Author: Robert Hyatt
Date: 11:55:40 02/19/04
Go up one level in this thread
On February 19, 2004 at 14:25:07, Dann Corbit wrote: >On February 19, 2004 at 14:14:11, Robert Hyatt wrote: >[snip] >>Clusters can work. 1000 nodes is a totally different problem from using 1-64, >>but nothing I see says it is an impossible task. Impossible to "some" perhaps >>so. Lots of things are impossible to "some". Fortunately, for the rest of us, >>what is "impossible" is something that just takes a bit longer to do... :) > >A cluster notion (e.g. 64000 machines maybe using the network): > >Divide the machines into clumps that work on forward nodes, and multiple >possible best moves. I already do this on large SMP platforms. IE crafty has a "group" concept, so that with (say) 64 nodes, you don't get 64 nodes working on one split point. That is bad for multiple reasons. Usually there are not 64 moves to search so the split overhead is just pure overhead. Also, on NUMA boxes, you don't want 64 processors banging on one "split block" even though they don't do it that much unless the split is too deep in the tree (this is also controllable in current crafty). I suspect the same idea will be the right way for clusters. Divide it into groups, sub-groups, and so forth. The "groups" have a "head" and the initial split is between the "group heads". Then each group head can further split its tree farther down the path, with sub-group members. Of course each of these might be a "sub-group-head" that could later further split the tree among sub-sub-group members. I think some such hierarchy is necessary for truly large numbers of processors, and the potential for search overhead is going to be significant. > >For instance, on the first move of the game, there are maybe 7 moves that are >statistically better than the others. So divide the cluster into 7 clumps and >have each cluster work on one forward move. For each of these clumps, have the >same technique take place until you are out of processors. > >If you can gain 5 plies by that technique, it would be worthwhile.
This page took 0.01 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.