Mr Share 11 Sep 2010

Post on 09-Jul-2015

1.520 views 4 download

description

Large-scale data analysis lies in the core of modern enterprises and scientific research. With the emergence of cloud computing, the use of an analytical query processing infrastructure (e.g., Amazon EC2) can be directly mapped to monetary value. MapReduce has been a popular framework in the context of cloud computing, designed to serve long running queries (jobs) which can be processed in batch mode. Taking into account that different jobs often perform similar work, there are many opportunities for sharing. In principle, sharing similar work reduces the overall amount of work, which can lead to reducing monetary charges incurred while utilizing the processing infrastructure. In this paper we propose a sharing framework tailored to MapReduce. Our framework, MRShare, transforms a batch of queries into a new batch that will be executed more efficiently, by merging jobs into groups and evaluating each group as a single query. Based on our cost model for MapReduce, we define an optimization problem and we provide a solution that derives the optimal grouping of queries. Experiments in our prototype, built on top of Hadoop, demonstrate the overall effectiveness of our approach and substantial savings.

Transcript of Mr Share 11 Sep 2010

MRShare: Sharing Across Multiple Queries in MapReduce

Tomasz Nykiel (University of Toronto)

Michalis Potamias (Boston University)

Chaitanya Mishra (University of Toronto, currently Facebook)

George Kollios (Boston University)

Nick Koudas (University of Toronto)

1

Data management landscape

efficiency

flex

ibili

ty

σ π

2

• Time performance

• Arbitrary data• Large scale setups

MRShare – sharing framework for MR

MRShare – a sharing framework for Map Reduce

• MRShare framework:

– Inspired by sharing primitives from relational domain

– Introduces a cost model for Map Reduce jobs

– Searches for the optimal sharing strategies

– Does not change the Map Reduce computational model

3

Outline

• Introduction

• Map Reduce recap.

• MRShare – Sharing primitives in Map-Reduce

• MRShare – Cost based approach to sharing

• MRShare Evaluation

• Summary

4

Outline

• Map Reduce recap.

5

Map Reduce recap.

I

I

I

I

Map Reduce

6

Outline

• MRShare - Sharing primitives in Map-Reduce

7

Sharing primitives – sharing scans

• SELECT COUNT(*) FROM user GROUP BY hometown

• SELECT AVG(age) FROM user GROUP BY hometown

8

User_id Hometown Occupation Age

MRShare – sharing scans (map).

9

MRShare – sharing scans (reduce)

J1 J2 J3 J4 key value

Toronto 1

Toronto 1

Toronto 1

Toronto 17

Toronto 19

Toronto 2

Toronto 5

10

Outline

• MRShare - Sharing primitives in Map-Reduce

– Sharing scans

– Sharing intermediate data

11

Sharing primitives - Sharing intermediate data.

• SELECT COUNT(*) FROM user WHERE occupation=‘student’ GROUP BY hometown

• SELECT COUNT(*) FROM user WHERE age > 18 GROUP BY hometown

User_id Hometown Occupation Age

Occupation ?= ‘student’ Age ?> 18

12

MRShare – sharing intermediate data (map).

13

MRShare – sharing intermediate data (reduce).

J1 J2 J3 J4 key value

Toronto 1

Toronto 4

Toronto 1

Toronto 1

Toronto 2

Toronto 2

Toronto 5

14

Outline

• MRShare – Cost based approach to sharing

– Cost model for finding the optimal sharing strategy

– SplitJobs – cost based algorithm for sharing scans

– MultiSplitJobs – an improvement of SplitJobs

– γ-MultiSplitJobs – the algorithm for sharing intermediate data

15

Cost model for Map Reduce (single job)

• Reading – f(input size)

• Sorting – f(intermediate data size)

• Copying – f(intermediate data size)

• Writing – f(output size)

16

Cost of executing a group of jobs

17

Finding the optimal sharing strategy

18

• An optimization problem

“NoShare”

“GreedyShare”

Outline

• MRShare – Cost based approach to sharing

– Cost model for finding the optimal sharing strategy

– SplitJobs – cost based algorithm for sharing scans

– MultiSplitJobs – an improvement of SplitJobs

– γ-MultiSplitJobs – the algorithm for sharing intermediate data

19

Sharing scans - cost based optimization

• Savings come from reduced number of scans• The sorting cost might change• The costs of copying and writing the output do not

change

20

• We prove NP-hardness of the problem of finding the optimal sharing strategy

SplitJobs – a DP solution for sharing scans.

• We reduce the problem of grouping to the problem of splitting a sorted list of jobs – by approximating the cost of sorting.

22

• Using our cost model and the approximation, we employ a DP algorithm to find the optimal split points.

Outline

• MRShare – Cost based approach to sharing

– Cost model for finding the optimal sharing strategy

– SplitJobs – cost based algorithm for sharing scans

– MultiSplitJobs – an improvement of SplitJobs

– γ-MultiSplitJobs – the algorithm for sharing intermediate data

23

MultiSplitJobs – an improvement of SplitJobs

24

Outline

• MRShare – Cost based approach to sharing

– Cost model for finding the optimal sharing strategy

– SplitJobs – cost based algorithm for sharing scans

– MultiSplitJobs – an improvement of SplitJobs

– γ-MultiSplitJobs – the algorithm for sharing intermediate data

25

Sharing intermediate data - cost based optimization

• The sorting and copying costs change – depending on the size of the intermediate data

26

We need to estimate the size of the intermediate data of all combinations of jobs.

γ-MultiSplitJobs – the solution for sharing intermediate data

• Approximate the size of the intermediate data

27

• γ –MultiSplitJobs – applies MultiSplitJobs with modified cost function

• γ set heuristically

Outline

• MRShare Evaluation

28

Evaluation setup

• 40 EC2 small instance virtual machines

• Modified Hadoop engine

• 30 GB text dataset consisting of blogs

• Multiple grep-wordcount queries

– Counts words matching a regular expression

– Allows for variable intermediate data sizes

– Generic aggregation Map Reduce job

29

Evaluation goals

• Sharing is not always beneficial.

– ‘GreedyShare’ policy

• How much can we save on sharing scans?

– MRShare - MultiSplitJobs evaluation

• How much can we save on sharing intermediate data?

– MRShare - γ-MultiSplitJobs evaluation

30

Is sharing always beneficial?- ‘GreedyShare’ policy

Group of jobs

Group size

d=|intermediate data| / |input data|

H1 16 0.3 < d <0.7

H2 16 0.7 < d

H3 16 0.9 < d

31

How much we save on sharing scans –MRShare MultiSplitJobs

Group of jobs

Group size

d=|intermediate data| / |input data|

G1 16 0.7 < d

G2 16 0.2 < d < 0.7

G3 16 0.0 < d < 0.2

G4 16 0.0 < d < max

G5 64 0.0 < d < max

32

How much we save on sharing intermediate data -

MRShare - γ-MultiSplitJobs

33

Group of jobs

Group size

d=|intermediate data| / |input data|

G1 16 0.7 < d

G2 16 0.2 < d < 0.7

G3 16 0.0 < d < 0.2

Summary

• We introduced MRShare – a framework for automatic work sharing in Map Reduce.

• We identified sharing primitives and demonstrated the implementation thereof in a Map Reduce engine.

• We established a cost model and solved several work sharing optimization problems.

• We demonstrated vast savings when using MRShare.

34

Thank you!!!

Questions?

35

Ongoing work – sharing expensive computation

• Sharing across multiple Map Reduce jobs with expensive predicates.

36

Ongoing work – dynamic sharing

• Dynamic sharing.

37

time

pro

gres

s