Author: Robert Hyatt
Date: 09:01:02 09/10/03
Go up one level in this thread
On September 09, 2003 at 16:32:30, Vincent Diepeveen wrote: >On September 08, 2003 at 18:14:56, Robert Hyatt wrote: > >>On September 08, 2003 at 17:32:45, Matthew Hull wrote: >> >>>. >> >> >>Don't get too excited. Using a dual Opteron is easy enough. But using >>a cluster is another thing. IE this won't be a 24-node NUMA machine. It >>will be a cluster of 24 dual-cpu NUMA boxes but the nodes will be connected >>with something that does message-passing. Gigabit initially but something >>faster later (ie the current cluster uses the cLAN hardware but cLAN is >>no longer a player in that arena). Myrinet is not bad. But there are >>better options. > >DIEP runs NUMA, so diep runs within 1 week at a cluster too if i get an offer to >run at a cluster. Just simplistic replacement of shared memory references by MPI >calls. That's all. Scales till around 4000 processors i guess. Could be 8000 >too. > >Oh MPI has some overhead of course that slows it down, but that's trivial. > >Best regards, >Vincent Aren't you the _same_ person that has for years claimed that using a network (100mbig or gigabit) simply will _not_ produce any reasonable speedup for chess? I can re-post some of your claims if you have forgotten. A cluster is a cluster, which means _no_ shared memory. You are confusing many things. MPI is a message passing protocol. Using it on a NUMA machine is a last resort for someone that doesn't know what they are doing. NUMA supports shared memory, but it has to be used correctly within the application to avoid remote memory accesses that are slow. Using MPI is a ridiculous kludge to maximize local memory accesses. There are _much_ better ways to do this. Just ask someone who knows...
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.