Author: Robert Hyatt
Date: 18:09:57 12/23/03
Go up one level in this thread
On December 23, 2003 at 18:45:21, Peter Berger wrote: >On December 23, 2003 at 11:32:12, Robert Hyatt wrote: > >>On December 23, 2003 at 06:39:22, Peter Berger wrote: >> >>>On December 22, 2003 at 22:59:01, Robert Hyatt wrote: >>> >>Several _possible_ machines. One here in Alabama, 64-way Itanium. Another >>was a 64-way alpha. Another choice was a 16-way opteron. I really didn't >>investigate what I might use, because if you read the rules, they _really_ >>want the author present, else it gets even _more_ expensive to enter which is >>silly. >> >>There were other possibilities for machines, but it is pointless to start >>bugging people about "what if" to see if a machine might be usable, when I >>really was not considering going due to the length of the event.. >> >>The 16-way opteron is probably the slowest of the above machines, and it >>would ring in at 32M+ nodes per second, up to 44M+ in endgames, based on >>the 8-11M nps I saw on the 4-way box I played with for a while on ICC. > >It seems you have done additional work on parallel speed-up with many >processors. I did some tests a longer time ago (17.11 ??) on Sparc platform ( I >know you really hate it) with many processors, and Crafty *really* didn't scale >too well when the number exceeded 4. The point here is that the sparc is really lousy anyway. :) However, the opteron, the alpha, and the Itanium boxes look pretty good. I've not tried any HP stuff in years so I don't know what they've done with their PA processor. However, they have always been good at technology, so I would assume they have done their homework also... The main issue here was not large numbers of processors, it was/is NUMA architectures. However, Crafty (at present) won't work well on all NUMA architectures, until I actually get to run on them and figure out what their NUMA API is for things like local/global/interleaved memory allocation, setting processor affinity so that a thread runs where its local data was allocated, etc... > >Of course I am aware that this is a purely theoretical discussion. > >The ICGA requirement to be present as an author doesn't seem to get enforced >that strongly anyway. Sending a book author would be sufficient I suppose. The ICGA seems to have a _real_ problem with their rules, IMHO. Either you have them and enforce them, or you get rid of them. You don't have them but not enforce them, that makes no sense. However... :) > >Peter
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.