Author: Sven Reichard
Date: 01:57:58 12/08/03
Go up one level in this thread
On December 07, 2003 at 23:00:10, Robert Hyatt wrote: >On December 07, 2003 at 20:40:51, Sean Mintz wrote: > >>Discussions about running chess programs efficiently on clusters have been going >>on here for as long as I can remember. With the recent (impressive!) work by >>Hyatt and Nalimov, I wonder if this brings us any closer towards efficient >>searches by clusters. Although there are many major differences left, NUMA >>seems to me much closer to clusters than SMP. Are we getting closer, or are the >>differences still to great/many? >> >>Sean Mintz > > >Still a long way apart. NUMA is _still_ shared memory. Clusters are still >message-passing. Handling NUMA is just a special-case of shared memory stuff >that simply offers more performance if done right... A while ago I read about an algorithm specifically geared towards clusters; it had some cryptachrostic name like APHID or something like that... Anyway, it didn't depend on shared memory, just on low bandwidth message passing. The basic idea was the following: The master process searches to a shallow depth (say, 4 ply) and distributes the encountered nodes to the slaves. Each slave investigates its nodes independently, feeding back its findings to the master. The main advantage was that if one slave has finished all work on a given ply, it can immediately continue with the next ply, without waiting for the other slaves to finish. Of course there are still many issues to resolve, such as efficient distribution of the subtrees and providing reasonably search bounds to the slaves. Also, it is not clear how this scales to large clusters. Does anybody have practical experience with this approach? Sven.
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.