Author: Vincent Diepeveen
Date: 08:41:03 09/26/01
Go up one level in this thread
On September 26, 2001 at 11:39:05, Vincent Diepeveen wrote: please also print for each depth the number of nodes you need! depths 12/13 with R=2 or R=3 you get within a few seconds or so. All the 2.0 speedups i get here i never get within a few seconds. it's always levels which i can't realize in tournaments. >On September 26, 2001 at 11:19:04, Robert Hyatt wrote: > >>On September 26, 2001 at 10:48:14, Vincent Diepeveen wrote: >> >>>On September 25, 2001 at 23:44:09, Robert Hyatt wrote: >>> >>>you didn't turn off futility bob and i cannot see outputs >>>in number of nodes and such to see whether sequential overhead >>>mattered! >> >> >>I don't do futility pruning. I can't turn it off... >> >> >> >>> >>>Kopec positions are the worst testpositions ever to use anyway >>>for obvious reasons that the score only goes up and up and up. >> >> >>Not true. Crafty changes its mind several times per test position... >> >>The data shows that the speedup is pretty constant regardless of null-move >>or not. It would _also_ be constant with or without futility. Older versions >>had futility pruning and/or razoring. And it had no effect on the parallel >>search. >> >>If you want to make a wager, I will find the CB test positions and run >>them in the same way. You will see _no_ change in my speedup results, >>however... > >please run the positions and write down the speedups for 8,9,10,11,12 ply >of course for both 1 and 2 processors. 4 processors frommy viewpoint is >less interesting. > >game positions are *better* than positions where a few strategical >principles dominate and where some of the positions are too easy >to be true. you only get a score higher and higher! > >What you did here is compare apples with peanuts. depth 9 compared to >depth 10. That's stupid. > >What we need is the crafty OUTPUT simply from each run at this game! > >Easy to automate too. > >> >>> >>>>Here are the results. First the tests. I took the last 8 kopec positions, >>>>and searched them to a fixed depth for each null-move setting. 9 plies for >>>>R=0, 10 plies for R=1, 11 plies for R=2 and 12 plies for R=2~3. >>>> >>>>I ran the tests with 1, 2 and 4 processors, and computed the speedup for >>>>each. The data: >>>> >>>>null move R=0----------------------------- >>>> 1cpu 2cpu 4cpu >>>>pos17 115 67 40 >>>>pos18 267 146 77 >>>>pos19 61 32 17 >>>>pos20 106 56 30 >>>>pos21 126 71 36 >>>>pos22 116 63 33 >>>>pos23 108 59 31 >>>>pos24 337 176 90 >>>> sum 1236 670 354 >>>> S/U 1.0 1.8 3.5 >>>> >>>> >>>>null move R=1----------------------------- >>>> 1cpu 2cpu 4cpu >>>>pos17 42 22 15 >>>>pos18 76 34 21 >>>>pos19 32 16 9 >>>>pos20 35 20 11 >>>>pos21 30 15 9 >>>>pos22 51 28 16 >>>>pos23 68 36 19 >>>>pos24 144 74 40 >>>> sum 478 245 140 >>>> S/U 1.0 1.9 3.4 >>>> >>>> >>>>null move R=2----------------------------- >>>> 1cpu 2cpu 4cpu >>>>pos17 39 19 12 >>>>pos18 121 55 18 >>>>pos19 27 16 8 >>>>pos20 34 19 13 >>>>pos21 20 11 6 >>>>pos22 43 22 12 >>>>pos23 58 29 15 >>>>pos24 83 44 28 >>>> sum 425 215 112 >>>> S/U 1.0 1.9 3.8 >>>> >>>> >>>>null move R=2~3--------------------------- >>>> 1cpu 2cpu 4cpu >>>>pos17 67 41 26 >>>>pos18 265 99 60 >>>>pos19 36 21 12 >>>>pos20 90 52 27 >>>>pos21 40 26 15 >>>>pos22 74 41 21 >>>>pos23 107 66 39 >>>>pos24 194 106 51 >>>> sum 873 452 251 >>>> S/U 1.0 1.9 3.5 >>>> >>>> >>>> >>>>The conclusions: >>>> >>>>1. Crafty gets roughly 1.9X faster using two processors, regardless of >>>>the null-move setting. R=0 (no null move at all) to r=2-3, the most >>>>aggressive setting I use. >>>> >>>>2. It averaged a 3.5 speedup for 4 cpus, with R=2 having a slightly better >>>>speedup for random reasons. >>>> >>>>3. Null-move has _zero_ influence on the speedup of a parallel search, as I >>>>have said _many_ times. All this nonsense about saying that the old programs >>>>got better speedups without null-move, or better speedups with null-move is >>>>total baloney. >>>> >>>>Anybody else is free to run the same tests... But I prefer to do things a >>>>bit scientifically by running a test, rather than wild speculation without >>>>any testing at all. >>>> >>>>I can provide the raw data if needed, but it would be a very large post since >>>>I ran the tests several times to average them.
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.