Author: John Merlino
Date: 15:56:01 03/17/02
Go up one level in this thread
On March 17, 2002 at 16:13:39, Dieter Buerssner wrote: >On March 17, 2002 at 15:55:01, John Merlino wrote: > >>Once again, a fabulous report and many thanks to everybody who was involved in >>running this tournament. It was a joy to watch and was incredibly exciting as it >>neared the finish. > >I agree! > >>One correction, though. In your report, you state: >> >>------------------------ >>It must be emphasized that in all our tests and the 4500 games we have played at >>40'/40, it was the original TheKing312 engine that has been used that means >> >>TheKing 3.12 >>123 KB (125'952 Byte) >>18th October 2000 >> >>and neither TheKing 3.12c [124 KB=126'976 Byte, 12th February 2001] nor TheKing >>3.12d [124 KB=127'488 Byte, 5th July 2001]. We do not know whether this may be >>of any significance regarding playing strength as the three versions should more >>or less be identical. As far as we know, the orginal engine [18th October 2000] >>does clear the hashtables after each move whereas the versions 3.12c and 3.12d >>keep the hashtables in memory. >>------------------------ >> >>This is not correct. The King 3.12c and 3.12d also clear their hashtables after >>every move. However, they clear them much FASTER. This was in response to many >>users who were using 32MB or even 64MB of hash table memory in blitz games >>(obviously, useless). > >I think, this is not obvious. I believe, quite some engines can fill (say) 32 Mb >of hashtables in Blitz games. This is certainly true for some other engines, but not for The King. The King, as far as NPS goes, is relatively slower than most other engines. And that, coupled with clearing the hash tables after every move, means that a very large hash table is unnecessary, even in tournament time controls! The last time Johan entered The King in a tournament (40/120 time controls) , he had an Athlon 1.2GHz machine and only used a 15MB hash table. Therefore, in a blitz game, it is almost certainly useless to have anything more than 8MB for The King. >>The original King 3.12 could take as long as 2 seconds to >>clear a 64MB hash table (on a slower machine), which could really be a problem >>in a blitz game (i.e. losing 2 seconds per move). So, Johan found a way to clear >>the hash tables about 30 times faster in version 3.12c. > >Very interesting. 2 seconds for 64 Mb, 1 second for 32 Mb then. 30 times faster >means about 1 Gb/s. Hmmm - this really sounds impossible, if you mean the same >by "clear" than I do. I can imagine to make this situation (two large HTs for >this engine at this fast time control) much faster - with some compromise. One >easy method would be, to have a move counter in the HTs, and allways ignore HT >entries with a non up to date move counter. This would be essentially the same >as clearing. But then, "30 times faster" would make no sense. Because the effort >for changing and comparing the move counter would be very dependant on the >actual hash table size and the time control. What I mean - with 32 Mb HTs and 1 >second per game, the factor would be infinity, while it gets smaller with longer >time control. With 2 min + 12 s (which many will consider as Blitz), this factor >would be much smaller than 2. So, you really seem to mean some real clearing. >Now, would you enlighten us, how this can be done (the 1 Gb/s)? >If not - I can understand it as well ... Well, here's what Johan said in his e-mail to me concerning the update: ------------------ CHANGED cleaning of transposition table. Old: resetting all the entries (bound by RAM bandwith). New: resetting array of valid bits, resetting entries only when needed. The bit array is only 1/320 of the TT size. ------------------ Basically, this means that 1/320th of the amount of memory is being cleared when the hash tables are being cleared. It was my independent testing (on a PIII-600) which led me to the 30-fold speed increase. This may have been influenced by the debugging version that I was running, or many other factors. jm p.s. I was also wrong about one small thing. This fix was only in The King 3.12d; it was not in 3.12c.
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.