INT404 - Big Data Analytics

Lecture 6 and 7: Algorithms Using Map

Reduce

Most of the slides are from the Mining of Massive Datasets book.
These slides have been modified for Big Data Analytics Course. The original slides can be accessed at: www.mmds.org
Example Application: Join
Join Operation
◼ Compute the natural join R(A,B) ⋈ S(B,C)
◼ R and S are each stored in files
◼ Tuples are pairs (a,b) or (b,c)

A B B C A B C
a1 b1 b2 c1 a3 b2 c1
a2
a3
b1
b2
⋈ b2 c2 = a3 b2 c2
b3 c3 a4 b3 c3
a4 b3
S
R
 Join with MapReduce: Map

◻ Map:
For each input tuple R(a, b):
Generate <key = b, value = (‘R’, a)>
For each input tuple S(b, c):
Generate <key = b, value = (‘S’, c)>
Think of ‘R’ and ‘S’ as bool variables that indicate where the pair originated from

A B Key-value pairs
a1 b1 B C
<b1, (R, a1)> <b2, (S, c1)>
a2 b1 b2 c1 <b1, (R, a2)> <b2, (S, c2)>
a3 b2 b2 c2 <b2, (R, a3)> <b3, (S, c3)>
a4 b3 b3 c3 <b3, (R, a4)>

R S

4
Join with MapReduce: Shuffle & Sort

Output of Map

<b1, (R, a1)> <b2, (S, c1)>

<b1, (R, a2)> <b2, (S, c2)>
<b2, (R, a3)> <b3, (S, c3)>
<b3, (R, a4)>

Input of Reduce

<b1, [(R, a1); (R, a2)]>

<b2, [(R, a3); (S, c1); (S, c2)]>
<b3, [(R, a4); (S, c3)]>

5
 Join with MapReduce: Reduce

◻ Reduce:
Input: <b, value_list>
In the value list:
■ Pair each entry of the form (‘R’, a) with each entry (‘S’, c), and output:
<a, b, c>

Input of Reduce Output of Reduce

<b1, [(R, a1); (R, a2)]> <a3, b2, c1>

<b2, [(R, a3); (S, c1); (S, c2)]> <a3, b2, c2>
<b3, [(R, a4); (S, c3)]> <a4, b3, c3>

6
Example Application:
Matrix-Vector Multiplication
Matrix-Vector Multiplication

M V X

mij . = xi
vj

8
Simple Case: Vector fits in memory

◻ For simplicity, assume that n is not too large and V fits into
main memory of each node.
◻ First, read V into an array accessible from Map tasks
◻ Map:
For each mij, generate <key = i, value = mij * vj>

◻ Reduce:
Input: <key = i, values = [mi1 * v1 ; mi2 * v2 ; … ; min * vn]
Sum up all values, and output <key = i, value = sum>

◻ What if V does not fit into main memory?

Still works, but very slow.
9
 General Case: Vector does not fit into memory

M V

• Vertical stripes for M

. • Horizontal stripes for V
chunk • Each stripe stored in a file.

◻ Each map task:

is assigned a chunk of one of the M files.
reads one stripe of V completely, and stores in local node’s memory.
◻ Map and reduce function definitions same as in previous slide.

10
Example Application:
Matrix-Matrix Multiplication
2 Map-Reduce Steps
Matrix – Matrix Multiplication

j k k

i mij i pik
X =
j njk

M N P

12
Two-Step MapReduce

◻ Two matrices in two separate input files

◻ Two steps of MapReduce:

Step 1: “Join” the matrix entries that need to be multiplied with each other
■ Similar to the Join operation we implemented using MapReduce

Step 2: Accumulate all results and compute the output matrix entries

13
First MapReduce Step

◻ Objective: “Join” mij and njk entries

In this case, “join” corresponds to multiplication

◻ Map:
For each mij value of matrix M
Generate <key = j, value = (“M”, i, mij) >

For each njk value of matrix N

Generate <key = j, value = (“N”, k, njk) >

14
Example: Map Output

j k
0 0 m13 0 0 n12 0 n14
m21 0 0 m24 0 0 0 0
i j
0 0 0 0 0 n32 0 n34
m41 0 m43 0 0 n42 0 0

Map output:
<1, (M, 2, m21)> <1, (N, 2, n12)>
<1, (M, 4, m41)> <1, (N, 4, n14)>
<3, (M, 1, m13)> <3, (N, 2, n32)>
<3, (M, 4, m43)> <3, (N, 4, n34)>
<4, (M, 2, m24)> <4, (N, 2, n42)>

15
 Intuition 1: Joining entries with same j values

j k
0 0 m13 0 0 n12 0 n14
m21 0 0 m24 0 0 0 0
i j
0 0 0 0 0 n32 0 n34
m41 0 m43 0 0 n42 0 0

Map output:
<1, (M, 2, m21)> <1, (N, 2, n12)>
<1, (M, 4, m41)> <1, (N, 4, n14)>
<3, (M, 1, m13)> <3, (N, 2, n32)>
<3, (M, 4, m43)> <3, (N, 4, n34)>
<4, (M, 2, m24)> <4, (N, 2, n42)>

16
 Intuition 1: Joining entries with same j values

j k
0 0 m13 0 0 n12 0 n14
m21 0 0 m24 0 0 0 0
i j
0 0 0 0 0 n32 0 n34
m41 0 m43 0 0 n42 0 0

Map output:
<1, (M, 2, m21)> <1, (N, 2, n12)>
<1, (M, 4, m41)> <1, (N, 4, n14)>
<3, (M, 1, m13)> <3, (N, 2, n32)>
<3, (M, 4, m43)> <3, (N, 4, n34)>
<4, (M, 2, m24)> <4, (N, 2, n42)>

17
 Intuition 2: Partial sums

j k
0 0 m13 0 0 n12 0 n14
m21 0 0 m24 0 0 0 0
i j
0 0 0 0 0 n32 0 n34
m41 0 m43 0 0 n42 0 0

m21 . n12 will contribute to the

<1, (M, 2, m21)> <1, (N, 2, n12)>
partial sum of p22
<1, (M, 4, m41)> <1, (N, 4, n14)>
<3, (M, 1, m13)> <3, (N, 2, n32)>
<3, (M, 4, m43)> <3, (N, 4, n34)>
<4, (M, 2, m24)> <4, (N, 2, n42)>

18
 Intuition 2: Partial sums

j k
0 0 m13 0 0 n12 0 n14
m21 0 0 m24 0 0 0 0
i j
0 0 0 0 0 n32 0 n34
m41 0 m43 0 0 n42 0 0

m24 . n42 will contribute to the

<1, (M, 2, m21)> <1, (N, 2, n12)>
partial sum of p22
<1, (M, 4, m41)> <1, (N, 4, n14)>
<3, (M, 1, m13)> <3, (N, 2, n32)>
<3, (M, 4, m43)> <3, (N, 4, n34)>
<4, (M, 2, m24)> <4, (N, 2, n42)>

19
 Intuition 2: Partial sums

j k
0 0 m13 0 0 n12 0 n14
m21 0 0 m24 0 0 0 0
i j
0 0 0 0 0 n32 0 n34
m41 0 m43 0 0 n42 0 0

m21 . n14 will contribute to the

<1, (M, 2, m21)> <1, (N, 2, n12)>
partial sum of p24
<1, (M, 4, m41)> <1, (N, 4, n14)>
<3, (M, 1, m13)> <3, (N, 2, n32)>
<3, (M, 4, m43)> <3, (N, 4, n34)>
<4, (M, 2, m24)> <4, (N, 2, n42)>

20
 First MapReduce Step: Reduce

Reduce input: <1, [ (M, 2, m21); (M, 4, m41); (N, 2, n12); (N, 4, n14) ] >
<3, [ (M, 1, m13); (M, 4, m43); (N, 2, n32); (N, 4, n34) ] >
<4, [ (M, 2, m24); (N, 2, n42) ] >

Reduce(key, value_list):
Put all entries in value_list of form (M, i, mi,key) into LM
Put all entries in value_list of form (N, k, nkey,k) into LN
for each entry (M, i, mi,key) in LM
for each entry (N, k, nkey,k) in LN
output <key = (i, k); value = mi,key . nkey,k >

21
 Example: Reduce Output

Reduce input:
<1, [ (M, 2, m21); (M, 4, m41); (N, 2, n12); (N, 4, n14) ] >
<3, [ (M, 1, m13); (M, 4, m43); (N, 2, n32); (N, 4, n34) ] >
<4, [ (M, 2, m24); (N, 2, n42) ] >

Reduce output:

< (2, 2), (m21.n12) > < (1, 2), (m13.n32) >
< (2, 4), (m21.n14) > < (1, 4), (m13.n34) >
< (4, 2), (m41.n12) > < (4, 2), (m43.n32) >
< (4, 4), (m41.n14) > < (4, 4), (m43.n34) >

< (2, 2), (m24.n42) >

22
 Second MapReduce Step: Map

Map input:
< (2, 2), (m21.n12) > < (1, 2), (m13.n32) >
< (2, 4), (m21.n14) > < (1, 4), (m13.n34) >
< (4, 2), (m41.n12) > < (4, 2), (m43.n32) >
< (4, 4), (m41.n14) > < (4, 4), (m43.n34) >

< (2, 2), (m24.n42) >

◻ Map:
for each (key, value) pair in the input
generate (key, value)
◻ Identity function
◻ The system will most likely assign the map tasks on the same node as
the reduce that produced these outputs. Hence, no communication cost.
23
Second MapReduce Step: Reduce

Reduce input:
< (2, 2), (m21.n12) > < (1, 2), (m13.n32) >
< (2, 4), (m21.n14) > < (1, 4), (m13.n34) >
< (4, 2), (m41.n12) > < (4, 2), (m43.n32) >
< (4, 4), (m41.n14) > < (4, 4), (m43.n34) >

< (2, 2), (m24.n42) >

◻ Reduce(key, value_list):
sum = 0
foreach v in value_list
sum += v
output <key, sum>

24
 Example: MapReduce Step 2 - Reduce

j k
0 0 m13 0 0 n12 0 n14
m21 0 0 m24 0 0 0 0
i j
0 0 0 0 0 n32 0 n34
m41 0 m43 0 0 n42 0 0

Reduce output:
< (2, 2), (m21.n12) > < (1, 2), (m13.n32) >
< (2, 4), (m21.n14) > < (1, 4), (m13.n34) >
< (4, 2), (m41.n12) > < (4, 2), (m43.n32) >
< (4, 4), (m41.n14) > < (4, 4), (m43.n34) >

< (2, 2), (m24.n42) >

25
 Example: MapReduce Step 2 - Reduce

j k
0 0 m13 0 0 n12 0 n14
m21 0 0 m24 0 0 0 0
i j
0 0 0 0 0 n32 0 n34
m41 0 m43 0 0 n42 0 0

Reduce output:
< (2, 2), (m21.n12) > < (1, 2), (m13.n32) >
< (2, 4), (m21.n14) > < (1, 4), (m13.n34) >
< (4, 2), (m41.n12) > < (4, 2), (m43.n32) >
< (4, 4), (m41.n14) > < (4, 4), (m43.n34) >

< (2, 2), (m24.n42) >

26
 Example: MapReduce Step 2 - Reduce

j k
0 0 m13 0 0 n12 0 n14
m21 0 0 m24 0 0 0 0
i j
0 0 0 0 0 n32 0 n34
m41 0 m43 0 0 n42 0 0

Reduce output:
< (2, 2), (m21.n12) > < (1, 2), (m13.n32) >
< (2, 4), (m21.n14) > < (1, 4), (m13.n34) >
< (4, 2), (m41.n12) > < (4, 2), (m43.n32) >
< (4, 4), (m41.n14) > < (4, 4), (m43.n34) >

< (2, 2), (m24.n42) >

27
 Example: MapReduce Step 2 - Reduce

j k
0 0 m13 0 0 n12 0 n14
m21 0 0 m24 0 0 0 0
i j
0 0 0 0 0 n32 0 n34
m41 0 m43 0 0 n42 0 0

Reduce output:
< (2, 2), (m21.n12) > < (1, 2), (m13.n32) >
< (2, 4), (m21.n14) > < (1, 4), (m13.n34) >
< (4, 2), (m41.n12) > < (4, 2), (m43.n32) >
< (4, 4), (m41.n14) > < (4, 4), (m43.n34) >

< (2, 2), (m24.n42) >

28
Summary: Two-Step MapReduce Algorithm

◻ Step1: Map (input):

For each mij value of matrix M
generate <key = j, value = (“M”, i, mij) >
For each njk value of matrix N
generate <key = j, value = (“N”, k, mjk) >
◻ Step1: Reduce(key, value_list):
for each entry (M, i, mi,key) in value_list
for each entry (N, k, nkey,k) in value_list
output <key = (i, k); value = mi,key . nkey,k >
◻ Step2: Map (key, value):
generate (key, value)
◻ Step2: Reduce(key, value_list):
sum ← accumulate the values in value_list
output (key, sum)
29
Example Application:
Matrix-Matrix Multiplication
Single Map-Reduce
Single-Step MapReduce: Intuition

◻ To compute pik, we need mij and njk values for all j.

◻ In other words:
mij entry is needed to compute pik values for all k.
njk entry is needed to compute pik values for all i.

◻ Intuition: Send each input matrix entry to all reducers that need it.

31
An Entry of Matrix M

j k

i mij
X
j njk

M N

◻ Each mij needs to be paired with all entries in row j of matrix N

32
An Entry of Matrix N

j k

i mij X
j njk

M N

◻ Each njk needs to be paired with all entries in column j of matrix M

33
Map Operation

◻ Reminder:
mij entry is needed to compute pik values for all k.
njk entry is needed to compute pik values for all i.

◻ Map:
for each mij entry from matrix M:
for k=1 to n
generate <key = (i, k), value = (‘M’, j, mij) >
for each njk entry from matrix N:
for i=1 to n
generate <key = (i, k), value = (‘N’, j, njk) >

34
 Example: Map Output for Matrix M Entries

0 0 m13 0 0 n12 0 n14

m21 0 0 m24 0 0 0 0
. =
0 0 0 0 0 n32 0 n34
m41 0 m43 0 0 n42 0 0

Map output:
<(1,1), (M, 3, m13) >

35
 Example: Map Output for Matrix M Entries

0 0 m13 0 0 n12 0 n14

m21 0 0 m24 0 0 0 0
. =
0 0 0 0 0 n32 0 n34
m41 0 m43 0 0 n42 0 0

Map output:
<(1,1), (M, 3, m13) >
<(1,2), (M, 3, m13) >

36
 Example: Map Output for Matrix M Entries

0 0 m13 0 0 n12 0 n14

m21 0 0 m24 0 0 0 0
. =
0 0 0 0 0 n32 0 n34
m41 0 m43 0 0 n42 0 0

Map output:
<(1,1), (M, 3, m13) >
<(1,2), (M, 3, m13) >
<(1,3), (M, 3, m13) >
<(1,4), (M, 3, m13) >

37
 Example: Map Output for Matrix M Entries

0 0 m13 0 0 n12 0 n14

m21 0 0 m24 0 0 0 0
. =
0 0 0 0 0 n32 0 n34
m41 0 m43 0 0 n42 0 0

Map output:
<(1,1), (M, 3, m13) > <(2,1), (M, 1, m21) >
<(1,2), (M, 3, m13) > <(2,2), (M, 1, m21) >
<(1,3), (M, 3, m13) > <(2,3), (M, 1, m21) >
<(1,4), (M, 3, m13) > <(2,4), (M, 1, m21) >

38
 Example: Map Output for Matrix M Entries

0 0 m13 0 0 n12 0 n14

m21 0 0 m24 0 0 0 0
. =
0 0 0 0 0 n32 0 n34
m41 0 m43 0 0 n42 0 0

Map output:
<(1,1), (M, 3, m13) > <(2,1), (M, 1, m21) > <(2,1), (M, 4, m24) >
<(1,2), (M, 3, m13) > <(2,2), (M, 1, m21) > <(2,2), (M, 4, m24) > …
<(1,3), (M, 3, m13) > <(2,3), (M, 1, m21) > <(2,3), (M, 4, m24) >
<(1,4), (M, 3, m13) > <(2,4), (M, 1, m21) > <(2,4), (M, 4, m24) >

39
 Example: Map Output for Matrix N Entries

0 0 m13 0 0 n12 0 n14

m21 0 0 m24 0 0 0 0
. =
0 0 0 0 0 n32 0 n34
m41 0 m43 0 0 n42 0 0

Map output:
<(1,2), (N, 1, n12) >

40
 Example: Map Output for Matrix N Entries

0 0 m13 0 0 n12 0 n14

m21 0 0 m24 0 0 0 0
. =
0 0 0 0 0 n32 0 n34
m41 0 m43 0 0 n42 0 0

Map output:
<(1,2), (N, 1, n12) >
<(2,2), (N, 1, n12) >
<(3,2), (N, 1, n12) >
<(4,2), (N, 1, n12) >

41
 Example: Map Output for Matrix N Entries

0 0 m13 0 0 n12 0 n14

m21 0 0 m24 0 0 0 0
. =
0 0 0 0 0 n32 0 n34
m41 0 m43 0 0 n42 0 0

Map output:
<(1,2), (N, 1, n12) > <(1,4), (N, 1, n14) >
<(2,2), (N, 1, n12) > <(2,4), (N, 1, n14) >
<(3,2), (N, 1, n12) > <(3,4), (N, 1, n14) >
<(4,2), (N, 1, n12) > <(4,4), (N, 1, n14) >

42
 Example: Map Output for Matrix N Entries

0 0 m13 0 0 n12 0 n14

m21 0 0 m24 0 0 0 0
. =
0 0 0 0 0 n32 0 n34
m41 0 m43 0 0 n42 0 0

Map output:
<(1,2), (N, 1, n12) > <(1,4), (N, 1, n14) > <(1,2), (N, 3, n32) >
<(2,2), (N, 1, n12) > <(2,4), (N, 1, n14) > <(2,2), (N, 3, n32) > …
<(3,2), (N, 1, n12) > <(3,4), (N, 1, n14) > <(3,2), (N, 3, n32) >
<(4,2), (N, 1, n12) > <(4,4), (N, 1, n14) > <(4,2), (N, 3, n32) >

43
Summary: Map Operation

<key=(i,1), value = (M, j, mij)>

…
mij

<key=(i,n), value = (M, j, mij)>

<key=(1,k), value = (N, j, njk)>

…

njk

<key=(n,k), value = (N, j, njk)>

44
Reduce Operation

◻ Input:
key = (i,k), value_list = [ … (M, j, mij); … (N, j, njk) … ]
an entry exists for any non-zero mij or njk
◻ Objective: Multiply mij and njk values with matching j values, and sum
up all products to compute pik
◻ Reduce(key, value_list)
put all entries of form (M, j, mij) into LM
sort entries in LM based on j values
put all entries of form (N, j, njk) into LN
sort entries in LN based on j values
sum ← 0
for each pair (M, j, mij) in LM and (N, j, njk) in LN
sum += mij . njk
output (key, sum)

45
 Example: Reduce

0 0 m13 0 0 n12 0 n14

m21 0 0 m24 0 0 0 0
. =
0 0 0 0 0 n32 0 n34
m41 0 m43 0 0 n42 0 0

Reduce input: key = (4, 2), value_list = { (M, m41, 1); (M, m43, 3);
(N, n12, 1); (N, n32, 3); (N, n42, 4) }

Reduce output: key = (4, 2), value = m41. n12 + m43. n32

46
Summary: Single-Step MapReduce Algorithm

◻ Map(input):
for each mij entry from matrix M:
for k=1 to n
generate <key = (i, k), value = (‘M’, j, mij) >
for each njk entry from matrix N:
for i=1 to n
generate <key = (i, k), value = (‘N’, j, njk) >

◻ Reduce(key, value_list)
sum ← 0
for each pair (M, j, mij) and (N, j, njk) in value_list
sum += mij . njk
output (key, sum)

47
 References

◻ Jure Leskovec, Anand Rajaraman, Jeff Ullman, Mining of Massive

Datasets, Cambridge University Press, Second Edition, 2014.
◻ http://mmds.org/
◻ http://www.cs.bilkent.edu.tr/~mustafa.ozdal/cs425/

48