Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: The Strongest Linux Supercomputer !

Author: Jay Urbanski

Date: 23:17:12 08/08/03

Go up one level in this thread


On August 08, 2003 at 07:53:57, Vincent Diepeveen wrote:

>On August 07, 2003 at 17:45:24, Jorge Pichard wrote:
>
>>http://www.cpuplanet.com/news/article.php/2242421
>
>putting a sporthal full of dual xeon or dual opteron machines is not very
>difficult for a government of course. Also very cheap.
>
>The question is how they connect the nodes to each other. That's the pricey part
>of a supercomputer. Then they need to get to work 4 more things to call it a
>supercomputer IMHO.

Myrinet "D" adatpers for the Opteron portion, dual-rail Myrinet "E" adapters for
the Itanium portino.

>First of all OpenMP and MPI. Secondly an OS that without crashing runs at it
>and i bet that default (experimental) NUMA linux kernel will go only to 64
>processors so i wonder how you can
>call this worlds largest supercomputer anyhow if each accessible node is just 64
>cpu's which right now is the maximum linux can work with.

Of course it supports OpenMP and MPI.  As for an OS, it's standard Linux. The
biggest SMP node is a 4-way for the Itaniums, all the Opterons are two-ways.
You don't need esoteric kernels unless you are doing large SMPs, which this is
not.  Most "Supercomputers" are not NUMA systems, they are clusers of small
SMP's.  Take a look at #3 on the current TOP500 list.  Nothing but dual
processor Xeons runing Linux.

>Third is they need a huge bandwidth and fast latency and in such a way that we
>can work with more than 64 processors at a time. We can put in of course 100mbit
>network cards and 2 routers or so then we have to work 32 cpu's.
>
>but suppose we want to run a job of 500 cpu's or something similar and we cannot
>let each cpu communicate fast with the other cpu's.
>
>You still want to call that a supercomputer?

It is by any definition of the word.  But it will probably only be #4 or #5 by
the time it's installed.

>Now the only good thing we know is that the clusters of IBM usually are not so
>bad.

Thanks. :)



This page took 0 seconds to execute

Last modified: Thu, 15 Apr 21 08:11:13 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.