Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: What's the difference between NUMA, SMP and MPI for chess?

Author: Roberto Nerici

Date: 08:55:33 04/15/04

Go up one level in this thread


On April 15, 2004 at 11:26:28, Joachim Rang wrote:

>On April 15, 2004 at 08:59:24, Robert Hyatt wrote:
>
>>
>>NUMA -> Non-Uniform Memory Access. Just like SMP except that each CPU has local
>>memory that can be accessed faster than memory on other processors.
>>
>>MPI -> Message Passing Interface.  It is a message passing library similar to
>>PVM that works mainly on clusters.  Those not knowing what they are doing might
>>use it on a NUMA box but it is not the best approach there.
>>
>>NUMA offers some problems in that it is more efficient for a CPU to access some
>>parts of memory than it is to access others.  If you don't plan for this, you
>>simply run slower than optimal.
>
>
>So for a chessengine a SMP-System is the best and a NUMA-System must be
>addressed with special code, right? MPI works for cluster but cluster are not
>efficient for chess because of latency and sharing problems, right?

Not necessarily. The NUMA-system needs special code to work at its best, but it
may well (probably will in fact) work better than the SMP-system *if* it has the
code to address the NUMA issues. Also NUMA may become more attractive as the
number of CPUs increases; it is likely to scale better than SMP.

As for clusters, two single-processor machines (making a cluster) will probably
give poor performance compared to a 2-cpu SMP machine, but two 2-cpu machines
clustered is obviously a possibility against one 2-cpu machine.

Roberto/.





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.