Author: Odd Gunnar Malin
Date: 00:01:01 04/26/02
Go up one level in this thread
On April 26, 2002 at 01:05:01, Uri Blass wrote: >On April 25, 2002 at 23:14:38, Jeremiah Penery wrote: > >>On April 25, 2002 at 13:33:03, Robert Hyatt wrote: >> >>>take Crafty, use the default hash, and search for (say) 5 minutes, a typical >>>search for 40/2hr time controls. Then double the hash and try again. And >>>double and try again. The time required for the search will vary from 2x-3x >>>(or more)... In middlegame positions. >>> >>>I posted some test results like this several years ago in r.g.c.c when >>>KomputerKorner asked me to run the test to give some real answers on how >>>hashing affected middlegame positions. The answer was more than I thought. >> >>I actually remember it being less than expected - I've found the link: >> >>http://groups.google.com/groups?q=crafty+hash&hl=en&selm=7b1o5j%24qhm%241%40juniper.cis.uab.edu&rnum=3 >> >>[D]3rr1k1/1p2b1pp/p1b1p3/8/PnBBq3/2Q2N2/1PP3PP/R4R1K b - - 0 1 >> >>This position was searched to 11 ply (Crafty version unspecified), here are the >>results for different hash sizes: >> >> hash table sizes >>time bytes entries nodes searched >>------------------------------------------- >>13:13 96kb 6K 157,085,451 >>12:03 192kb 12K 142,633,162 >>10:31 384kb 24K 123,762,238 >> 9:28 768kb 49K 110,838,220 >> 8:38 1.5M 98K 100,802,339 >> 7:48 3M 196K 90,979,000 >> 7:22 6M 392K 85,975,960 >> 6:53 12M 800K 80,347,212 >> 6:32 24M 1.5M 76,465,119 >> 6:22 48M 3.0M 74,738,532 >> 6:13 96M 6.0M 73,253,374 >> 6:05 192M 12.0M 71,581,397 >> 6:03 384M 24.0M 71,156,722 >> >> >>So about 12 doublings in size give a speedup of slightly over 2. That's not >>such a huge amount. :) > >Thenks for the information. >It suggests that 6-7 elo from doubling the hash size is really close to the >truth. > >There is a big difference between 6K and 24M but there is not a very big >difference between 6M and bigger hash tables and in my case I guess that there >is not very big difference between 8M hash for chezzz and bigger hash tables. > >Uri You are probably right. Dennis M. Breuker did a test with several different replacement scheme in his 'Ph.D. thesis: Memory versus Search in Games' and found out that the gain start with much gain but begin to level out around 512K entries (If 1 entry=16byte, 16*512K=8M) and at 1024K or before the gain is only 3% for each doubling. Ref: http://www.breuker.demon.nl/thesis/index.html , Chapter 2, page 35. This is for midlegame positions. Most gain from transposition tables is in the endgame, but here just a little table would gain many plies, so that the curve is maybe even more precipitous for endgames. Odd Gunnar Odd Gunnar
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.