Author: Robert Hyatt
Date: 21:10:22 04/15/03
Go up one level in this thread
On April 15, 2003 at 09:03:21, Vincent Diepeveen wrote: >On April 15, 2003 at 04:54:11, Tony Werten wrote: > >>On April 14, 2003 at 17:43:12, Robert Hyatt wrote: >> >>>On April 14, 2003 at 17:15:41, Vincent Diepeveen wrote: >>> >>>>On April 13, 2003 at 22:39:39, Robert Hyatt wrote: >>>> >>>>>On April 13, 2003 at 11:49:28, Vincent Diepeveen wrote: >>>>> >>>>>>On April 13, 2003 at 11:27:53, Robert Hyatt wrote: >>>>>> >>>>>>I said initially. It drops back to 10 splits a second in DIEP after a while. >>>>>>Search depth matters. >>>>>> >>>>>>Let's compare 2 things. >>>>>> >>>>>> time=45.98 cpu=464% mat=0 n=37870294 fh=88% nps=823k >>>>>> ext-> chk=638414 cap=249442 pp=9588 1rep=32966 mate=223 >>>>>> predicted=0 nodes=37870294 evals=14565859 >>>>>> endgame tablebase-> probes done=0 successful=0 >>>>>> hashing-> trans/ref=28% pawn=93% used=28% >>>>>> SMP-> split=431 stop=57 data=6/64 cpu=3:33 elap=45.98 >>>>>> >>>>>>MT 2 crafty 18.10 which i have here. 431 splits at 45 seconds. I guess you must >>>>>>limit in crafty the number of splits a lot as splitting is expensive in crafty >>>>>>when compared to the costs of a single node. >>>>> >>>>>I'm not sure how expensive it is compared to a node. I'll run a test where >>>>>I do the split overhead at every node to compare, however... >>>>> >>>>> >>>>> >>>>>I don't limit them at all. The only limit is the YBW algorithm. But I split >>>>>at the root also, which reduces them signficantly... >>>> >>>>I can split at the root nowadays, but i have turned it off for diep. it gives >>>>too poor speedup for me. The interesting thing which searching SMP can give is >>>>transpositions at a big depth which possibly are overwritten by a sequential >>>>search. i don't want to miss them. >>> >>>Maybe you don't split at the root correctly. I limit this with some intelligent >>>guesswork, so that if it appears that I might change my mind this iteration, >>>then >>>I don't split at the root until I have searched all moves that I think might >>>replace >>>the best move... >> >>Just trying to understand. Are you talking about the case where the best move in >>the root got a fail low ? > >this has not so much to do with alfabeta but with how you parallel split. > >After the best move has been searched (YBW property) so after you have a >mainline, then the rest of the moves can be searched in parallel according to >YBW algorithm. Bob is doing that in crafty (with usually 3 moves for the YBW >property and not 1, as nullmove counts too as a move. I'm using 2, that includes >nullmove). No I'm not doing it that way. I have explained this before. When I complete an iteration, I look at the node counts for each root move (I keep them separately so that I can re-order the root moves based on this node count.) If any look promising (large node count compared to others) then I search it before I drop into YBW and search the rest in paralel. The number of moves I do this to is not fixed. It depends on how many root moves produce big node counts... > >I'm not doing that in Diep as it gives me a bad parallel performance. For >fritz&co it didn't give a good speedup either. > >Of course i did mainly dual cpu tests here. not many 16 processor tests. I can >get a 16 processor NUMA machine for > 90% busy effectively even with 2 cpu's at >1 position, so no need at all to split at root at all. So no need to try if it >already gives a poor speedup at 2 processors. It only gives a poor speedup if you do it _wrong_. > >However with 128 processors or so it might be a good idea to re-experiment with >this as it is hard to get cpu's busy. > >Crafty on other hand doesn't have a very good speedup at 2 processors and the >price of a split is expensive, so probably Bob concluded something else than i >did. I'm doing pretty well at 2 processors... > >> >>When that happens, your testresults indicate that's it's better to split lower >>than to search 2 rootmoves parallel in order to get an established score asap ? >>( So not breaking off seacrh when 1 gets a first failhigh, but only when the >>score is resolved ) > >>Tony >> >>> >>> >>>>
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.