Author: Robert Hyatt
Date: 17:09:52 10/08/02
Go up one level in this thread
On October 08, 2002 at 19:22:44, Omid David wrote: >On October 08, 2002 at 14:37:39, Robert Hyatt wrote: > >>On October 08, 2002 at 14:29:02, Rex wrote: >> >>>I agree that dual or more processors are faster. My question is the effeciency >>>of search may be decrease. The weight of a given move may be decreased when its >>>way down the tree. The differences of move weight gets less efficiant when the >>>tree search gets deeper and deeper. Calculating what would be the best move >>>based on score may not be the best move at all. Figuring this calculation deep >>>in the search tree is probably imposible until that particular move arrives 1 to >>>2 moves deep. Then it may be too late. >> >> >>By faster, I don't mean "faster NPS only". >> >>I mean that if it takes one 1ghz processor 3 minutes to search to depth X, then >>the two >>1ghz machine should take significantly less than 3 minutes to reach that same >>depth X. >> >>For Crafty, the rough number is about 1.7. IE if it takes 3 minutes for one >>cpu, then it will >>take around 1.7 minutes to do the same search using two cpus. For Crafty, the >>NPS will >>be closer to 2x faster using two cpus, but the "search efficiency" you mentioned >>drops >>the overall performance gain down to 1.7X or so... (again, this is an _average_ >>number, >>for those that like to question it. Some positions speed up more than 2.0 times >>faster, >>other speed up significanly less than 2.0 times faster. The overall _average_ >>seems to >>hover around 1.7 for Crafty. YMMV on other programs or on different hardware.) > > >And how will Crafty perform on say 8 processors? Is there any limit for number >of processors, or the more the processors the better the performance of Crafty? I have run it on 16, but I have been unable to test it enough to see what happens to the speedup= 1 + (N-1)*.7 formula. I'm pretty sure it won't hold true up at that end of the number of processors. It works fine for 1-4, and it seemed to work ok for an 8-cpu test I ran, but non-intel-based hardware. I will answer this question one day. But one thing is for sure, I don't ever expect to see more = worse performance, although it is possible that more != better performance...
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.