15
Beowulf Cluster, building cheap powerful hpc By Akash Sahoo

Dad i want a supercomputer on my next

Embed Size (px)

Citation preview

Beowulf Cluster, building cheap powerful hpcBy Akash Sahoo

Beowulf

www.top500.org

Speed

•16 Pentium Ppro 200Mhz each•2GB Memory•1.25 Glops•130 Mflops individual speed•Made by Caltech in ‘96

FLOPS (or flops or flop/s, for floating-point operations per second) is a measure of a computer's performance, especially in fields of scientific calculations. It is instruction per second

i7-920 has 40Gflops

In 2007 a Microwulf Cluster (less nodes) with just 4 nodes gave a performance of 26.25 Gflops (Wikipedia)

Why SpeedComputer simulations are far cheaper and faster than

physical experimentsComputers can solve a much wider range of problems

than specific laboratory equipment canComputational approaches are only limited by

computer speed and memory capacity, while physical experiments have many practical constraints

Uses: Predictive Modeling and Simulations, Engineering Design and Automation, Military, Visualization

Why BeowulfCommodity Hardware, No need to throw old PCs !!Ease of addition of nodesWork if nodes failVVVV Cheap (Supercomputer costs million)World Fastest is a cheap Beowulf ClusterCan be used in Homes, educational InstitutesStudents will study and write parallel, optimized codeSimulation Weather, BLAST (comparing primary

biological sequence information), BIONIC (Cancer), distributed Matlab, DNA, Image processing, Web server, own software

Software can run on one/many node as specifiedINDIA has outdated supercomputer - PARAM

Parallel ComputingShared Memory NFSMessage PassingEfficiencyLatencyTopologyAdmin and maintainaceRedundancyWorkload, Unbalancing, ArchEnergy, cost, scalable

Building ItInstall Linux in each node (MPICH2), Can be build

on Windows (PVM)SSH EquivalenceSSH Hostkeys – RSA, DSAInstallation of PDSHSetting NFSCreation of MPI usersInstallation of compat-g77,gccMPI install, LINPACK BenchmarkOSCAR, ROCKS, SCARY- maintainance

Mpdboot –n 3

Starting cluster

Mpdtrace

Know which nodes are running

Mpicc

Compiling prog

Mpiexec –n 1 ./prog

Execute a sample prog on 1 node

Same prog on 2 nodes

Same prog on 3 nodes

Improvements Add fibre optic cableDual Lan cardPowerful Gateway, RouterFaster computers – Cost High, Performance HighKeep in mind-Total memory, Memory bandwidth,

Storage bandwidth, System interconnect latency, The effect of interconnect topologies such as meshes, cubes, and fat trees on application design

SSD Harddisk, RAID, Graphics Card

NotesGave a performance boost of > 200% on 3 nodesOriginality : Security both RSA, DSA. PDSH makes

run commands in all nodes at onceCommercialization: giving manuals on prog, software,

setting cluster, maitainaceProj done in 3rd year

Patents Dynamically allocated cluster system Charles Evert Prael  ;  (Patent No: US5093912)The invention provides a system and method for managing clusters of parallel processors 

for use by groups and individuals requiring supercomputer level computational power.Given Scheduling algo PSB, job queue… 28 claims mostly on scheduling

Method and apparatus for securing and managing cluster David Bruce Kumhyr 2002

ReferenceBeowulf Cluster Computing with Linux (2001)- Thomas Sterling

White papers by HP - Considerations in Specifying Beowulf Clusters 2000MPI: The Complete Reference , Ewing Lusk 

?????????

THANK YOU