Computer Chess Club Archives




Subject: Re: Nalimov TB caching ?

Author: Robert Hyatt

Date: 13:58:24 08/06/99

Go up one level in this thread

On August 06, 1999 at 14:11:09, Eugene Nalimov wrote:

>On August 06, 1999 at 08:47:52, Robert Hyatt wrote:
>>On August 06, 1999 at 06:28:55, Ulrich Tuerke wrote:
>>>I am wondering how (and if) to use the TB caching offered by Eugene's
>>>I assume that almost all chess programs use one or more
>>>transposition hash tables, this way caching the results of table base
>>>probes anyway. Isn't it redundant to use the caching of the TB module
>>>(FTbSetCacheSize() call) too ?
>>>If not, what values for the cache size can be suggested ?
>>>Thanks in advance, Uli
>>It's pretty important, unless you run a system like Linux that really does well
>>caching on its own.  However, Eugene's LRU cache is very efficient and  can
>>speed things up significantly.
>>In the case of Crafty, I generally do this:
>>hash = 3/4 of all memory.
>>pawn hash = 1/2 of what is left, up to 16mb.
>>cache = whatever is left, leaving enough room to prevent paging.
>Based on your own experiments, larger TB cache speeds up probing on a Linux,
>too. My guess it's because there is no necessity to go to OS, switch protection
>levels, and copy 8Kb of data to program's space...

Definitely right.  I didn't mean to imply cache was unneeded in Linux...
although linux seems to handle this better than windows 95/98 (I haven't
fiddled with testing NT, which is probably much better than 95/98 with
respect to disk cache).

That was why I suggested 3/4 of memory for hash, up to 16mb for pawn hash,
and the rest for cache...

This page took 0.03 seconds to execute

Last modified: Thu, 07 Jul 11 08:48:38 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.