Author: Robert Hyatt
Date: 07:28:09 11/28/04
Go up one level in this thread
On November 27, 2004 at 18:05:07, Barry Culp wrote: >I know there was a long thread earlier on hyperthreading and engines that >support multiple processors. I ran some tests from the Chessbase Testsets folder >using the Marathon base test 1622 ECM, game #208 (solution ..Nxh7). My system is >a P4 3.6ghz with 2gb RAM. All tests used Shredder 8, 1gb hash, default engine >options except learning disabled and style=aggressive. I ran each test twice to >see if results were consistent. > >Test 1 - Hyperthreading disabled in BIOS - Shredder threads = 1 > time 11:53 ply 18/50 nodes 303621 > time 11:39 ply 18/50 nodes 303621 > >Test 2 - Hyperthreading enabled in BIOS - Shredder threads = 2 > time 10:45 ply 18/43 nodes 335338 > time 11:08 ply 18/46 nodes 345877 > >Test 3 - Hyperthreading enabled in BIOS - Shredder threads = 1 > time 11:27 ply 18/50 nodes 303621 > time 11:35 ply 18/50 nodes 303621 > >I was thinking off turning hyperthreading OFF on my system but after these tests >I dont think I will :) > >Barry Your test is flawed. To see how it helps, run the program with 1 thread and with 2 threads, over a _set_ of test positions. Look at the elapsed time for HT on and HT off. Often a parallel search will produce varying times anyway, and since hyper-threading only offers maybe 30% increase in NPS for programs that have some severe memory issues, it is possible that with threads=2 that the NPS will go up somewhat, but the time to solution will go up as well, which is not desired. You have to run a set of positions and look carefully. For Crafty HT=off is better, because the memory / cache footprint was tweaked to be more efficient on AMD NUMA boxes and as a result, hyper-threading doesn't add that much speed, and the parallel search loss will make the overall search speed slower. Also, hyper-threading is moot for a program that doesn't support threads. Your numbers suggested slightly faster search with hyper-threading on. That means something is broken in your testing methodology, that simply isn't possible. Unless you are running two things at once, in which case none of your numbers will be useful anyway...
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.