Author: Brian Richardson
Date: 05:57:01 03/29/03
Go up one level in this thread
I have tested Tinker with up to 2GB hash tables and seen similar trends, which of course vary with search time and NPS rate. Perhaps the code pages are competing too much with hash table memory blocks. Bob posted some interesting somewhat related data about the resiliency of the search despite hashing "errors", IIRC. It is disturbing to me when tactical positions are solved significantly faster with _smaller_ hash tables. Intuitively this should not be the case. Unless one is passing the whole table between iterations or something, I suspect that it is due to other search bugs, but have not been able to track them down. I have a 6GB dual Itanium2 system to test with, but have not been able to tweak things to enable larger 4+GB table support. I have also done some informal tests between hash replacement schemes. At least for Tinker, one depth preferred table is generally "better" than the two table (depth preferred and always replace) approach. Incidentally, Tinker hashes in q-search. I suspect the older ICCA hashing replacement article was limited by the configurations at that time (1M entries or 1MB, I don't recall which).
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.