Author: Robert Hyatt
Date: 12:49:23 01/15/00
Go up one level in this thread
On January 15, 2000 at 14:08:42, Vincent Lejeune wrote: >On January 15, 2000 at 01:20:28, Robert Hyatt wrote: > >>It is time for my next project. Today I finished the PO for a new beowulf >>cluster machine here. This machine will have 8 nodes, with each node being >>a quad xeon 550mhz machine. The nodes will be connected by a gigabit/sec >>switch. And no, it won't be crafty's permanent machine. But look for it to >>do some interesting matches on ICC later this year when I get to the distributed >>search. :) >> >>This is an interesting 'cluster' since it has 8 machines with 4 cpus per >>node. We are adding 8 more machines this Summer. It will be a horse of >>a box, and the initial 8-node box ought to be able to search at 8M nodes >>per second easily... > >Somes basic questions after this breaking news : > >Will you split the tree as same way done in Crafty or closer to the root ? >Is it important to have some power at each node or can you add a 4x Xeon800 node >later (for example) ? >Will this project keep the name 'Crafty' ? :) > >Waohw, Crafty on the footsteps of Cray Blitz ... First, this will continue to be known as "crafty". As far as the near the root question, I would certainly suspect so, but this will probably be a dynamic run-time decision, as this cluster will talk at gigabit speeds with latency less than 1 microsecond per message. That won't be a common cluster speed obviously. Also this will be a combo distributed/ parallel search algorithm as I worked on in Cray Blitz. I suspect that the distributed stuff will be restricted to reasonably close to the root, but I haven't run on a cluster with this kind of speed/low latency before, so I won't know until I try it... As far as asymmetry with respect to speed, that should be no problem, as the cray blitz search was tested on a C90 (16 cpus, 4+ns clock cycle) plus two YMPs (8 cpus, 6ns cycle times) + several XMPs (4 cpus 8 ns cycle time). It is better to know something about the speed of each cluster node to make scheduling decisions... but over time, this ought to be adaptive...
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.