CS8493 OPERATING SYSTEMS

UNIT – III
STORAGE MANAGEMENT

Dr.N.Saravanan, Professor, Computer Science & Engg.

M N M Jain Engineering College, Chennai 1
Unit – III : STORAGE MANAGEMENT

Main Memory – Background, Swapping, Contiguou

s Memory Allocation, Paging, Segmentation,
with paging, 32 and 64 bit architecture Examples; Virtual
Segmentation
Memory – Background, Demand Paging, Page Replacement,
Allocation, Thrashing; Allocating Kernel Memory, OS
Examples.

Abraham Silberschatz, Peter Baer Galvin and Greg Gagne, “Operating

System
Concepts”, 9 th Edition, John Wiley and Sons Inc., 2
 Main Memory
 Program must be brought (from disk) into memory
and placed within a process for it to be run
 Main memory and registers are only storage CPU
can access directly
 Memory unit only sees a stream of addresses +
read requests, or address + data and write requests
 Register access in one CPU clock (or less)
 Main memory can take many cycles, causing a stall
 Cache sits between main memory and CPU registers
 Protection of memory required to ensure
correct operation
3
 Base and Limit Registers
 A pair of base and limit registers define the
logical address space
 CPU must check every memory access generated in user
mode to be sure it is between base and limit for that user

Hardware Address Protection

4
 Address Binding
 Programs on disk, ready to be brought into memory to execute form an
input queue
 Without support, must be loaded into address 0000
 Inconvenient to have first user process physical address always at 0000
 How can it not be?
 Further, addresses represented in different ways at different stages of a
program’s life
 Source code addresses usually symbolic
 Compiled code addresses bind to relocatable addresses
 i.e. “14 bytes from beginning of this module”
 Linker or loader will bind relocatable addresses to absolute
addresses i.e. 74014
 Each binding maps one address space to another

5
Binding of Instructions and Data to Memory
 Address binding of instructions and data to
memory addresses can happen at three different
stages
 Compile time: If memory location known a
priori, absolute code can be generated; must
recompile code if starting location changes
 Load time: Must generate relocatable code if
memory location is not known at compile
time
 Execution time: Binding delayed until run
time if the process can be moved during its
execution from one memory segment to
another
 Need hardware support for address maps Multistep Processing
(e.g., base and limit registers) of a User Program
6
 Logical vs. Physical Address Space
 The concept of a logical address space that is bound to a separate
physical address space is central to proper memory management
 Logical address – generated by the CPU; also referred to as
virtual address
 Physical address – address seen by the memory unit
 Logical and physical addresses are the same in compile-time and
load-time address-binding schemes; logical (virtual) and physical
addresses differ in execution-time address-binding scheme
 Logical address space is the set of all logical addresses generated
by a program
 Physical address space is the set of all physical addresses
generated by a program

7
 Memory-Management Unit (MMU)
 Hardware device that at run time maps virtual to physical address
 Many methods possible, covered in the rest of this chapter
 To start, consider simple scheme where the value in the relocation
register is added to every address generated by a user process at
the time it is sent to memory
 Base register now called relocation register
 MS-DOS on Intel 80x86 used 4 relocation registers
 The user program deals with logical addresses; it never sees the
real physical addresses
 Execution-time binding occurs when is made to
reference location in memory
 Logical address bound to physical addresses

8
 Dynamic relocation using a relocation

register
Routine is not loaded until it is called
 Better memory-space utilization; unused
routine is never loaded
 All routines kept on disk in relocatable load
format
 Useful when large amounts of code are
needed to handle infrequently occurring
cases
 No special support from the
operating system is required
 Implemented through program
design
 OS can help by providing
libraries to
implement dynamic loading 9
 Dynamic Linking
 Static linking – system libraries and program code combined by the
loader into the binary program image
 Dynamic linking –linking postponed until execution time
 Small piece of code, stub, used to locate the appropriate memory-
resident library routine
 Stub replaces itself with the address of the routine, and executes
the routine
 Operating system checks if routine is in memory
processes’
address
 If not in address space, add to address space
 Dynamic linking is particularly useful for libraries
 System also known as shared libraries
 Consider applicability to patching system 10
 Swapping
 A process can be swapped temporarily out of memory to a
backing store, and then brought back into memory for
continued execution
Total physical memory space of processes can exceed
physical memory
 Backing store – fast disk large enough to accommodate copies
of all memory images for all users; must provide direct access
to these memory images
 Roll out, roll in – swapping variant used for priority-based
scheduling algorithms; lower-priority process is swapped out
so higher-priority process can be loaded and executed
 Major part of swap time is transfer time; total transfer time is
directly proportional to the amount of memory swapped
 System maintains a ready queue of ready-to-run processes
which have memory images on disk

11
 Swapping
 Does the swapped out process
need to swap back in to same
physical addresses?
 Depends on address binding
method
 Plus consider pending I/O to /
from process memory space
 Modified versions of swapping are
found on many systems (i.e., UNIX,
Linux, and Windows)
 Swapping normally disabled
 Started if more than threshold
amount of memory allocated
Schematic View of
 Disabled again once
Swapping
demand memory reduced
threshold 12
below
 Context Switch Time including Swapping
 If next processes to be put on CPU is not in memory, need to swap
out a process and swap in target process
 Context switch time can then be very high
 100MB process swapping to hard disk with transfer rate
of 50MB/sec
 Swap out time of 2000 ms
 Plus swap in of same sized process
 Total context switch swapping component time of
4000ms (4 seconds)
 Can reduce if reduce size of memory swapped – by knowing how
much memory really being used
 System calls to inform OS of memory use via request_memory()
and release_memory()
13
 Context Switch Time and Swapping
 Other constraints as well on swapping
Pending I/O – can’t swap out as I/O would
occur to wrong process
Or always transfer I/O to kernel space, then
to I/O device
 Known as double buffering, adds overhead
 Standard swapping not used in modern
operating systems
 But modified version common
 Swap only when free memory extremely low

14
 Swapping on Mobile Systems
 Not typically supported
 Flash memory based
 Small amount of
space
 Limited number
of write cycles
Poor throughput between flash memory and CPU on mobile
platform
 Instead use other methods to free memory if low
Read-only data thrown out and reloaded flash
 iOS asks apps
fromtoneeded
voluntarily relinquish allocated memory if
 Failure to free can result in termination
 Android terminates apps if low free memory, but first
writes
 application state topaging
Both OSes support flash for
as fast restartbelow
discussed 15
 Contiguous Allocation
 Main memory must support both OS and user processes
 Limited resource, must allocate efficiently
 Contiguous allocation is one early method
 Main memory usually into two partitions:
 Resident operating system, usually held in low
memory with interrupt vector
 User processes then held in high memory
 Each process contained in single contiguous section of memory
 Relocation registers used to protect user processes from each
other, and from changing operating-system code and data
 Base register contains value of smallest physical address

16
 Contiguous Allocation
 Limit register contains
range of logical addresses
– each logical address
must be less than the limit
register
 MMU maps logical
address dynamically
 Can then allow actions
such as kernel code being
Hardware Support for
transient and kernel
Relocation and Limit Registers
changing size

17
 Multiple-partition allocation
 Multiple-partition allocation
 Degree of multiprogramming limited by number of partitions
 Variable-partition sizes for efficiency (sized to a given process’ needs)
 Hole – block of available memory; holes of various size are
scattered throughout memory
 When a process arrives, it is allocated memory from a hole large enough to
accommodate it
 Process exiting frees its partition, adjacent free partitions combined
 Operating system maintains information
about:
a) allocated partitions b) free partitions (hole)

18
Dynamic Storage-Allocation Problem
First-fit: Allocate the first hole that is big
enough
First-fit and best-fit better than worst-fit in terms of speed and storage utilization

Best-fit: Allocate the smallest hole that is

big enough; must search entire list, unless
ordered by size
 Produces the smallest leftover hole
How to satisfy a request of size n from a list of free holes?

Worst-fit: Allocate the largest hole; must

also search entire list
 Produces the largest leftover hole
19
 Fragmentation
 External Fragmentation – total memory space exists to
satisfy a request, but it is not contiguous
 Internal Fragmentation – allocated memory may be
slightly larger than requested memory; this size
difference is memory internal to a partition, but not
being used
 First fit analysis reveals that given N blocks allocated, 0.5
N blocks lost to fragmentation
 1/3 may be unusable -> 50-percent rule

20
 Fragmentation
 Reduce external fragmentation by compaction
Shuffle memory contents to place all free
memory together in one large block
Compaction is possible only if relocation is dynamic,
and is done at execution time
 I/O problem
 Latch job in memory while it is involved in I/O
 Do I/O only into OS buffers
 Now consider that backing store has
same fragmentation problems

21
 Segmentation
 Memory-management scheme that supports user view of memory
 A program is a collection of segments
 A segment is a logical unit such as:
main program
procedure
function
method
object
local variables, global variables
common block
stack
symbol table
arrays User’s View of a Program
22
 Segmentation Architecture
 Logical address consists of a two tuple:
<segment-number, offset>,
Logical View of
 Segment table – maps two-dimensional
physical addresses; each table entry has:
Segmentation
 base – contains the starting physical
address where the segments reside in
memory
 limit – specifies the length of the
segment
 Segment-table base register (STBR)
points to the segment table’s location in
use space
memory
 Segment-table length register (STLR)
indicates number of segments used by a physical memory space
program;
segment number s is legal if s < STLR 23
 Segmentation Architecture
 Protection
 With each entry in
segment table associate:
 validation bit = 0  illegal
segment
 read/write/
execute
privileges
 Protection bits associated with
segments; code sharing occurs at
segment
 Since level vary in length,
segments
memory allocation is a dynamic Segmentation Hardware
storage-allocation problem
 A segmentation example is shown
in the following diagram
24
 Paging
 Physical address space of a process can be noncontiguous;
process is allocated physical memory whenever the latter is
available
 Avoids external fragmentation
 Avoids problem of varying sized memory chunks
 Divide physical memory into fixed-sized blocks called frames
 Size is power of 2, between 512 bytes and 16 Mbytes
 Divide logical memory into blocks of same size called pages
 Keep track of all free frames
 To run a program of size N pages, need to find N free frames and
load program
 Set up a page table to translate logical to physical addresses

 Backing
Still havestore likewise
Internal split into pages
fragmentation
25
 Address Translation Scheme
 Address generated by CPU is divided into:
Page number (p) – used as an index into a page table
which contains base address of each page in physical
memory
Page offset (d) – combined with base address to
define the physical memory address that is sent to the
memory unit page number page offset
p d

m -n n

 For given logical address space 2m and page size 2n

26
 Paging Hardware

 Calculating internal fragmentation

 Page size = 2,048 bytes
 Process size = 72,766 bytes
 35 pages + 1,086 bytes
 Internal fragmentation of 2,048 - 1,086 = 962 bytes
 Worst case fragmentation = 1 frame – 1 byte

27
Paging Model of Logical and Physical Memory

 On average fragmentation = 1 / 2 frame size

 So small frame sizes desirable?
 But each page table entry takes memory to track
 Page sizes growing over time
 Solaris supports two page sizes – 8 KB and 4 MB
 Process view and physical memory now very different
 By implementation process can only access its own memory

28
 Paging Example
 Page table is kept in
main
memory
 Page-table base register
(PTBR)
points to the page table
 Page-table length register (PTLR)
indicates size of the page table

 In this scheme every data/instruction access requires two

memory accesses
 One for the page table and one for the data / instruction
 The two memory access problem can be solved by the use of a
special fast-lookup hardware cache called associative memory or
translation look-aside buffers (TLBs)

29
 Free Frames
 Some TLBs
store
identifiers (ASIDs)
address-space
in
each –
TLB
uniquely entry
identifies
each process to
provide address-space
protection for that
process
 Otherwise need to
flush at
every switch
context
 TLBs typically small (64 to 1,024 entries)
 On a TLB miss, value is loaded into the TLB for faster access next
time

 Replacement policies must down
be considered
Some entries can be wired for permanent fast
access 30
 Associative Memory
Associative memory – parallel search

Page # Frame #

Address translation (p, d)

 If p is in associative register, get frame # out
 Otherwise get frame # from page table in memory

31
Paging Hardware With TLB

32
 Effective Access Time
 Associative Lookup =  time unit
 Can be < 10% of memory access time
 Hit ratio = 
 Hit ratio – percentage of times that a page number is found in
the associative registers; ratio related to number of associative
registers
 Consider  = 80%,  = 20ns for TLB search, 100ns for memory access
 Effective Access Time (EAT) = (1 + )  + (2 + )(1 – ) = 2 +  – 
 Consider  = 80%,  = 20ns for TLB search, 100ns for memory
access
 EAT = 0.80 x 100 + 0.20 x 200 = 120ns
 Consider more realistic hit ratio ->  = 99%,  = 20ns for TLB search,
100ns for memory access
 EAT = 0.99 x 100 + 0.01 x 200 = 101ns
33
 Memory Protection
 Memory protection implemented by associating protection bit
with each frame to indicate if read-only or read-write access is
allowed
Can also add more bits to indicate page execute-only, and
so on
 Valid-invalid bit attached to each entry in the page table:
“ valid ” indicates that the associated page is
in the process’ logical address space, and is thus a legal
page
“invalid” indicates that the page is not in the
process’logical address space
 Or use page-table length register (PTLR)
 Any violations result in a trap to the kernel
34
Valid (v) or Invalid (i) Bit In A Page
Table

35
 Shared Pages
 Shared code
One copy of read-only (reentrant) code shared
among processes (i.e., text editors, compilers, window
systems)
Similar to multiple threads sharing the same
process space
Also useful for interprocess communication if sharing of
read-write pages is allowed
 Private code and data
Each process keeps a separate copy of the
code and data
The pages for the private code and data can
appear anywhere in the logical address space
36
Shared Pages Example

37
 Structure of the Page Table
 Memory structures for paging can get huge using straight-forward
methods
 Consider a 32-bit logical address space as on modern computers
 Page size of 4 KB (212)
 Page table would have 1 million entries (232 / 212)
 If each entry is 4 bytes -> 4 MB of physical address
space / memory for page table alone
 That amount of memory used to cost a lot
 Don’t want to allocate that contiguously in main memory
 Hierarchical Paging
 Hashed Page Tables
 Inverted Page Tables

38
 Hierarchical Page Tables
 Break up the logical address space into multiple
page tables
 A simple technique is a two-level page table
 We then page the page table

Two-Level Page-Table Scheme

39
 Two-Level Paging Example
A logical address (on 32-bit machine with 1K page
size) is divided into:
 a page number consisting of 22 bits
 a page offset consisting of 10 bits

Since the page table is paged, the page number is

further divided into:
 a 12-bit page number
 a 10-bit page offset

Thus, a logical address is as follows:

40
 Address-Translation Scheme

Three-level Paging
Scheme

41
 64-bit Logical Address Space
 Even two-level paging scheme not sufficient
 If page size is 4 KB (212)
 Then page table has 252 entries
 If two level scheme, inner page tables could be 210 4-byte
entries
 Address would look like

 Outer page table has 242 entries or 244 bytes

 One solution is to add a 2nd outer page table
 But in the following example the 2nd outer page table is still 234
bytes in size
And possibly 4 memory access to get to one physical
memory location
42
 Hashed Page Tables
 Common in address spaces > 32 bits
 The virtual page number is hashed into a page table
 This page table contains a chain of elements hashing to the same
location
 Each element contains (1) the virtual page number (2) the value of
the mapped page frame (3) a pointer to the next element
 Virtual page numbers are compared in this chain searching
fora match
 If a match is found, the corresponding physical frame is extracted
 Variation for 64-bit addresses is clustered page tables
 Similar to hashed but each entry refers to several pages (such as
16) rather than 1
 Especially useful for sparse address spaces memory
(where references are non-contiguous and scattered) 43
Hashed Page Table

44
 Inverted Page Table
 Rather than each process having a page table and keeping track of
all possible logical pages, track all physical pages
 One entry for each real page of memory
 Entry consists of the virtual address of the page stored in that real
memory location, with information about the process that owns that
page
 Decreases memory needed to store each page table, but increases
time needed to search the table when a page reference occurs
 Use hash table to limit the search to one — or at most a few —
page-table entries
 TLB can accelerate access
 But how to implement shared memory?
 One mapping of a virtual address to the shared physical address
45
Inverted Page Table Architecture

46
 Oracle SPARC Solaris
 Consider modern, 64-bit operating system example with
tightly integrated HW
 Goals are efficiency, low overhead
 Based on hashing, but more complex
 Two hash tables
 One kernel and one for all user
processes
 Each maps memory addresses
from virtual to physical memory
 Each entry represents a contiguous area of mapped
virtual memory,
More efficient than having a separate hash-table entry for
each page
 Each entry has base address and span (indicating the
47
 Oracle SPARC Solaris
 TLB holds translation table entries (TTEs) for fast hardware lookups
 A cache of TTEs reside in a translation storage buffer (TSB)
 Includes an entry per recently accessed page
 Virtual address reference causes TLB search
 If miss, hardware walks the in-memory TSB looking for the TTE
corresponding to the address
If match found, the CPU copies the TSB entry into the TLB and
translation completes
 If no match found, kernel interrupted to search the hash table
The kernel then creates a TTE from the appropriate hash
table and stores it in the TSB, Interrupt handler returns
control to the MMU, which completes the address
translation.
48
 Example: The Intel 32 and 64-bit
Architectures
 Dominant industry chips

 Pentium CPUs are 32-bit and called IA-32 architecture

 Current Intel CPUs are 64-bit & called IA-64 architecture

 Many variations in the chips, cover the main ideas here

49
Example: The Intel IA-32 Architecture
 Supports both segmentation and segmentation with paging
 Each segment can be 4 GB
 Up to 16 K segments per process
 Divided into two partitions
First partition of up to 8 K segments are private to process
(kept in local descriptor table (LDT))
Second partition of up to 8K segments shared
among all processes (kept in global descriptor table (GDT))
 CPU generates logical address
 Selector given to segmentation unit
 Which produces linear addresses

50
Example: The Intel IA-32 Architecture
 Linear address given to paging unit
 Which generates physical address in main memory
 Paging units form equivalent of MMU
 Pages sizes can be 4 KB or 4 MB

51
Intel IA-32 Segmentation

52
Intel IA-32 Paging Architecture

53
 Intel IA-32 Page Address Extensions
 32-bit address limits led Intel to create page address extension (PAE), allowing
32-bit apps access to more than 4GB of memory space
 Paging went to a 3-level scheme
 Top two bits refer to a page directory pointer table
 Page-directory and page-table entries moved to 64-bits in size
 Net effect is increasing address space to 36 bits – 64GB of physical memory

54
 Intel x86-64
 Current generation Intel x86 architecture
 64 bits is ginormous (> 16 exabytes)
 In practice only implement 48 bit addressing
 Page sizes of 4 KB, 2 MB, 1 GB
 Four levels of paging hierarchy
 Canalso use PAE so virtual addresses are 48 bits and
physical addresses are 52 bits

55
 Example: ARM Architecture
 Dominant mobile platform 32
chip (Apple iOS and Google outer
bits
inner offse
Android devices for example) page page t

 Modern, energy efficient, 32- 4-KB

bit CPU or
16-
KB
 4 KB and 16 KB pages page

 1 MB and 16 MB pages 1-MB

(termed sections) 16-
or

MB
 One-level paging for sections, section

two-
level for smaller pages
 Two levels of TLBs
 Outer level has two micro
TLBs
 First inner is checked, on miss outers are checked, and on miss page
(one data, one instruction)
table walk performed by CPU
 Inner is single main TLB
56
 Virtual Memory
 Code needs to be in memory to execute, but entire program rarely
used
 Error code, unusual routines, large data structures
 Entire program code not needed at same time
 Consider ability to execute partially-loaded program
 Program no longer constrained by limits of physical memory
 Each program takes less memory while running ->
more programs run at the same time
Increased CPU utilization and throughput with no increase in
response time or turnaround time
 Less I/O needed to load or swap programs into memory -> each
user program runs faster

57
 Background (Cont.)
 Virtual memory – separation of user logical
memory from physical memory
Only part of the program needs to be in memory for
execution
Logical address space can therefore be much
larger than physical address space
Allows address spaces to be by several
shared processes
 Allows for more efficient process
creation
 More programs running concurrently
 Less I/O needed to load or swap
processes 58
 Background (Cont.)
 Virtual address space – logical view of how process
is stored in memory
Usually start at address 0, contiguous addresses until
end of space
Meanwhile, physical memory in
organized frames page
 MMU must map logical to physical
 Virtual memory can be implemented via:
 Demand paging
 Demand segmentation

59
Virtual Memory That is Larger Than Physical Memory

60
 Virtual-address Space
 Usually design logical address space for stack to start at
Max logical address and grow “down” while heap grows
“up”
 Maximizes address space use
 Unused address space between the two is hole
 No physical memory needed until heap or stack
grows to a given new page
 Enables sparse address spaces with holes left for growth,
dynamically linked libraries, etc
 System libraries shared via mapping into virtual address
space
 Shared memory by mapping pages read-write into virtual
address space
 Pages can be shared during fork(), speeding
process
creation
61
Shared Library Using Virtual Memory

62
 Demand Paging
 Could bring entire process into memory at
load time
 Or bring a page into memory only when it is
needed
 Less I/O needed, no unnecessary I/O
 Less memory needed
 Faster response
 More users
 Similar to paging system with swapping
(diagram on right)
 Page is needed  reference to it
 invalid reference  abort
 not-in-memory  bring to memory
 Lazy swapper – never swaps a page into
memory unless page will be needed
 Swapper that deals with pages is a
pager

63
 Basic Concepts
 With swapping, pager guesses which pages will be used before
swapping out again
 Instead, pager brings in only those pages into memory
 How to determine that set of pages?
 Need new MMU functionality to implement demand paging
 If pages needed are already memory resident
 No difference from non demand-paging
 If page needed and not memory resident
 Need to detect and load the page into memory from storage
 Without changing program behavior
 Without programmer needing to change code

64
 Valid-Invalid Bit
 With each page table entry a valid–invalid bit is associated (v  in-
memory – memory resident, i  not-in-memory)
 Initially valid–invalid bit is set to i on all entries
 Example of a page table snapshot:

Page Table When Some Pages Are

Not in Main Memory

65
 Page Fault
 If there is a reference to a page, first reference to that
page will trap to operating system: page fault
 Operating system looks at another table to decide:
 Invalid reference  abort
 Just not in memory
 Find free frame
 Swap page into frame via scheduled disk operation
 Reset tables to indicate page now in memory
Set validation bit = v
 Restart the instruction that caused the page
fault

66
Steps in Handling a Page Fault

67
 Aspects of Demand Paging
 Extreme case – start process with no pages in memory
 OS sets instruction pointer to first instruction of process, non-
memory-resident -> page fault
 And for every other process pages on first access
 Pure demand paging
 Actually, a given instruction could access multiple pages -> multiple
page faults
 Consider fetch and decode of instruction which adds 2 numbers
from memory and stores result back to memory
 Pain decreased because of locality of reference
 Hardware support needed for demand paging
 Page table with valid / invalid bit
 Secondary memory (swap device with swap space)
 Instruction restart 68
 Instruction Restart
 Consider an instruction that could access several different
locations
 Block move

 auto increment/decrement location

 Restart the whole operation?
 What if source and destination overlap?

69
 Performance of Demand Paging
 Stages in Demand Paging (worse case)
1. Trap to the operating system
2. Save the user registers and process state
3. Determine that the interrupt was a page fault
4. Check that the page reference was legal and determine the location of the page on the disk
5. Issue a read from the disk to a free frame:
1. Wait in a queue for this device until the read request is serviced
2. Wait for the device seek and/or latency time
3. Begin the transfer of the page to a free frame
6. While waiting, allocate the CPU to some other user
7. Receive an interrupt from the disk I/O subsystem (I/O completed)
8. Save the registers and process state for the other user
9. Determine that the interrupt was from the disk
10. Correct the page table and other tables to show page is now in memory
11. Wait for the CPU to be allocated to this process again
12. Restore the user registers, process state, and new page table, and then resume the interrupted
instruction
70
Performance of Demand Paging (Cont.)
 Three major activities
 Service the interrupt – careful coding means just several
hundred instructions needed
 Read the page – lots of time
 Restart the process – again just a small amount of time
 Page Fault Rate 0  p 
1
 if p = 0 no page faults
 if p = 1, every
reference is a fault
 Effective Access Time
(EAT)
= (1 – p) x memory
access
 Demand Paging Example
 Memory access time = 200 nanoseconds
 Average page-fault service time = 8 milliseconds
 EAT = (1 – p) x 200 + p (8 milliseconds)
= (1 – p x 200 + p x 8,000,000
= 200 + p x 7,999,800
 If one access out of 1,000 causes a page fault, then EAT = 8.2
microseconds.
 This is a slowdown by a factor of 40!!
 If want performance degradation < 10 percent
 220 > 200 + 7,999,800 x
p 20 > 7,999,800 x p
 p < .0000025
 < one page fault in every
72
400,000 memory
 Demand Paging Optimizations
 Swap space I/O faster than file system I/O even if on the same device
 Swap allocated in larger chunks, less management needed than file system
 Copy entire process image to swap space at process load time
 Then page in and out of swap space
 Used in older BSD Unix
 Demand page in from program binary on disk, but discard rather than paging
out when freeing frame
 Used in Solaris and current BSD
 Still need to write to swap space
 Pages not associated with a file (like stack and heap) – anonymous
memory
 Pages modified in memory but not yet written back to the file system
 Mobile systems
 Typically don’t support swapping
 Instead, demand page from file system and reclaim read-only pages (such as
code) 73
 Copy-on-Write
 Copy-on-Write (COW) allows both parent and child processes to initially
share the same pages in memory
 If either process modifies a shared page, only then is the page copied
 COW allows more efficient process creation as only modified pages are
copied
 In general, free pages are allocated from a pool of zero-fill-on-demand
pages
 Pool should always have free frames for fast demand page execution
Don’t want to have to free a frame as well as other processing on
page fault
 Why zero-out a page before allocating it?
 vfork() variation on fork() system call has parent suspend and child using
copy-on-write address space of parent
 Designed to have child call exec()
 Very efficient 74
Before Process 1 Modifies Page C

75
After Process 1 Modifies Page C

76
What Happens if There is no Free Frame?
 Used up by process pages
 Also in demand from the kernel, I/O buffers, etc
 How much to allocate to each?
 Page replacement – find some page in memory, but not
really in use, page it out
 Algorithm – terminate? swap out? replace the page?
Performance – want an algorithm which will result in
minimum number of page faults
 Same page may be brought into memory several times

77
 Page Replacement
Prevent over-allocation of memory by modifying
page-fault service routine to include page
replacement
Use modify (dirty) bit to reduce overhead of page
transfers – only modified pages are written to
disk
Page replacement completes separation between
logical memory and physical memory – large
virtual memory can be provided on a smaller
physical memory
78
Need For Page Replacement

79
 Basic Page Replacement
1. Find the location of the desired page on disk
2. Find a free frame:
- If there is a free frame, use it
-If there is no free frame, use a page replacement algorithm to
select a victim frame
- Write victim frame to disk if dirty
3. Bring the desired page into the (newly) free frame; update
the page and frame tables
4. Continue the process by restarting the instruction that caused
the trap
Note now potentially 2 page transfers for page fault – increasing
EAT

80
Page Replacement

81
Page and Frame Replacement Algorithms
 Frame-allocation algorithm determines
 How many frames to give each process
 Which frames to replace
 Page-replacement algorithm
 Want lowest page-fault rate on both first access and re-access
 Evaluate algorithm by running it on a particular string of memory
references (reference string) and computing the number of page
faults on that string
 String is just page numbers, not full addresses
 Repeated access to the same page does not cause a page fault
 Results depend on number of frames available
 In all our examples, the reference string of referenced
page numbers is 7,0,1,2,0,3,0,4,2,3,0,3,0,3,2,1,2,0,1,7,0,1
82
Graph of Page Faults Versus The Number of
Frames

83
 First-In-First-Out (FIFO) Algorithm
 Reference string: 7,0,1,2,0,3,0,4,2,3,0,3,0,3,2,1,2,0,1,7,0,1
 3 frames (3 pages can be in memory at a time per process)

 Can vary by reference string: consider 1,2,3,4,1,2,5,1,2,3,4,5

 Adding more frames can cause more page faults!
 Belady’s Anomaly
 How to track ages of pages?
 Just use a FIFO queue

84
FIFO Illustrating Belady’s Anomaly

85
 Optimal Algorithm
Replace page that will not be used for
longest period of time
 9 is optimal for the example
How do you know this?
 Can’t read the future
Used for measuring how well your
algorithm performs

86
 Least Recently Used (LRU) Algorithm
 Use past knowledge rather than future
 Replace page that has not been used in the most amount of time
 Associate time of last use with each page

 12 faults – better than FIFO but worse than OPT

 Generally good algorithm and frequently used
 But how to implement?

87
 LRU Algorithm (Cont.)
 Counter implementation
 Every page entry has a counter; every time page is
referenced
through this entry, copy the clock into the counter
 When a page needs to be changed, look at the counters
to find smallest value
 Search through table needed
 Stack implementation
 Keep a stack of page numbers in a double link form:
 Page referenced:
 move it to the top
 requires 6 pointers to be changed
 But each update more expensive
 No search for replacement
 LRU and OPT are cases of stack algorithms that don’t have Belady’s88
Anomaly
Use Of A Stack to Record Most Recent Page
References

89
 LRU Approximation Algorithms
 LRU needs special hardware and still slow
 Reference bit
 With each page associate a bit, initially = 0
 When page is referenced bit set to 1
 Replace any with reference bit = 0 (if one exists)
 We do not know the order, however
 Second-chance algorithm
 Generally FIFO, plus hardware-provided reference bit
 Clock replacement
 If page to be replaced has
 Reference bit = 0 -> replace it
 reference bit = 1 then:
 set reference bit 0, leave page in memory
 replace next page, subject to same rules
90
Second-Chance (clock) Page-Replacement Algorithm

91
 Enhanced Second-Chance Algorithm
 Improve algorithm by using reference bit and modify bit
(if available) in concert
 Take ordered pair (reference, modify)
1. (0, 0) neither recently used not modified – best page to replace
2. (0, 1) not recently used but modified – not quite as good, must
write out before replacement
3. (1, 0) recently used but clean – probably will be used again soon
4. (1, 1) recently used and modified – probably will be used again
soon and need to write out before replacement
 When page replacement called for, use the clock scheme but
use the four classes replace page in lowest non-empty class
 Might need to search circular queue several times

92
 Counting Algorithms
Keep a counter of the number of references that
have been made to each page
 Not common
Lease Frequently Used (LFU) Algorithm:
replaces page with smallest count
Most Frequently Used (MFU) Algorithm: based
on the argument that the page with the smallest
count was probably just brought in and has yet
to be used

93
 Page-Buffering Algorithms
 Keep a pool of free frames, always
 Then frame available when needed, not found at fault time
 Read page into free frame and select victim to evict and add to
free pool
 When convenient, evict victim
 Possibly, keep list of modified pages
 When backing store otherwise idle, write pages there and set to
non-dirty
 Possibly, keep free frame contents intact and note what is in them
 If referenced again before reused, no need to load
contents again from disk
 Generally useful to reduce penalty if wrong victim
frame selected
94
 Applications and Page Replacement
 All of these algorithms have OS guessing about
future page access
 Some applications have better knowledge – i.e. databases
 Memory intensive applications can cause double buffering
 OS keeps copy of page in memory as I/O buffer
 Application keeps page in memory for its own work
 Operating system can given direct access to the
disk, getting out of the way of the applications
 Raw disk mode
 Bypasses buffering, locking, etc

95
 Allocation of Frames
 Each process needs minimum number of frames
 Example: IBM 370 – 6 pages to handle SS
MOVE instruction:
 instruction is 6 bytes, might span 2 pages
 2 pages to handle from
 2 pages to handle to
 Maximum of course is total frames in the
system
 Two major allocation schemes
 fixed allocation
 priority allocation
 Many variations
96
 Fixed Allocation
 Equal allocation – For example, if there are 100 frames (after
allocating frames for the OS) and 5 processes, give each process
20 frames
 Keep some as free frame buffer pool
 Proportional allocation – Allocate according to the size of process
 Dynamic as degree of multiprogramming, process sizes change
m
si  size of process pi 64
ss12 

S   si 10
127 10
m  total number of a1   62 
137
frames
i  allocation for pi  S
a i m 127
4a  62 
2
s 
137 57

97
 Priority Allocation
Use a proportional allocation using
scheme priorities rather than
size
If process Pi generates a page fault,
 select for replacement one of its frames
select for replacement a frame from a process
with lower priority number

98
 Global vs. Local Allocation
Global replacement – process selects a
replacement frame from the set of all frames; one
process can take a frame from another
 But then process execution time can vary greatly
 But greater throughput so more common
Local replacement – each process selects from
only its own set of allocated frames
 More consistent per-process performance
 But possibly underutilized memory

99
 Non-Uniform Memory Access
 So far all memory accessed equally
 Many systems are NUMA – speed of access to memory varies
 Consider system boards containing CPUs and
memory, interconnected over a system bus
 Optimal performance comes from allocating memory “close to” the
CPU on which the thread is scheduled
 And modifying the scheduler to schedule the thread on the same
system board when possible
 Solved by Solaris by creating lgroups
 Structure to track CPU / Memory low latency groups
 Used my schedule and pager
When possible schedule all threads of a process and allocate
all memory for that process within the lgroup
100
 Thrashing
 If a process does not have “enough” pages, the page-fault rate is
very high
 Page fault to get page
 Replace existing frame
 But quickly need replaced frame back
 This leads to:
 Low CPU utilization
Operating system thinking that it needs to
increase the degree of multiprogramming
 Another process added to the
system

 Thrashing  a process is busy swapping pages in and out

101
 Demand Paging and Thrashing
Why does demand paging work?
Locality model
 Process migrates from one locality to another
 Localities may overlap
Why does thrashing occur?
 size of locality > total memory size
 Limit effects by using local or priority page replacement

102
 Working-Set Model
   working-set window  a fixed number of page references
Example: 10,000 instructions
 WSSi (working set of Process Pi) =
total number of pages referenced in the most recent  (varies in time)
 if  too small will not encompass entire locality
 if  too large will encompass several localities
 if  =   will encompass entire program
 D =  WSSi  total demand frames
 Approximation of locality
 if D > m  Thrashing
 Policy if D > m, then suspend or swap out one of the processes

103
 Keeping Track of the Working Set
 Approximate with interval timer + a reference bit
 Example:  = 10,000
 Timer interrupts after every 5000 time units
 Keep in memory 2 bits for each page
 Whenever a timer interrupts copy and sets the values of all
reference bits to 0
 If one of the bits in memory = 1  page in working set
 Why is this not completely accurate?
 Improvement = 10 bits and interrupt every 1000 time units

104
 Page-Fault Frequency
 More direct approach than WSS
 Establish “acceptable” page-fault frequency (PFF) rate
and use local replacement policy
 If actual rate too low, process loses frame
 If actual rate too high, process gains frame

105
 Working Sets and Page Fault Rates
 Direct relationship between working set of a process and
its page-fault rate
 Working set changes over time
 Peaks and valleys over time

106
 Memory-Mapped Files
 Memory-mapped file I/O allows file I/O to be treated as routine
memory access by mapping a disk block to a page in memory
 A file is initially read using demand paging
 A page-sized portion of the file is read from the file system into
a physical page
 Subsequent reads/writes to/from the file are
treated as ordinary memory accesses
 Simplifies and speeds file access by driving file I/O through memory
rather than read() and write() system calls
 Also allows several processes to map the same file allowing
the pages in memory to be shared
 But when does written data make it to disk?
 Periodically and / or at file close() time
 For example, when the pager scans for dirty pages 107
Memory-Mapped File Technique for all I/O
 Some OSes uses memory mapped files for standard I/O
 Process can explicitly request memory mapping a file via
mmap() system call
 Now file mapped into process address space
 For standard I/O (open(), read(), write(), close()), mmap anyway
 But map file into kernel address space
 Process still does read() and write()
 Copies data to and from kernel space and user space
 Uses efficient memory management subsystem
 Avoids needing separate subsystem
 COW can be used for read/write non-shared pages
 Memory mapped files can be used for shared memory (although
again via separate system calls)
108
Memory Mapped Files

109
Shared Memory via Memory-Mapped I/O

110
 Shared Memory in Windows API
First create a file mapping for file to be mapped
Then establish a view of the mapped file in process’s
virtual address space
Consider producer / consumer
Producer create shared-memory object using memory
mapping features
 Open file via CreateFile(), returning a
HANDLE
 Create mapping via CreateFileMapping()
creating a named shared-memory object
 Create view via MapViewOfFile()
Sample code in Textbook 111
 Allocating Kernel Memory
Treated differently from user memory
Often allocated from a free-memory pool
Kernel requests memory for
structures of varying sizes
Some kernel memory needs to be contiguous
I.e. for device I/O

112
 Buddy System
 Allocates memory from fixed-size segment consisting of physically-contiguous
pages
 Memory allocated using power-of-2 allocator
 Satisfies requests in units sized as power of 2
 Request rounded up to next highest power of 2
 When smaller allocation needed than is available, current chunk split into
two buddies of next-lower power of 2
 Continue until appropriate sized chunk available
 For example, assume 256KB chunk available, kernel requests 21KB
 Split into AL and AR of 128KB each
 One further divided into BL and BR of 64KB
 One further into CL and CR of 32KB each – one used to satisfy
request
 Advantage – quickly coalesce unused chunks into larger chunk
 Disadvantage - fragmentation
113
Buddy System Allocator

114
 Slab Allocator
 Alternate strategy
 Slab is one or more physically contiguous pages
 Cache consists of one or more slabs
 Single cache for each unique kernel data structure
 Each cache filled with objects – instantiations of
the data structure
 When cache created, filled with objects marked as free
 When structures stored, objects marked as used
 If slab is full of used objects, next object allocated from empty slab
 If no empty slabs, new slab allocated
 Benefits include no fragmentation, fast memory request satisfaction

115
Slab Allocation

116
 Slab Allocator in Linux
 For example process descriptor is of type struct task_struct
 Approx 1.7KB of memory
 New task -> allocate new struct from cache
 Will use existing free struct task_struct
 Slab can be in three possible states
 Full – all used
 Empty – all free
 Partial – mix of free and used
 Upon request, slab allocator
 Uses free struct in partial slab
 If none, takes one from empty slab
 If no empty slab, create new empty
117
 Slab Allocator in Linux (Cont.)
Slab started in Solaris, now wide-spread for both
kernel mode and user memory in various OSes
Linux 2.2 had SLAB, now has both SLOB
and SLUB allocators
 SLOB for systems with limited memory
Simple List of Blocks – maintains 3 list objects for small,
medium, large objects
SLUB is performance-optimized SLAB
removes per- CPU queues, metadata stored in
page structure

118
 Other Considerations - Prepaging
 Prepaging
 To reduce the large number of page faults that
occurs at process startup
 Prepage all or some of the pages a process will need, before
they are referenced
 But if prepaged pages are unused, I/O and
memory was wasted
 Assume s pages are prepaged and α of the pages is used
 Is cost of s * α save pages faults > or < than the cost of
prepaging
s * (1- α) unnecessary pages?
 α near zero  prepaging loses

119
 Other Issues – Page Size
 Sometimes OS designers have a choice
 Especially if running on custom-built CPU
 Page size selection must take into consideration:
 Fragmentation
 Page table size
 Resolution
 I/O overhead
 Number of page faults
 Locality
 TLB size and effectiveness
 Always power of 2, usually in the range 212 (4,096 bytes) to 222
(4,194,304 bytes)
 On average, growing over time
120
 Other Issues – TLB Reach
 TLB Reach - The amount of memory accessible from the TLB
 TLB Reach = (TLB Size) X (Page Size)
 Ideally, the working set of each process is stored in the TLB
 Otherwise there is a high degree of page faults
 Increase the Page Size
This may lead to an increase in fragmentation as not all
applications require a large page size
 Provide Multiple Page Sizes
This allows applications that require larger page sizes the
opportunity to use them without an increase in
fragmentation

121
 Other Issues – Program Structure
 Program structure
 int[128,128] data;
 Each row is stored in one page
 Program 1
for (j = 0; j <128; j++)
for (i = 0; i < 128; i++)
data[i,j] = 0;
128 x 128 = 16,384 page faults
 Program 2
for (i = 0; i < 128; i++)
for (j = 0; j < 128; j+
+)
data[i,j] = 0;
128 page faults 122
 Other Issues – I/O interlock
 I/O Interlock Pages must
– sometimes
be memory locked into
 Consider I/O - Pages that are
used for copying a file from a
device must be locked from
being selected for eviction by a
page replacement algorithm
 Pinning of pages to lock into
memory

123
 Operating System Examples

Windows

Solaris

124
 Windows
 Uses demand paging with clustering. Clustering brings in pages
surrounding the faulting page
 Processes are assigned working set minimum and working set
maximum
 Working set minimum is the minimum number of pages the
process is guaranteed to have in memory
 A process may be assigned as many pages up to its working set
maximum
 When the amount of free memory in the system falls below a
threshold, automatic working set trimming is performed to
restore the amount of free memory
 Working set trimming removes pages from processes that have
pages in excess of their working set minimum
125
 Solaris
 Maintains a list of free pages to assign faulting processes
 Lotsfree – threshold parameter (amount of free memory) to
begin paging
 Desfree – threshold parameter to increasing paging
 Minfree – threshold parameter to being swapping
 Paging is performed by pageout process
 Pageout scans pages using modified clock algorithm
 Scanrate is the rate at which pages are scanned. This ranges
from slowscan to fastscan
 Pageout is called more frequently depending upon the amount
of free memory available
 Priority paging gives priority to process code pages

126
Solaris 2 Page Scanner

127
Questions?

128