Distributed MapReduce Team B Presented by: Christian Bryan Matthew Dailey Greg Opperman Nate Piper Brett Ponsler Samuel Song Alex Ostapenko Keilin Bickar.

Distributed MapReduce Team B

Presented by:• Christian Bryan• Matthew Dailey• Greg Opperman• Nate Piper• Brett Ponsler• Samuel Song• Alex Ostapenko• Keilin Bickar

Introduction

Functional Languages

What makes MapReduce Special?

Map function Lisp – McCarthy et al in 1958 Reduce function (paper example) = Summing up occurrences The combination? Behind the scenes action One user to n computers, where the only insight into n is the speed at which computation is completed.

Example of an abstraction(define appendEven (lambda (x) (cond ((empty? x) empty) (else (begin (cond ((= 0 (remainder (car x) 2)) (cons (car x) (appendEven (cdr x)))) (else (cons (* 2 (car x)) (appendEven (cdr x)) ))))))))

(define appendEvenMap (lambda (x) (cond ((= 0 (remainder x 2)) x) (else (* 2 x)))))

(appendEven myL)

(map appendEvenMap myL)

(list 2 2 6 4 10 6 14 8 18 10 22 12)(list 2 2 6 4 10 6 14 8 18 10 22 12)

SCALABLE???

Goals of Distributed System• Transparency

• Scalable

• More fault tolerant than standalone system

• Monotonicity – Can’t retract statements

• Which computer is correct?

• Many points of failure

Problems when scaling

Naturally Distributable

Why?

• The 'map' and 'reduce' functions themselves.• 'map' takes in a function and a set of data.• That set of data is partitioned and ready to go.• Function + Data = Convenient

More...

• 'reduce' is less convenient.• Takes in an operation and a dataset.• GFS helps out alot.

Distributing Map and Reduce

• User writes Map functiono (k1,v1) → list(k2,v2)

• Next, user rights Reduce programo (k2,list(v2)) → list(v2)

• Specification file defines inputs, outputs, and tuning parameterso Passed to MapReduce function

• MapReduce library handles the rest!

Productivity Improvements

• Programmers no longer have to program for the network

• Simplified library to make a program distributed, can be reused

• Can focus on problem instead of distributed implementation of it

• Quote from Google: "Fun to use" oProgrammers having fun are more

productive

MapReduce Performance

Measured Performance

• ~1800 2GHz processors with 4Gb of RAM used

• First test task – search through ~1Tb of data for a particular pattern

• Second test task – sort ~1Tb of data

Test 1 (searching)

• Input split into 64Mb pieces• Machines assigned until all are working @55sec• Sources of delay: startup, opening files, locality

optimization

Test 2 (sorting)

Criticism from Database Systems Community

• Very old concepts used• Poor implementation (indices)• Limited set of features (idea of views)

Fault Tolerance

Worker Failure

• Master pings workers periodically• Worker “fails” if it does not respond within

a certain amount of time.• All map tasks completed or in progress by

worker are reset to idle state • Eligible for rescheduling

Worker Failure

• Completed Reduce tasks are not reset because their output is stored in a global file system and not locally on the Failed Machine.

• All workers are notified of the changes in workers

• Resilient to large scale worker failure.

Master Failure

• Periodic checkpoints• Upon Failure: a new copy starts from last

checkpoint• Failure of master is unlikely• Current implementation aborts upon

Master Failure

Google Cluster Configuration

• Large clusters of commodity PCs connected together with switched Ethernet

• Typically dual-processor x86 processors running Linux, 2-4 GB of memory

• Inexpensive IDE disks attached directly to individual machines

• Commodity networking hardware is used. Typically either 100 megabits/second or 1 gigabit/second at the machine level

Google Cluster Operation

• Users submit jobs to a scheduling system. Each job consists of a set of tasks, and is mapped by the scheduler to a set of available machines within a cluster.

• A distributed file system (GFS) is used to manage the data stored on the disks.

• Uses replication to provide availability and reliability on top of unreliable hardware.

Networking

Cost Efficiency

• April 2004, Google spent about $250 million on hardware equipmento includes other equipment than CPUs such as

routers and firewallso Approximately

63, 272 machines 126,554 CPUs 253, 088 GHz of processing power 126,544 Gb of RAM 5,062 TB of Hard Drive Space About 253 teraflops (trillion floating point

operations per second)

Cost Efficiency

• January 2005, Japan's NEC's Earth Simulator supercomputero$250 milliono41 teraflops

• Much more expensive compared to a large cluster of personal computers

Cost Efficiency

• 2003, Virgina Tech used 1,100 Apple computersocost $5 milliono10 teraflopso3rd most powerful at the timeosupercomputer would have cost much

more

Cost Efficiency

• Disadvantagesodeal with network bandwidthoconstantly monitor for hardware failure

Conclusion

Questions?