cots hardware platform (commercial off-the-shelf)sokocalo.engr.ucdavis.edu/~jeremic/cots hardware...
TRANSCRIPT
COTS Hardware Platform(Commercial Off-The-Shelf)
Boris JeremicDepartment of Civil and Environmental Engineering
University of California, Davis
MRCCS/NSF Summer School
High Performance Computing in Finite Element Analysis
1st - 5th September 2003,
University of Manchester
Supported in part by the NSF, PEER, Caltrans, and Cal–EPA.
Collaborators: Professors Mike Kleeman (UCD), Drs. Francis McKenna (UCB), and graduate
and undergraduate students Ritu Jain (UCD), Guanzhou Jie (UCD), Mark Olton (UCD), Kevin
Murakoshi (UCD).
Jeremic, Manchester, Sept. 2003 1
Motivation
• In house parallel platform for learning, development, production
runs
• Inexpensive (save 30% per box compared to a commercial PC,
save 60commercial PC cluster)
• Maintenance (hardware, software) takes time (money)
• It’s fun and students love to be able to break it and fix it
Jeremic, Manchester, Sept. 2003 2
Parallel Computer System
• Computer performance steadily increase – Price decreasing
• Beowulf parallel computer systems: off–the–shelf PC components
• Performance comparable to commercial supercomputers
• Available in every networked computer environment (PCs, Macs,
UNIX)
• GeoWulf at UC Davis (+ CML cluster + FatCat cluster)
node
001
node
002
node
003
node
004
node
005
node
006
node
007
node
008
node
009
node
010
node
011
node
012
node
013
node
014
node
015
node
016
EthernetFastSwitch
40 portsco
mpu
ter
serv
ice
com
pute
rco
ntro
ller
com
pute
rco
ntro
ller console
outside world Internet
Jeremic, Manchester, Sept. 2003 3
GeoWulf
• 2 Controller computers: dual AMD 2400 machine with 2GB RAM
and 50 and 70GB disk space (recent upgrade)
• Service machine (large disk space (IDE) for backups, can do work
as well...)
• Node computers (heterogeneous):
– 8 x single PIII 400 machines with 128MB RAM and 7GB disks
– 8 x single AMD 600 machines with 128MB RAM and 9GB disks
• Connection: fast Ethernet switch (100T, HP ProCurve 4000).
• Cost: < $500 per node + $50 per switched port + $3500 per
controller
Jeremic, Manchester, Sept. 2003 4
GeoWulf
• In house assembly
• space issues
Jeremic, Manchester, Sept. 2003 5
GeoWulf: Maintenance
Jeremic, Manchester, Sept. 2003 6
GeoWulf: Space Problem
Jeremic, Manchester, Sept. 2003 7
Commercial Alternative
Jeremic, Manchester, Sept. 2003 8
Summary
• It is worth the effort
• Source code compatible with large DMP machines (can be used
for large scale runs)
• If funding permits, go for packaged deals (scyld, linux networx...)
• Always save some funds to build couple machines in house (for fun
and to show students that there is nothing special about piece of
hardware...)
Jeremic, Manchester, Sept. 2003 9
References
[1] Sterling, T. L., Salmon, J., Becker, D. J., and Savarese, D. F. How to Build a Beowulf: A Guideto the Implementation and Application of PC Clusters. Scientific and Engineering Computations Series. The MITPress, 1999. ISBN 0-262-69218-X ; QA 76.58.S854 1998.
Jeremic, Manchester, Sept. 2003 10