Author: Dann Corbit
Date: 11:52:51 12/22/03
Go up one level in this thread
On December 20, 2003 at 08:23:24, Vincent Diepeveen wrote: >On December 19, 2003 at 22:05:11, Dann Corbit wrote: > >>On December 19, 2003 at 21:07:40, Uri Blass wrote: >> >>>I am interested if programmers can do the following experiment. >>> >>>Take the GCP test suite(delete endgame positions from it) and test your program >>>twice. >>> >>>one time with 1 mbyte hash tables and one time with 2 mbytes hash tables. >>> >>>of course 2 mbytes is better but the question is if 2 mbytes gives 6-7% speed >>>improvement for all problems (except easy problems that are solved in less time >>>that is needed for the program to fill the entries of 2 mbytes hash table) or if >>>it gives bigger improvement for hard problems that the program needs some >>>minutes to solve. >>> >>>I asked to use very small hash tables in order to enable programs to fill the >>>hash tables in a very small time. >>> >>>I agree that bigger hash table probably means better branching factor at the >>>time that is enough only to fill the small hash and not enough to fill the >>>bigger hash but the question is what happens later. >> >>The curve looks like c0*log(c1*x+1), where c0 anbd c1 are constants and x is >>hash size. Y is performance increase. Stupendous at first, and never goes >>away, but the added strenght becomes less and less. >> >>You can make too large of a hash table if you clear hash frequently (e.g. >>between moves) and the hash clear time becomes significant. > >I see a formula above. Can you please write down your formula out for next 2 >experiments: > >experiment A) I ran diep at 460 processors with 115MB hashtable *in total* >experiment B) Same diep version at 460 processors with 115GB hashtables. > >Note hashtable means transpositiontable here. Each processor had local 4.2MB >pawnhashtable and each processor had local 32MB evaluation table. > >Please write down according to your formula how much deeper B was supposed to >search than A after 10 hours of search. I have no idea. I have never seen experimental data for MP search. I have never made experiments for that model of MP search. Therefore, any sort of extrapolation would be nothing more than a wild guess. What have you measured with your program?
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.