Author: Robert Hyatt
Date: 09:43:39 05/31/98
Go up one level in this thread
On May 31, 1998 at 03:32:56, Don Dailey wrote: >On May 30, 1998 at 09:56:25, Robert Hyatt wrote: > >>On May 30, 1998 at 02:28:02, Steven Juchnowski wrote: >> >>>How much does hash table size effect performance? >>> >>>On one hand we hear that Fritz5 can gain about 40 ELO points by >>>increasing the hash size from 44MB to 100MB. >>> >>>On the other hand there are claims that doubling the hash size will >>>only improve performance by a few ELO points. >>> >>>I assume the above claims are based on tournament time controls. >>> >>>So which is the real picture, or is it perhaps that the sensitivity of >>>hash >>>size on performance varies from program to program. >>> >>>Regards >> >> >>there are at least a couple of things that could make Fritz far more >>sensitive to hash table size than other programs: >> >>(1) a poor replacement strategy. If this is true, then a larger table >>reduces replacement, which would produce better performance. >> >>(2) using the table for other things besides the normal score/best move/ >>etc. If this is true, replacing *any* entry could be bad, depending on >>what is stored in the table. >> >>no one knows what Fritz does, but one of the above reasons is almost >>certain to be correct. I'd suspect (2) myself, since replacement >>strategies are well-known now. > >Is Fritz actually more sensitive to hash table size, or does it just >need bigger hash tables than the other programs? My impression >was that it probably hashed into quies and was so fast it needed >bigger tables than most others. > >If Crafty was 100X faster (on the same hardware), would you consider >your program more "sensitive" to hash table size than the others? > >My opinion is that Fritz's speed just makes it need more memory. > >I have a hard time believing reason 2 is likely. I agree with you >that reason 1 is also unlikely. My opinion is that Fritz's speed >just makes it need more memory and this makes it appear to be >"sensitive" to hash table size. > >If you were fed 200 calories a day, people might think you were >calorie sensitive. He seems to perform much better with just >an extra 200 calories added to his diet! > >Let me think some more about your second suggestion as a possibility. > >- Don I don;t think it is an "overflow" issue. IE I ran a few tests last year with Cray Blitz on a T90, searching about 5M nodes per second... and the variability I saw was about 30% from way too small to way too big. So bigger was better, but not a lot better. I suspect they store other things (ala hsu's "sticky transposition table" only more-so) that costs them dearly to compute when they don't find it in the hash... but that's only a "SWAG"...
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.