Author: Gerd Isenberg
Date: 14:14:34 07/24/03
Go up one level in this thread
On July 23, 2003 at 18:54:39, Vincent Diepeveen wrote: >Hello, > >Dr. Hyatt suggests to compare the difference between using big hashtables and >small hashtables in order to see how much of the system time is going to >hashtables. > >In itself not such a bad idea under a few conditions > a) Pawnhashtable should be put to 64KB at most > b) Hashtable should be put to around 64KB at most too > c) in not a single case the pawnhashtable should cause a cutoff > d) in not a single case the hashtable should give a cutoff nor > a bestmove to search first > >So basically the hashtables get called but do not give cutoffs. It is trivial >that bigger hashtables mean higher cutoff %. Those cutoffs in itself are a very >small % (in case of transpositiontable) but they take care that the search >process searches in the same game space. > >No one can deny this. Especially not Dr Hyatt. He has written similar >observations in old articles of himself in ICCA journal. > >So complaining that the above conditions aren't fair won't help if this reveals >that using big hashtables is quite a bit slower than small hashtables. > >For the big test conditions c and d as above apply and for > a1) I advice 32M for pawnhashtable > a2) i advice 768M for hashtable > >This basically only measures the influence of hashtables. Not even from the big >rotated bitboard move generation which is a megabyte or so if not more in size. > >Note that this experiment is giving just a part of the truth. Just hashtables. >There is very good tools to exactly measure the real price one pays to RAM. One >of them you can download for free. If you have intel processors then intel has >stuff for free to download, if you have AMD processors then AMD has stuff of his >own too. > >Best regards, >Vincent Hi Vincent, quick try with IsiChess on Athlon xp2.8+, not using move or score from hash with 8-fold replacement scheme: 64K Hash 506kn/s 514k/ps 312M Hash 493kn/s 503k/ps with 5r1k/1P4pp/3P1p2/4p3/1P5P/3q2P1/Q2b2K1/B3R3 w - - bm=a2f7 ; bt2630-4 2bq3k/2p4p/p2p4/7P/1nBPPQP1/r1p5/8/1K1R2R1 b - - bm=c8e6 ; bt2630-13 so about 2%-2.6%, but still few MB Pawn- and Eval-Hash. What about Diep? Regards, Gerd
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.