[Gllug] Thanks for the awesome talk

John Hearns hearnsj at googlemail.com
Fri May 14 12:46:05 UTC 2010

The very large systems with high core counts in the Top 500 lists -
from SGI and from other manufacturers - are Beowulf style clusters.
You take many identical small(ish) systems and install them - either
1U systems in racks, or blade systems.
you boot Linux onto them, and 'farm out' a program to run on them all
- they communicate across a fast message passing system,
which can be Infiniband, Gigabit ethernet, 10 gigabit ethernet, or
less common now Quadrics or Myrinet.
So you get 10 000 plus cores all working on the same problem by
running one executable on multiple systems and having it pass messages
across to the others.

The Altix UV system which was being demonstrated last night was a new
system from SGI - this is based on Nehalem EX processors,
and you can boot a single system with up to 2048 real processor cores
(4096 hyperthread).
As Eng Lim demonstrated last night - you can log in and type 'cat
/proc/cpuinfo' and get 2048 entries flashing past!
Gllug mailing list  -  Gllug at gllug.org.uk

More information about the GLLUG mailing list