 Classes for Representing Bucket

and Directory Objects.
Our extendible hashing scheme consists of a set of buckets stored in file
and a directory that references them. Each bucketis a record that contains
a particular set of keys and information associated with the keys. A direc-
tory is primarily an array containing the record addresses of the buckets.
We have chosen to use a data record reference for the information
associated with the keys in this implementation. Thatis, the set of buckets
forms an indexof the file of actual data records. Thisis certainly not the
only way to configure the bucket file. We have left it as an exercise to
extend the bucket class so that the associated information could betherest
of the data record.In this case, the bucket file is the data file. For the rest of
the discussion of the implementation of extended hashing,we will treat
the buckets as sets of key-reference pairs.
The basic operations on buckets are exactly the same.as those of index

M
records: add a key-reference pair toa bucket, search for a key andreturnits
CO
reference, and removea key. Hence, we have chosen to make class Bucket
P.

a derived class of the class Text Index from Chapter 5 and Appendix F.
O

The definition of class Bucket is given in Fig. 12.10 and file bucket .h
LO

in Appendix J. These bucket records are stored in a file; we retrieve them as

TU

necessary. Each Bucket is connected to a directory and can be accessed

only in that context. This access restriction is enforced by making the
.V

membersof the class protected so that no outside accessis allowed, then

W

granting special access rights to class Directory by including the

W

friend class Directorystatement. Making class Directory a

W

friend of Bucket allows methodsof class Directoryto access all of

the private and protected members of class Bucket. The included meth-
ods of class Bucket will be explained below.
Class Directory is given in Fig. 12.11 and in file direct .h of
Appendix J. Each cell in the directory consists of the file address of a
Bucket record. Because we use direct accessto find directory cells, we
implement.the directory as an arrayof these cells in memory. The address
values returned by MakeAddress are treated as subscriptsfor this array,
ranging from 0 to one less than the number of cells in the directory.
Additional members are included to supportthe file operations required
to store the directory andtherelated buckets. You may notice a striking
similarity with classes IndexedFile (Chapter 5) and BTree (Chapter
9). Each of these classes supports open, create, and close operations as well
as insert, search, and remove. Objects of class BufferFile are used to
provide theI/O operations.

https://hemanthrajhemu.github.io
534 Chapter 12 Extendibie Hashing

class Bucket: protected TextIndex

{protected:
// there are no public members,
// access to Bucket members is only through class Directory
Bucket (Directory & dir, int maxKeys = defaultMaxKeys) ;
int Insert (char * key, int recAddr);
int Remove (char * key);
Bucket * Split ();// split the bucket and redistribute the keys
int NewRange (int & newStart, int & newEnd) ;
// calculate the range of a new (split) bucket
int Redistribute (Bucket & newBucket); // redistribute keys
int FindBuddy ();// find the bucket that is the buddy of this
int TryCombine (); // attempt to combine buckets
int Combine (Bucket * buddy, int buddyIndex); //combine buckets
int Depth; //number of bits used ‘in common’ by keys in bucket
Directory & Dir; // directory that contains the bucket

M
int BucketAddr; // address in file CO
friend class Directory;
friend class BucketBuffer;
P.
O

};
LO

Figure 12.10 Main members of class Bucket.

TU
.V
W

In order to use a Directory object, it must be constructed and then

W

attached to file for the directory and one for the buckets. Fig 12.12(page
W

536) is a sample test program, tsthash. cpp. Thetwostepsforinitial-

ization are the declaration of the Directory object, and the call to
method Create, that creates the twofiles and theinitial empty bucket.
The program proceeds to insert a set of key-reference pairs, Notice that the
reference values have no particularsignificance in this program.
The constructor and method Create are given in Fig. 12.13 (page
537). The constructor creates all of the objects that support the I/O oper-
ations: a buffer anda file for the directory and a buffer and file for the
buckets. The directory is stored in memory while the directory is open..
The Open method readsthe directory from thefile and the Close writes.
it back intothefile,

https://hemanthrajhemu.github.io
 Implementation 535

class Directory
{public:
Directory (int maxBucketKeys = -1);
~Directory ();
int Open (char * name);
int Create (char *- name);
int Close ();
int Insert (char * key, int recAddr};
int Delete (char * key, int recAddr = -l);
int Search (char * key); // return RecAddr for key
ostream’& Print (ostream & stream);
‘protected:
int Depth; // depth of directory
int NumCells; // number of cells, = 2**Depth
int * BucketAddr; // array of bucket addresses

M
CO
// protected methods
int DoubleSize (); // double the size of the directory
P.

int Collapse (); // collapse, halve the size

O
LO

int InsertBucket (int bucketAddr, int first, int last);

int Find (char * key); // return BucketAddr for key
TU

int StoreBucket (Buckét * bucket); ,

// update or append bucket in file
.V

int LoadBucket (Bucket * bucket, int bucketAddr) ;

W

// load bucket from file

W

// members to support directory and bucket files

W

int MaxBucketKeys;
BufferFile * DirectoryFile;
LengthFieldBuffer * DirectoryBuffer;
Bucket * CurrentBucket;// object to hold one bucket
BucketBuffer * theBucketBuffer;// buffer for buckets
BufferFile * BucketFile;
int Pack () const;
int Unpack ();
Bucket. * PrintBucket;// object to hold one bucket for printing
friend class Bucket;
};

Figure 12.11 Definition of class Directory.

https://hemanthrajhemu.github.io
536 . Chapter 12 -Extendible Hashing

main ()
{
int result;
Directory Dir (4);
result = Dir . Create ("hashfile"):
if (result == 0) {return 0;} // unable to create files
char * keys{]J={"bill", "lee", "pauline", "alan", "julie",
"mike", "elizabeth", “mark", “ann", "peter"?
"christina", "john", "charles", "mary", "“emily"};
const int numkeys = 15;
for (int i = 0; i<numkeys; i++)
{
result = Dir . Insert’ (keys[{i], 100 + i);
if (result == 0).
cout << “insert for "<<keys[i]<<" failed"<<endl;
Dir . Print (cout);

M
}
CO
return 1;
P.

}
O
LO

Figure 12.12 Test program tsthash.cpp inserts a sequence of key-reference pairs into
a directory.
TU
.V

Note that member Depth is directlyrelated to the size of the directo-

W

ry, since
W

2Depth = the numberof cells in the directory.

W

If.we are starting a new hashdirectory, the directory depthis 0, which

means that weare using no bits to distinguish between addresses;all the
keys go into the same bucket; no matter what their address. We get the
address of the initial, everything-goes-here bucket and assignit to the
single directorycell in this line from Directory: :Create:
BucketAddr[0) = StoreBucket (CurrentBucket) ;

The method StoreBucket appends the newbucket to the bucketfile

and returns its address.

12.3.3 Directory and Bucket Operations

Now that we have a way to open and closethe file, we are ready to add
records to the directory. The Insert, Search, andFind methodsare

https://hemanthrajhemu.github.io
 Implementation 537

Directory: :Directory (int maxBucketKeys)

{
Depth = 0; // depth of directory
NumCells = 1; // number of entries, = 2**Depth
BucketAddr = new int [NumCells]; // array of bucket addresses
// create I/O support objects
MaxBucketKeys = maxBucketKeys;
DirectoryBuffer = new LengthFieldBuffer; // default size
DirectoryFile = new BufferFile({*DirectoryBuffer) ;
CurrentBucket = new Bucket (*this, MaxBucketKeys) ;
theBucketBuffer = new BucketBuffer (MaxKeySize, MaxBucketKeys);
-BucketFile = new BufferFile (*theBucketBuffer} ;
PrintBucket = new Bucket (*this, MaxBucketKeys) ;
} ‘
int Directory: :Create (char * name)
{ // create the two files, create a single bucket

M
// and add it to the directory and the bucket file
CO
int result;
char * directoryName, * bucketName;
P.
O

makeNames (name, GirectoryName, bucketName);// create file names

LO

result.= DirectoryFile->Create(directoryName, ios: :inlios::out);

if (!result) return 0;
TU

result = BucketFile->Create(bucketName,ios::inlios::out);
.V

if ('result) return 0;
// store the empty bucket in the BucketFile; add to Directory
W

BucketAddr[0] = StoreBucket (CurrentBucket);

W

return result;
W

Figure 12.13 Constructor and method Createof class Directory.

shown in Fig. 12.14. The Insert method first searches for the key.
Search arranges for the CurrentBucket member to contain the
proper bucket for the key. If the key is not already in the bucket, then the
Bucket: : Insert method 1s called to performthe insertion. In method
Directory: : Search, as in most search functions we have seen, the
Find method determines where the key would be if it were in the struc-
ture. In this case, Find determines which bucket is associated with the
key. As noted previously, MakeAddress finds the array index of the
directory cell that contains the file address of the appropriate bucket.

https://hemanthrajhemu.github.io
538 Chapter 12 Extendible Hashing

int Directory::Insert (char * key, int recAddr)

{
int found = Search (key);
1£ (found != -1) return 0; // key already in directory
return CurrentBucket->Insert (key, recAddr);
}
int Directory::Search (char * key)
// return RecAddr for key, also put current bucket into variable
{ '
int bucketAddr = Find(key);
LoadBucket (CurrentBucket, bucketAddr) ;
‘return CurrentBucket->Search(key) ;
}
int Directory::Find (char * key)
// find BucketAddr associated with key
{ return BucketAddr[(MakeAddress (key, Depth) ];}

M
CO
Figure 12.14 Methods Insert, Search, and Find of class Directory.
P.
O
LO

int Bucket::Insert (char * key, int recAddr)

TU

{
if (NumKeys < MaxKeys)
.V

{
W

int result = TextIndex::Insert (key, recAddr) ;

W

Dir.StoreBucket (this);
W

return result;
}
else // bucket is full
{
Split ();
return Dir.insert (key, recAddr);
}
}

Figure 12.15 MethodInsert of class Bucket adds the key to the existing
bucketif thereis room. If the bucketis full,it splits it and then adds the key.

Method Bucket::Insert, given in Fig. 12.15 and in file

buffer.cpp of Appendix J, is called with a key-reference pair. If the
-bucket is not full, Insert simply calls TextIndex: : Insert to add

https://hemanthrajhemu.github.io
Implementation 539

the key-reference pair to the bucket and stores the bucket in thefile. A full
bucket, however, requires asplit, which is where things start to get inter-
esting. After the split is done, the Directory: :Insert is called
(recursively) to try again to insert the key-referencepair.
What we do when we split a bucket depends on the relationship
between the numberof address bits used in the bucket and the number
used in the directory as a whole. The two numbersare often not the same.
To see this, look at Fig. 12.6(a). The directory uses 3 bits to define its
address space (8 cells). The keys in bucket A are distinguished from keysin
other buckets by. having aninitial 0 bit. All the other bits in the hashed key
values in bucket A can be any value;it is only the first bit that matters.
Bucket A is using only J bit and has depth 1.
The keys in bucket C all share a common first 2 bits; they all begin
with 11. The keys in buckets B and D use 3 bits to establish their identi-
ties and, therefore, their bucket locations. If you look at Fig. 12.6(c), you

M
can see how using moreor fewer address bits changes the relationship
CO
between the directory and the bucket. Buckets that do not use as many
address bits as the directory have more than onedirectorycell pointing
P.

to them.
O

If we split one of the buckets that is using fewer address bits than the
LO

directory, and therefore is referenced from more than one directorycell,

TU

we can usehalf of the directorycells to point to the new bucket after the
split. Suppose, for example, that wesplit bucket A inFig. 12.6(c). Before
.V

the split only 1 bit, the initial 0, is'used to identify keys that belongin
W

bucket A. After thesplit, we use 2 bits. Keys starting with 00 (directorycells

W

000 and 001) go in bucket A; keys starting with 01 (directory cells 010 and
W

011) go in the new bucket. We do not have to expand thedirectory because

the directory already has the capacity to keep track of the additional
address information requiredfor the split.
If, on the other hand, wesplit a bucket that has the same address depth
as the directory, such as buckets B or D in Fig. 12.6(c), then there are no
additional directory cells that we can use to reference the newbucket.
Before we can split the bucket, we have to double thesize of thedirectory,
creating a new directory entry for every one that is currently there so we
can accommodate the new address information.
Figure 12.16 gives an implementation of method Split. First we
compare the numberofbits used for the directory with the number used
for the bucket to determine whether we need to doublethedirectory.If the
depthsare the same, we doublethe directory before proceeding.

https://hemanthrajhemu.github.io
540 Chapter 12 Extendible Hashing

Bucket * Bucket::Split ()
{// split this into two buckets, store the new bucket, and
// xveturn (memory) address of new bucket
int newStart, newEnd;
if (Depth == Dir.Depth)// no room to split this bucket
Dir .DoubleSize();// increase depth of directory
Bucket * newBucket = new Bucket (Dir, MaxKeys);
Dir.StoreBucket (newBucket}; // append to file
NewRange (newStart, newEnd); // determine directory addresses
Dir. InsertBucket (newBucket->BucketAddr, newStart, newEnd);
Depth ++; // increment depth of this
newBucket->Depth = Depth;
Redistribute (*newBucket); // move some keys into new bucket
Dir.StoreBucket (this);
Dir.StoreBucket (newBucket) ;
return newBucket;

M
}
CO
P.

Figure 12.16 MethodSplit of class Bucket divides keys between an existing bucket and
O

a new bucket. If necessary, it doubles the size of the directory to accommodate the new
LO

bucket.
TU
.V

Next we create the new bucket that we need for thesplit. Then wefind
W

the range of directory addresses that we will use for the new bucket. For
W

instance, when wesplit bucket A in Fig. 12.6(c), the range of directory

addresses for the new bucketis from 010 to 011. We attach the new buck-
W

et to the directory over this range, adjust the bucket address depth infor-
“mationin both bucketsto reflect the use of an additional address bit, then
redistribute the keys from the original bucket across the two buckets.
The most complicated operation supporting the Split method is
NewRange, which finds the rangeof directory cells that should point to
the new bucket instead of the old one after the split. It is given in Fig.
12.17. To see how it works, return, once again, to Fig. 12.6(c). Assume that
we need to split bucket A, putting some of the keys into a new bucketE.
Before thesplit, any address beginning with a 0 leadsto A. In other words,
the shared address of the keys in bucket is 0.
When wesplit bucket A we add another addressbit to the path leading
to the keys; addresses leading to bucket A nowshare an initial 00 while
those leading to Eshare an 01. So,the range of addresses for the new buck-
et is all directory addresses beginning with 01. Since the directory address-

https://hemanthrajhemu.github.io
‘Implementation 541

int Bucket::NewRange (int & newStart, int & newEnd)

{// make a range for the new split bucket
aint sharedAddr = MakeAddress(Keys[0), Depth);
int bitsToFill = Dir.Depth - (Depth + 1);
newStart = (sharedAddr << 1) | 1;
newEnd = newStart;
for {aint 5 = 0; j < bitsToFill: j++)
{
newStart = newStart << 1;
newEnd = (newEnd << 1} | 1;
}
return 1;
)

Figure 12.17 Method NewRange ofclass Bucketfinds the start and end

M
directory addresses for the new bucket by using information from the old
CO
bucket.
P.
O

es.use 3 bits, the new bucket is attached to the directorycells starting with
LO

010 and ending with 0171.

TU

Suppose that the directory used a 5-bit address instead of a 3-bit

.V

address. Then the range for the new bucket would start with 01000 and
end with 0111). This range coversall 5-bit addresses that share 01 as the
W

first 2 bits. The logic for finding the range of directory addresses for the
W

new bucket; then, starts by finding shared. address bits for the new bucket.
W

It then fills the address out with Os until we have the numberof bits used
in the directory. This is the start of the range. Filling the address out with
ls produces the end of the range.
The directory operations required to support Split are easy to
implement. They are given in Fig. 12.18. The first, Directory:
: DoubleSize, simply calculates the newdirectory size, allocates the
required memory, and writes the information from eachold directorycell
into two successive cells in the new directory.It finishes by freeing the old
space associated with member Buf ferAddrs, renaming the new space
as the Buf ferAdadrs, and increasing the Depth toreflect the fact that
the directory is nowusing an additional addressbit.
Method InsertBucket, used to attach a bucket address across a
range of directory cells, is simply a loop that works throughthe cells to
make the change.

https://hemanthrajhemu.github.io
542 Chapter 12 Extendible Hashing

int Directory: :DoubleSize ()

// Gouble the size of the directory
{
int newSize = 2 * NumCells;
int * newBucketAddr = new int [newSize];
for (int 1 = 0; 1 < NumCells; i++)
{// double the coverage of each bucket
newBucketAddr(2*i] = BucketAddr[i];
newBucketAddr[2*i+1] = BucketAddr[{i];
}
delete BucketAddr;// delete old space for cells
BucketAddr = newBucketAddr;
Depth ++;
NumCells = newSize;
return 1;
}

M
CO
int Directory::InsertBucket (int bucketAddr, int first, int last)
{
P.

for (int i = first; i <= last; i++)

O

BucketAddr[{ijJ = bucketAddr;
LO

return 1; .
TU

}
.V

Figure 12.18 Methods DoubleSize and Directory InsertBucket ofclass Directory.

W
W
W

12.3.4 Implementation Summary

Now that we have assembledall of the pieces necessary to add records to
an extendible hashing system,let’s see how the pieces work together.
The Insert method manages record addition. If the key already
exists, Insert returns immediately. Ifthe key does notexist, Insert
calls Bucket: : Insert, for the bucket into which the keyis to be added.
If Bucket::Insert finds that there is still room in the bucket, it
adds the key and the operation is complete. If the bucket is full,
Bucket: :Insert calls Split to handle the task of splitting the bucket.
The Split method starts by determining whether the directoryis
large enough to accommodate the new bucket.If the directory needs to be
larger, Split calls method Directory: :DoubleSizeto double the
directory size. Split thenallocates a new bucket, attachesit to the appro-
priate directory cells, and divides the keys between the two buckets.

https://hemanthrajhemu.github.io
 Deletion 543

When Bucket: : Insert regains control after Split has allocated a

newbucket,it calls Directory: :Insert to try to placethe key into the
new,revised directory structure. The Directory::Insert function,of
course, calls Bucket: : Insert again, recursively. This cycle continues
until there is a bucket that can accommodate the new key. A problem can
occur if there are many keys that have exactly the same hash address. The
process of double, split, and insert will never make room for the new key.

12.4 Deletion

12.4.1° Overview of the DeletionProcess

If extendible hashing is to be a truly dynamic system,like B-trees or AVL

trees, it must be able to shrink files gracefully as well as grow them. When

M
we delete a key, we need a wayto see if we can decrease thesize of the file
CO
system by combining buckets and,if possible, decreasing the size of the
P.

directory.
O

As with any dynamic system, the important question during deletion

LO

concernsthe definition of the triggering condition: When do we combine

TU

buckets? This question, in turn, leads us to ask, Which buckets can be

combined? For B-trees the answer involves determining whether nodes are
.V

. siblings. In extendible hashing we use a similar concept: buckets that are

W

buddy buckets.
W

Look again at the trie in Fig. 12.6(b). Which buckets could be com-
W

bined? Trying to combine anything with bucket A would mean collapsing

everything else in the trie first. Similarly, there is no single bucket that
could be combined with bucket C. But buckets B and D are in the same
configuration as buckets that have just split. They are ready to be
combined: they are buddy- buckets. We will take a closer look at finding
buddy buckets when we consider implementation of the deletion proce-
dure; for nowlet’s assume that we combine buckets B and D.
After combining buckets, we examine the directory to see-if we can
make changes there. Looking at the directory form of the trie in Fig.
12.6(c), we see that once we combine buckets Band D,directory cells 100
and 101 both point to the same bucket. In fact, each of the buckets has at
least a pair of directory cells pointing to it. In other words, none of the
buckets requires the depth of address information that is currently avail-
able in the-directory. That means that we can shrink the directory and
reduce the address spaceto halfits size.

https://hemanthrajhemu.github.io
544 Chapter 12 Extendible Hashing

Reducingthesize of the address space restores the directory and buck-

et structure to the arrangement shownin Fig. 12.4, before the additions
and splits that produced the structure in Fig, 12.6(c). Reduction consists of
collapsing each adjacent pair of directorycells into a single cell. Thisis
easy, because both cells in each pair point to the same bucket. Note that
this is nothing more than reversal of the directorysplitting procedure
that we use when weneed to add newdirectorycells.

12.4.2 A Procedure for Finding Buddy Buckets

Given this overview of how deletion works, we begin by focusing on
buddy buckets. Given a.bucket, how do wefind its buddy? Figure 12.19
contains the code for method Bucket: :FindBuddy. The method
works by checking to see whetherit is possible for there to be a buddy
bucket. Clearly, if the directory depth is 0, meaning that there is onlya

M
single bucket, there cannot be a buddy. CO
The next test compares the numberofbits used by the bucket with the
numberofbits used in the directory address space. A pair of buddy buck-
P.

ets is a set of buckets that are immediate descendantsof the same nodein
O
LO

the trie. Theyare, in fact, pairwise siblings resulting from a split. Going
back to Fig. 12.6(b), we see that asking whether the bucket usesall the
TU

address bits in the directory is another way of asking whetherthe bucketis

.V

at the lowest level of the trie. It is only when a bucketis at the outer edge
W

of the trie that it can have a single parent and a single buddy.
Once we determine that there is a buddy bucket, we needto findits
W

address. First we find the address used to find the bucket we haveat hand;
W

int Bucket: : FindBuddy ()

{// find the bucket that is paired with this
if (Dir.Depth == 0} return -1; // no buddy, empty directory

// unless bucket depth == directory depth, there is no single

// bucket to pair with
if (Depth < Dir.Depth) return -1;
int sharedAddress = MakeAddress(Keys[0], Depth);
// address of any key
return sharedAddress * 1; // exclusive or with low bit
}

Figure 12.19 Method FindBuddy ofclass Bucket returns a buddy bucket or —1 if none is
found,

https://hemanthrajhemu.github.io
 Deletion: 545

this is the shared address of the keys in the bucket. Since we knowthatthe
buddy bucket is the other bucket that was formed from a split, we know
that the buddy has the same address in all regards except for the lastbit.
Once again, this relationship is illustrated by buckets B and D in Fig.
12.6(b). So, to get the buddy address, weflip the last bit with an exclusive
or, We return directory address of the buddybucket.

12.4.3 Collapsing the Directory

The other important support function used to implementdeletion is the
function that handles collapsing the directory. Downsizing the directoryis
one of the principal potential benefits of deleting records. In our imple-
mentation we use one function to see whether downsizing is possible and,
if it is, to collapse the directory. |
Method Directory: :Collapse, given in Fig. 12.20, begins by

M
making sure that we are notat the lowerlimit of directorysize. By treating
CO
the special case of a directory with a single cell here, at the start of the
function, we simplify subsequent processing: with the exception of this
P.

case,all directory sizes are evenly divisible by 2.

O
LO

Thetestto see if the directory can be collapsed consists of examining

each pair of directorycells to see if they point to different buckets. As soon
TU
.V
W

int Directory: :Collapse (}

W

{// if collapse is possible, reduce size by half

W

(Depth == 0) return 0; // only 1 bucket

// look for buddies that are different, if found return
for (int i = 0; i < NumCells; 1 += 2)
if (BucketAddr[i] != BucketAddr[i+1]) return 0;
int newSize = NumCells / 2;
int * newAddrs = new int [newSize];
for (int j = 0; 3 < newSize; j++)
newAddrs{j) = BucketAddr[j*2];
delete BucketAddr;
BucketAddr = newAddrs;
Depth —-;
NumCells = newSize;
return 1; ,
}

Figure 12.20 Method Collapse of class Directory reduces the size of the directory,if
possible.

https://hemanthrajhemu.github.io
546 Chapter 12 Extendible Hashing

as we find such a pair, we know that we cannot collapse the directory and
the method returns. If we get all the way through the directory without
encountering such a pair, then we can collapse the directory.
The collapsing operation consists of allocating space for a new array
of bucket addresses that is half the size of the original and then copying
- the bucket references shared by each cell pair to a single cell in the new
directory.

12.4.4 Implementing the Deletion Operations

Nowthat we have an approach to thetwocritical support operationsfor
deletion, finding buddy buckets and collapsing the directory, we are ready
to construct the higher levels of the deletion operation.
The highest-level deletion operation, Directory: : Remove, is
very simple. We first try to find the key to be deleted. Ifwe cannotfindit,

M
we return failure; if we find it, we call Bucket : : Remove to remove the
CO
key from the bucket. We return the value reported back from that method.
Figure 12.21 gives the implementation of these two methods.
P.

Method Bucket : : Remove doesits work in two steps. Thefirst step,

O

removing the key from the bucket, is accomplished through the call to
LO

Text Index: :Remove, the base class Remove method. The second
TU
.V
W

int Directory::Remove (char * key)

W

{// remove the key and return its RecAddr

W

int bucketAddr = Find(key);

LoadBucket (CurrentBucket, bucketAddr) ;
return CurrentBucket -> Remove (key);
}
int Bucket::Remove (char * key}
{// remove the key, return its RecAddr
' int result = TextIndex::Remove (key);
if (!result) return 0; // key not in bucket
TryCombine (); // attempt to combine with buddy
// make the changes permanent
Dir.StoreBucket (this) ;
return 1;
} '

Figure 12.21 Remove methodsof classes Directory and Bucket.

https://hemanthrajhemu.github.io
 Deletion 547

step, which takes place only if a key is removed, consists of calling

TryCombineto see if deleting the key has decreased thesize of the buck-
et enoughto allow us to combineit withits buddy. |
Figure 12.22 shows the implementation of TryCombine and
Combine. Note that when we combine buckets, we reduce the address
depth associated with the bucket: combining bucketsmmeans that we use 1
less address bit to differentiate keys.

int Bucket: :TryCombine: ()

{// called after insert to combine buddies, if possible
int result;
int buddyIndex = FindBuddy ();
if (buddyIndex == -1) return 0;// no combination possible
// load buddy. bucket into memory
int buddyAddr- = Dir.BucketAddr [buddyIndex] ;

M
Bucket * buddyBucket = new Bucket (Dir, MaxKeys) ;
CO
Dir . LoadBucket (buddyBucket, buddyAddr);
P.

// if the sum of the sizes of the buckets is too big, return

O

if (NumKeys + buddyBucket->NumKkeys > MaxKeys) return 0;

LO

Combine (buddyBucket, buddyIndex}; .

result = Dir.Collapse (); // collapse the 2 buckets
TU

if (result) TryCombine(); //if collapse, may be able to combine

.V

return 1;
}
W

int Bucket::Combine (Bucket * buddy, int buddy Index)

W

{// combine this and buddy to make a single bucket

W

int result;
// move keys from buddy to this
for (int 1 = 0; i.< buddy->NumKeys; i++)
{// insert the key of the buddy into this
result = Insert (buddy->Keys[i] ,buddy->RecAddrs[i])
‘Lf (!result) return 0;// this should not happen
}
Depth ~ -;// reduce the depth of the bucket
Dir . RemoveBucket (buddyIndex, Depth);// delete buddy bucket
return 1;
}

‘Figure 12.22 Methods TryCombine and Combine ofclass Bucket. TryCombinetests to

see whether a bucket can be combined with its buddy. If the test succeeds, TryCombine
calls Combine to do-the combination.

https://hemanthrajhemu.github.io
o+90 Chapter 12 Extendible Hashing

After combiningthe buckets, we call Directory::Collapse()

to see if the decrease in the numberof buckets enables us to decrease the
size of the directory. If we do, in fact, collapse the directory, TryCombine
calls itself recursively. Collapsing the directory may have created a new
buddy for the bucket; it may be possible to do even more combination
and collapsing. Typically, this recursive combining and collapsing happens
only when the directory has a number of empty bucketsthat are await-
ing changes in the directory structure that finally produce a buddy to
combine with.

12.4.5 Summaryof the Deletion Operation

Deletion begins with a call to Directory: : Removethat passes the key
that is to be deleted. If the key cannot be found, there is nothing to delete.
If the key is found, the bucket containing the key is passed to

M
Bucket: : Remove. CO
The Bucket: :Remove method deletes the key, then passes the
bucket on to Directory: :TryCombine to seeif the smaller size of
P.

the bucket will now permit combination with a buddy bucket.

O
LO

TryCombinefirst checks to see if there is a buddy bucket. If not, we are

done. If there is a buddy,and if the sum of the keys in the bucket and its
TU

buddyis less than or equal to the size of a single bucket, we combine the.
.V

buckets.
W

The elimination of a bucket through combination might causethe

W

directory to collapse to halfits size. We investigate this possibility by call-

W

ing Directory: :Collapse. If collapsing succeeds, we mayhave a

new buddybucket, so TryCombine calls itself again, recursively.
File testdel.cpp in Appendix J opens the directory created by
testhash.cpp and proceeds to delete each element ofthe directory.
Using a debuggerto step through this program may help in understanding
the deletion process. .

12.5 Extendible Hashing Performance

Extendible hashingis an elegant solution to the problem of extending and

contracting the address space for a hash file as the file grows and shrinks.
How well doesit work? Asalways, the answerto this question must consid-
er the trade-off between time andspace.

https://hemanthrajhemu.github.io
Extendible Hashing Performance 549

The time dimension is easy to handle: if the directory for extendible

hashing can be kept in memory,a single accessis all that is ever required to
retrieve a record.If the directoryis so large that it must be paged in and
out of memory, two accesses may be necessary. The important pointis that
extendible hashing provides O(1) performance:since there is no overflow,
these access time values are truly independent ofthe size of thefile.
Questions about space utilization for extendible hashing are more
complicated than questions aboutaccess time. We need to be'concerned
about two uses of space: the space for the buckets and the space for the
directory.

12.5.1. Space Utilization for Buckets

In their original paper describing extendible hashing, Fagin, Nievergelt,
Pippenger, and Strong include analysis and simulation of extendible hash-

M
ing performance. Both the analysis and simulation show that the space
CO
utilization is strongly periodic, fluctuating between values of 0.53 and
0.94, The analysis portion of their papersuggests that for a given number
P.

of records r and a block size of b, the average numberof blocks N is

O
LO

approximated by the formula

r
TU

N=——— N
bin?
.V
W

Spaceutilization, or packing density,is defined as theratio of the actu-

W

al numberof records to the total numberofrecords that could bestored in

W

the allocated space:

Utilization = ——
bN
Substituting the approximation for N gives us:
Utilization = 1n 2 = 0.69
So, we expect average utilization of 69 percent. In Chapter 9, where we
looked at space utilization for B-trees, we found that simple B-trees tend
to have a utilization of about 67 percent, but this can be increased to more
than 85 percent byredistributing keys during insertion rather than just
splitting when a page is full. So, B-trecs tend to use less space than simple
extendible hashing, typically at a cost of requiring a few extra seeks.
The average spaceutilization for extendible hashing is onlypart of the
story; the other part relates to the periodic nature of the variations in

https://hemanthrajhemu.github.io
550 Chapter 12 Extendible Hashing ;

space utilization. It turns out that ifwe have keys with randomly distrib-
uted addresses, the buckets in the extendible hashing table tendto fill up at
about the sametime and therefore tend to split at the same time. This
explains the large fluctuations in space utilization. As the buckets fill up,
space utilization can reach past 90 percent. Thisis followed by a concen-
. trated series of splits that reduce the utilization to below 50 percent. As
these now nearly half-full bucketsfill up again, the cycle repeatsitself.

12.5.2 Space Utilization for the Directory

The directory used in extendible hashing grows by doublingits size. A
prudent designer setting out to implement an extendible hashing system
will want assurance that this doubling levels off for reasonable bucket
sizes, even. when the numberof keys is quite large. Just how large a direc-
tory should we expect to have, given an expected numberofkeys?

M
Flajolet (1983) addressed this question in a lengthy, carefully devel-
CO
oped paper that produces a numberofdifferent ways to estimate the diréc-
tory size. Table 12.1, which is taken from Flajolet’s paper, shows the
P.

expected value for the directory size for different numbers of keys and
O

different bucket sizes.

LO

Flajolet also provides the following formula for making roughesti-

TU

mates of the directory size for values that are not in this table. He notes
that this formula tends to overestimate directory size by a factorof 2 to 4.
.V
W

Estimated directory size = = (1+ 1/b)

W
W

Table 12.1 Expected directory size for a given bucket size b and total number of records fr

6 5 10 20 50 100 200
r | \

103 150K 0.30K 0.10K 0.00 K 0.00 K 0.00,K

108 25.60 K 4.80 K 1.70 K 0.50 K 0.20 K 0.00 K
105° 424.10K 68.20K 16.80 K 410K 2.00 K 1,00 K
108 6.90 M 1.02 M 0.26 M 62.50 K 16.80 K 18.10K
107 “112.11 M 12.64 M 2.25 M 0.52 M. 0.26 M 0.13 M

1K = 103, 1M 108,
From Flajolet, 1983,

https://hemanthrajhemu.github.io
 Alternative Approaches 551

12.6 Alternative Approaches

12.6.1 Dynamic Hashing

In 1978, before Fagin, Nievergelt, Pippenger, and Strong produced their
paperon extendible hashing, Larson published a paperdescribing a scheme
called dynamic hashing. Functionally, dynamic hashing and extendible
hashingare very similar. Both use a directory to track the addresses of the
buckets, and bothextend the directory through the useof tries.
The key difference between the approachesis that dynamic hashing,
like conventional, static hashing, starts with a hash function that covers an
address space of a fixed size. As buckets within that fixed address space
overflow, they split, forming the leaves of a trie that grows down from the
original address node. Eventually, after enough additions andsplitting, the

M
buckets are addressed through a forestof tries that have been seeded out of
the original static address space.
CO
Let’s look at an example. Figure 12.23(a) showsan initial address space
P.

of four and four buckets descending from the four addresses in the direc-
O

tory. In Fig. 12.23(b) we have split the bucket at address 4. We address the
LO

two buckets resulting from the split as 40 and 41. We change the shape of
TU

the directory node at address 4 from a square to a circle because it has

changed from an external node, referencing a bucket, to an internal node
.V

that points to two child nodes.

W

In Fig. 12.23(c) we split the bucket addressed by node 2, creating the

W

new external nodes 20 and 2]. Wealso split the bucket addressedby 41,
W

extending the trie downwardto include 410 and 411. Because the directo-
ry node 4J is nowan internal node rather than an externalone, it changes
from a square to a circle. As we continue to add keys and split buckets,
these directory tries continue to grow.
Finding a key in a dynamic hashing scheme.can involve the use of two
hash functions rather than just one. First, there is the hash function that
covers the original address space.If you find that the directory nodeis an
external node and therefore points to a bucket, the search is complete.
However,if the directory node is an internal node, then you need addi-
tional address information to guide you through the 1s and 0s that form
the trie. Larson suggests using a second hash function onthe key and using
_the result of this hashing as the seed for a random-numbergenerator that
produces a sequence ofIs and 0s for the key. This sequence describes the
path throughthetrie.

https://hemanthrajhemu.github.io
552 Chapter 12 Extendible Hashing

(a})-—— _—— 1 -——+ 2 b}——+4 3 F}——4 4) —~ Original address space

OCODOD oD
(b)— ——- ——- — + ) F—— SF 2 FS 8 — — — 4 Original address space

.| 40 41

CD OD CODED D
(c)- }- — — ——/f 9 }—— —— 4 3 +} —— — 4 Original address space

20 2] 40 41

M
410 41]
CO
SSE SEDEDODAD SD
y ¥ ¥ + y Y
P.

C
O
LO

Figure 12.23 The growth of index in dynamic hashing.

TU
.V

It is interesting to compare dynamic hashing and extendible hashing.

W

A brief, but illuminating, characterization of similarities and differences is

W

that while both schemes extend the hash function locally, as a binary
W

searchtrie, in order to handle overflow, dynamic hashing expresses the

extended directory as a Jinked structure while extendible hashing express-
‘es itas a perfect tree, whichis in turn expressible as an array.
Because of this fundamental similarity, it is not surprising that the
space utilization within the buckets is the same (69 percent) for both
approaches. Moreover, since the directories are essentially equivalent, just
expressed differently, it follows that the estimates of directory depth devel-
oped by Flajolet (1983) apply equally well to dynamic hashing and
extendible hashing. (In section 12.5.2 we talk about estimates for the
directorysize for extendible hashing, but we knowthat in extendible hash-
ing directory depth = log, diréctory sizc.)
The primarydifference between the two approachesis that dynamic
hashing allows for slower, more gradual growth of the directory, whereas
extendible hashing extends the directory by doubling it. However, because
the directory nodes in dynamic hashing mustbe capable of holding point-

https://hemanthrajhemu.github.io
: Alternative Approaches 553

ers to children, the size of a node in dynamic hashing is larger than a direc-
tory cell in extendible hashing, probably byat least a factor of 2. So, the
directory for dynamic hashing will usually require more space in memory.
Moreover, if.the directory becomesso large that it requires use of virtual
memory, extendible hashing offers the advantage of being able to access
the directory with no morethan a single pagefault. Since dynamic hash-
ing uses a linked structure for the directory, it may be necessary to incur
more than one page fault to move throughthe directory.

12.6.2 Linear Hashing

The key feature of both extendible hashing and dynamic hashing is that
they usea directory to access the buckets containing the keyrecords. This
directory makes it possible to expand and modify the hashed address space
without expanding the numberof buckets: after expanding the directory,

M
more than one directory node can point to the same bucket. However, the
CO
directory adds an additional layer of indirection which, if the directory
must be stored on disk, can result in an additional seek.
P.

Linear hashing, introduced by Litwin in 1980, does away with the

O

directory. An example, developed in Fig. 12.24, shows howlinear hashing

LO

works. This example is adapted from a description of linear hashing by

TU

Enbodyand Du (1988).
.V

Linear hashing,like extendible hashing, uses morebits of hashed value

W

as the address space grows. The example begins (Fig. 12.24{a]) with an
address space of four, which means that we are using an address function
W

that produces addresses with two bits of depth. In terms of the operations
W

that we developed earlier in this chapter, we are calling MakeAddress

with a key and a second argument of2. For this example wewill refer to
this as the h,(k) address function. Note that the address space consists of
four buckets rather than four directory nodes that can point to buckets.
As we add records, bucket b overflows. The overflow forces a split.
However,as Fig. 12.24(b) shows, it is not bucket b thatsplits, but bucket a.
The reason for this is that we are extending the address space linearly, and
bucket ais the next bucket that must split to create the next linear exten-
sion, which wecall bucket A. A 3-bit hash function, h;(k), is applied to
buckets aand A to divide the records between them. Since bucket b was
not the bucket that wesplit, the overflowing recordis placed into an over-
flow bucket w.
We add morerecords, and bucket d overflows. Bucket b is the next one
to split and extend the address space, so we use the t,(k) address function

https://hemanthrajhemu.github.io
 554 Chapter 12 Extendible Hashing

(a)| a b C d Figure 12.24 The growth of address

space in linear hashing. Adapted
oo OF 10 Ud from Enbody and Du (1988).
WwW

(b) a b c d A

a00— sO 10 il" 100

(c) a b c d A B

000 001 410 it" 100 191

M
CO
y
P.
O
LO

x
TU
.V

(d) a b c d A B Cc
W

000 001 O10 11 100 101° 110

W
W

Zz

(e){ a b c d A Bic! op
000 001 010 O11 100 101 110 211

to divide the records from bucket b andits overflow bucket w between 8

and the new bucket B. The record overflowing bucket d is placed in an
overflow bucket x. The resulting arrangementis illustrated in Fig. 12.24(c).
Figure 12.24(d) shows what happens when, as we add morerecords,
bucket d overflows beyond the capacity of the overflow bucket w. Bucket ¢
is the next in the extension sequence, so we use the ,(k) address function
to divide the records between cand C.

https://hemanthrajhemu.github.io
Alternative Approaches 555

Finally, assume that bucket B overflows. The overflow record is placed

in the overflow bucket z. The overflow also triggers the extension to buck-
et D, dividing the contents of d, x, and y between buckets d and D.At this
point all of the buckets use the ‘hy (k) address function, and we have
finished the expansion cycle. The pointer for the next bucket to be split
returns to bucket a to get ready for a new cycle that will use an h,(k)
address function to reach new buckets.
Because linear hashing uses two hash functions to reach the buckets
during an expansion cycle, an Ay(k) function for the buckets at the current
address depth and an hj, ,(k) function for the expansion buckets, finding
a record requires knowing which functionto use. If p is the pointer to the
addressof the next bucket to be split and extended, then the procedure for
finding the address of the bucket containing a key kis as follows:
1f£(hg(k) <= p)
address = hg(k);

M
else
CO
address = hg , ,(k);
P.

Litwin (1980) showsthat the access time performanceoflinear hash-

O

ing is quite good. Thereis no directory to access or maintain, and’since we

LO

extend the address space through splitting every time thereis overflow, the
TU

overflow chains do not become very large. Given a bucketsize of 50, the
average numberof disk accesses per search approachesveryclose to one.
.V

Spaceutilization, on the other hand,is lower thanit is for extendible hash-

W

ing or dynamic hashing, averaging around only 60 percent.

W
W

12.6.3 - Approaches to Controlling Splitting

We know from Chapter9 that we canincrease the storage capacity of B-
trees by implementing measuresthat tend to postponesplitting, redistrib-
uting keys between pages rather than splitting pages. We can apply similar
logic to the hashing schemesintroducedin this chapter, placing recordsin
chains.of overflow buckets to postponesplitting. .
Since linear hashinghas the lowest storage utilization of the schemes
introduced here, andsince it already includes logic to handle overflow
buckets,it is an attractive candidate for use of controlled splitting logic. in
its uncontrolled-splitting form, linear hashing splits a bucket and extends
the address space every time any bucket overflows. This choice of a trig-
gering event forsplitting is arbitrary, particularly when we considerthat
the bucket that splits is typically not the bucket that overflows. Litwin

https://hemanthrajhemu.github.io
 porter re erect Ha SHUG

(1980) suggests using the overall load factor of the file as an alternative
triggering event. Suppose we let the buckets overflow until the space
utilization reaches somedesired figure, such as 75 percent. Every time the
utilization exceeds that figure; we split a bucket and extend the address
space. Litwin simulatedthis kind of system and found thatfor load factors
of 75 percent and even 85 percent, the average numberof accesses for
successful and unsuccessful searchesstill stays below 2.
We can also use overflow buckets to defer splitting and increase space
utilization for dynamic hashing and extendible hashing. For these meth-
ods, which use directories to the buckets, deferring splitting has the addi-
tional attraction of keeping the directory size down. For extendible
hashingit is particularly advantageousto chain to an overflow bucket and
therefore avoid a split when the split would cause the directory to double
in size. Consider the example that we usedearly in this chapter, where we
split the bucket Bin Fig. 12.4(b), producing the expanded directory and

M
bucket structure shownin Fig. 12.6(c). If we had allowed bucket B to over-
CO
flow instead, we could have retained the smaller directory. Depending on
how much space weallocated for the overflow buckets, we might also have
P.

improvedspace utilization among the buckets. The cost of these improve-

O

ments, of course, is a potentially greater search length due to the overflow

LO

chains. .
TU

Studies of the effects of different overflow bucket sizes and chaining

.V

mechanisms supported a small industry of academic research during the

early and mid-1980s. Larson (1978) suggested the use of deferred splitting
W

‘in his original paper on dynamic hashing but found theresults of some
W

preliminary simulations of the idea to be disappointing. Scholl (1981)

W

developed a refinementofthis idea in which overflow buckets are shared.

Master’s thesis research by Chang (1985) tested Scholl’s suggestions empir-
ically and found thatit was possible to achieve storage utilization of about
81 percent while maintaining search performance in the range of 1.1 seeks
per search. Veklerov (1985) suggested using buddy buckets for overflow
rather than allocating chains of new buckets. This is an attractive sugges-
tion, since splitting buckets without buddies can never cause a doubling of
the directory in extendible hashing. Veklerov obtained storage utilization
of about 76 percent with a bucketsize of8.

https://hemanthrajhemu.github.io
Summary 557

SUMMARY
Conventional, static hashing does not adaptwellto file structures that are
dynamic, that grow and shrink over time. Extendible hashing is one of
several hashing systems that allow theaddress space for hashing to grow
and shrink along with the file. Because the size of the address space can
growas the file grows, it is possible for extendible hashing to provide
hashed access without the need foroverflow handling, even asfiles grow
manytimes beyondtheir original expected size.
The key to extendible hashing is using more bits of the hashed valueas
we need to cover more address space. The modelfor extending the use of
the hashed value is the trie: every time we use anotherbit of the hashed
value, we have added anotherlevel to the depth of a trie with a radix of 2.
In extendible hashingwefill out all the leaves of the trie until we have
a perfect tree, then we collapse that tree into a one-dimensional array. The

M
array forms a directory to the buckets, kept on disk, that hold the keys and
CO
records. The directory is managed in memory,if possible.
P.

If we add a record and there is no room for it in a bucket, we split the

O

bucket. We use } additional bit from the hash values for the keys in the
LO

bucket to divide the keys between the old bucket and the newone.If the
address space representéd in the directory can cover the use ofthis newbit,
TU

no more changes are necessary. If, however, the address space is using
.V

fewer bits than are needed by oursplitting buckets, then we double the
W

address space to accommodate the use of the newbit.

W

Deletion reverses the addition process, recognizing thatit is possible to

W

combine the records for two buckets only if they are buddy buckets, which
is to say that they are the pair of buckets that resulted from split.
Access performancefor extendible hashing is a single seekif the direc-
tory can be kept:in memory.If the directory must be paged off to disk,
worst-case performance1s two seeks. Space utilization for the bucketsis
approximately 69 percent. Tables and an approximation formula devel-
oped by Flajolet (1983) permit estimation of the probable directorysize,
given a bucket size and total numberof records.
There are a number of other approaches to the problem solved by
extendible hashing. Dynamic hashing uses a very similar approach but
expresses the directory as a linked structure rather than as an array. The
linked structure is more cumbersome but grows more smoothly. Space
utilization and’seek performance for dynamic hashingare the same asfor
extendible hashing.

https://hemanthrajhemu.github.io
558 Chapter 12 Extendible Hashing

Linear hashing does away with the directory entirely, extending the
address space by adding new buckets in a linear sequence. Although the
overflow of a bucket can be used to trigger extension of the address space
in linear hashing, typically the bucket that overflowsis not the one that is
split and extended. Consequently,linear hashing implies maintaining over-
flow chains and a consequent degradation in seek performance. The degra-
dationis slight, since the chains typically do not grow to be very long before
they are pulled into a new bucket. Space utilization is about 60 percent.
Space utilization for extendible, dynamic, and linear hashing can be
improved by postponingthesplitting of buckets. This is easy to implement
for linear hashing,since there are already overflow buckets. Using deferred
splitting, it is possible to increase space utilization for any of the hashing
schemes described here to 80 percent or better whilestill maintaining
search performance averagingless than two seeks. Overflow handling for
these approaches can use the sharing of overflow buckets.

M
CO
KEY TERMS
P.
O

Buddy bucket. Given a bucket with an address uvwxy, where u, v, w, x, and

LO

y have values of either 0 or I, the buddy bucket, if it exists, has the

TU

value uvwxz, such that

.V

z=yXOR 1
W

Buddy buckets are important in deletion operations for extendible

W

hashing because, if enough keys are deleted, the contents of buddy

W

buckets can be combined into a single bucket.

Deferred splitting. It is possible to improve space utilization for dynamic
hashing, extendible hashing, and linear hashing by postponing, or
deferring, the splitting of buckets, placing records into overflow buck-
ets instead. This is a classic space/time trade-off in which we accept
diminished performance in return for more compactstorage.
Directory. Conventional, static hashing schemes transform a key into a
bucket address. Both extendible hashing and dynamic hashing intro-
duce an additional layer of indirection, in which the key is hashed to a
directory address. The directory, in turn, contains information about
the location of the bucket. This additional indirection makes it possi-
ble to extend the address space by extending the directory rather than
having to work with an address space made up of buckets.

https://hemanthrajhemu.github.io
Key Terms 559

Dynamic hashing. Used in a generic sense, dynamic hashing can refer

to any hashing system that provides for expansion and contraction
of the address space for dynamicfiles where the numberof records
changesover time. In this chapter we use the term in a morespecif-
ic sense to refer to a system initially described by Larson (1978).
The system uses a directory to provide access to the buckets that
contain the records. Cells in the directory can be used as root nodes
of trie structures that accommodate greater numbers of buckets as
buckets split.
Extendible hashing. Like dynamic hashing, extendible hashing is some-
times used to refer to any hashing scheme that allows the address
space to grow andshrink so it can be used in dynamic file systems.
Used moreprecisely, as it is used in this chapter, extendible hashing
refers to an approach to hashedretrieval for dynamic files that was
first proposed by Fagin, Nievergelt, Pippenger, and Strong (1979).

M
Their proposal is for a system that uses a directory to represent
CO
the address space. Access to buckets containing the records is
through the directory. The directory is handled as an array; the
P.
O

size of the array can be doubled or halved as the number of buck-

LO

ets changes. .
TU

Linear hashing. An approach to hashing for dynamicfiles that wasfirst

proposed by Litwin (1980). Unlike extendible hashing and dynamic
.V

hashing, linear hashing does not use a directory. Instead, the address
W

spaceis extended one bucketat a time as buckets overflow. Because the

W

extension of the address space does not necessarily correspond to the

W

bucketthat is overflowing,linear hashing necessarily involves the use

of overflow buckets, even as the address space expands.
Splitting. The hashing schemes described in this chapter make room for
newrecordsbysplitting buckets to form new buckets, then extending
the address space to coverthese buckets. Coriventional, static hashing
schemesrely strictly on overflow buckets without extending the
address space.
Trie. A search tree structure in which each successive character of the key
is used to determine the direction of the search at each successive
level of the tree. The branching factor (the radix of the trie) at any
level is potentially equal to the number of values that the character
can take.

https://hemanthrajhemu.github.io
560 Chapter 12 Extendible Hashing

FURTHER READINGS
For information about hashing for dynamicfiles that goes beyond what we
present here, you must turn to journalarticles. The best summary ofthe
different approachesis Enbody and Du’s Computing Surveysarticle titled
“Dynamic Hashing Schemes,” which appeared in 1988.
The original paperon extendible hashing is “Extendible Hashing—
A Fast Access Method for Dynamic Files” by Fagin, Nievergelt, Pippen-
ger, and Strong (1979). Larson (1978) introduces dynamic hashing in an
article titled “Dynamic Hashing.” Litwin’s initial paper on linear hashing
is titled “Linear Hashing: A New Tool for File and Table Addressing”
(1980), All three of these introductory articles are quite readable;
Larson’s paper and Fagin, Nievergelt, Pippenger, and Strong are especial-
ly recommended.
Michel Scholl’s 1981 papertitled “New File Organizations Based on

M
Dynamic Hashing” provides another readable introduction to dynamic
CO
hashing. It also investigates implementations thatdefer splitting by allow-
P.

ing buckets to overflow.

O

Papers analyzing the performance of dynamic or extendible hashing

LO

" often derive results that apply to either of the two methods.Flajolet (1983)
TU

presents a careful analysis of directory depth and size. Mendelson (1982)

arrives at similar results and goes on to discuss the costs of retrieval and
.V

deletion as different design parameters are changed. Veklerov (1985)

W

analyzes the performance of dynamic hashing whensplittingis deferred by

W

allowing records to overflow into a buddy bucket. His results can be

W

applied to extendible hashing as well.

After introducing dynamic hashing, Larson wrote a number of
papers building on the ideas associated with linear hashing. His 1980
paper. titled “Linear Hashing with Partial Expansions” introduces an
approachto linear hashing that can avoid the unevendistribution of the
lengths of overflow chains across the cells in the address space. He
followed up with a performanceanalysis in a 1982 papertitled “Perfor-
mance Analysis of Linear Hashing with Partial Expansions.” A subse-
quent, 1985 paper titled “Linear Hashing with Overflow—Handling by
Linear Probing” introduces a method of handling overflow that does not
involve chaining.

https://hemanthrajhemu.github.io
 Exercises 561

EXERCISES
1. Briefly describe the differences between extendible hashing, dynamic
hashing, and linear hashing. What are the strengths and weaknesses of
each approach?
2. The tries that are the basis for the extendible hashing procedure
described in this chapter have a radix of 2. Howdoes performance
change if we use a larger radix?
3. In the MakeAddress function, what would happen if we did not
reverse the order ofthe bits but just extracted the required number of
low-orderbits in the same left-to-right order that they occur in the
address? Think about the way the directory location would change as
we extend the implicit trie structure to use yet anotherbit.
4, If the language that you are using to implement the MakeAddress

M
function does not support bit shifting and masking operations, how
CO
could you achieve the same ends, even if less elegantly andclearly?
P.

5. Inthe method Bucket: : Split, we redistribute keys between the

O

original bucket and a new one. Howdo you decide whether a key
LO

belongs in the new bucketor the original bucket?

6. Suppose the redistribution of keys in Bucket::Split does not
TU

result in moving any keys into the new bucket. Under what conditions
.V

could such an event happen? How do the methodsof classes Bucket

W

and Directory handlethis?

W

7. The Bucket: :TryCombine function is potentially recursive. In

W

section 12.4.4 we describeda situation in which there are empty buck-

ets that can be combined with other buckets through a series of recur-
sive calls to’ TryCombine. Describe two situations that could
produce empty buckets in the hashstructure.
8. Deletion occasionally results in collapsing the directory. Describe
the conditions that must be met before the directory can collapse.
What methods in classes Bucket and Directorydetect, these
conditions?
9, Deletion depends on finding buddy buckets. Why does the address
depth for a bucket have to be the same as the address depth for the
directory in order for a bucket to ave a buddy?

https://hemanthrajhemu.github.io
562 Chapter 12 Extendible Hashing

10. In the extendible hashing procedure described in this chapter, the

directory can occasionally point to empty buckets. Describe two situ-
ations that can produce empty buckets. How could we modify the
methods to avoid empty buckets?
ll. If buckets are large, a bucket containing only a few records is not
much less wasteful than an empty bucket. How could we minimize
nearly empty buckets?
12. Linear hashing makes use of overflow records. Assuming an uncon-
trolled splitting implementation in which we split and extend the
address space as soon as we have an overflow, whatis the effect of
using different bucketsizes for the overflow buckets? For example,
consider overflow buckets that are as large as the original buckets.
Now consider overflow buckets that can hold only one record. How
does this choice affect performancein termsof spaceutilization and
access time?

M
13. In section 12.6.3 we described an approach to linear hashing that
CO
controls splitting. For a load factor of 85 percent,the average number
P.

of accesses for a successful search is 1.20 (Litwin, 1980). Unsuccessful

O

searches require an average of1.78 accesses. Whyis the average search

LO

length greater for unsuccessful searches?

TU

14. Because linear hashing splits one bucket at a time, in order, until it
has reached the end of the sequence, the overflow chains forthe last
.V

buckets in the sequence can become muchlonger than those for the
W

earlier buckets. Read about Larson’s approachto solving this problem

W

through the use of “partial expansions,” originally described in Larson

W

(1980) and subsequently summarized in Enbody and Du (1988).

Write a pseudocodedescription of linear hashing with partial expan-
sions, paying particular attention to how addressing is handled.
15. In section 12.6.3 we discussed different mechanismsfor deferringthe.
splitting of buckets in extendible hashingin orderto increase storage
utilization. What is the effect of using smaller overflow buckets rather
than larger ones? How does using smaller overflow buckets compare
with sharing overflow buckets?

https://hemanthrajhemu.github.io
Programming Project 563

PROGRAMMING EXERCISES
16. Write a version of the MakeAddress function that prints out the
input key, the hash value, and the extracted, reversed address. Build a
driver that allows you to enter keys interactively for this function and
see the results. Study the operation of the function on different keys.
18. Implement method Directory: :Delete. Write a driver pro-
gram to verify that your implementation is correct. Experiment with
the program to see how deletion works. Try deleting all the keys. Try
to create situations in which the directory will recursively collapse
over more than one level.
19. Design and:‘implement a class HashedFile patterned after
class Text IndexedFile of Chapter 7 and Appendix G. A
HashedFile object is a data file and an extendible hash ditectory.

M
Theclass should have methods Create, Open, Close, Read (read
CO
record that matches key), Append, and Update.
P.
O

PROGRAMMING
LO
TU

This is'the last part of the programming project. We create a hashed index
of the student record files and the course registration files from the
.V

programmingproject of Chapter 4. This project depends on the successful

W

completion of exercise 19.

W
W

20. Use class HashedFile to create a hashed index ofa student record
file with student identifier as key. Note that the student identifier field
is not unique in a studentregistration file. Write a driver program to
create a hashed file from an existing student recordfile.
21. Use class HashedFile to create a hashed index of a courseregistra-
tion record file with student identifier as key. Write a driver program
to create a hashedfile from an existing course registration recordfile.
22. Write a program that opens a hashedstudent file and a hashed course
registrationfile and retrieves information on demand. Prompt user
for a studentidentifier and printall objects that matchit.

https://hemanthrajhemu.github.io
 M
CO
P.
O
LO
TU
.V
W
W
W

https://hemanthrajhemu.github.io