Author: Robert Hyatt
Date: 07:40:24 05/02/04
Go up one level in this thread
On May 02, 2004 at 10:31:00, Djordje Vidanovic wrote: >On May 02, 2004 at 00:04:49, Robert Hyatt wrote: > >>On May 01, 2004 at 22:36:35, Eugene Nalimov wrote: >> >>>On May 01, 2004 at 18:20:09, Vincent Diepeveen wrote: >>> >>>>On May 01, 2004 at 18:10:32, Robert Hyatt wrote: >>>> >>>>>On May 01, 2004 at 15:18:33, Joachim Rang wrote: >>>>> >>>>>>On May 01, 2004 at 15:06:24, Robert Hyatt wrote: >>>>>> >>>>>>>My results... on a quad 700mhz xeon >>>>>>> >>>>>>>1cpu 289 secs 1.0 speedup 294K nps >>>>>>>2cpu 172 secs 1.7 speedup 571K nps >>>>>>>3cpu 127 secs 2.4 speedup 833K nps >>>>>>>4cpu 100 secs 2.9 speedup 1079K nps >>>>>>> >>>>>>>Unfortunately this machine was running nfs for the entire department so it >>>>>>>wasn't totally idle, but the numbers seem within "reason" of what I would >>>>>>>expect. Note also that I didn't run multiple runs and average the times, so >>>>>>>there could be some random noise in the above as always... >>>>>> >>>>>>still below 3 even for such a SMP-Giant like Crafty. >>>>>>For an 8-way machine what do you expect? >>>>>> >>>>>>regards Joachim >>>>> >>>>> >>>>>My estimate is always speedup = 1 + (CPUS - 1) * .7 as a rough number. Some >>>>>positions go higher, some go lower... >>>>> >>>>>So figure about 6x on 8-way... I have tested that myself. 16 and beyond are >>>>>not so clear yet... >>>> >>>>8 ways are kind of NUMA, so forget it for crafty. >>> >>>As well as 32-way Itanium on which I tested Crafty some time ago... >>> >>>Thanks, >>>Eugene >> >> >>You should realize that our pitiful knowledge pales in the presence of the >>world's foremost authority on everything. I really don't see why anyone would >>enter a tournament where his program plays since they have no chance against >>his. >> >>Wait. Has he ever won one for real? >> >>Wait again. Seems that ICC is broken in that it shows that crafty vs diep has >>looked like this (from crafty's perspective): 468 wins. 156 losses. 166 >>draws. So maybe there _is_ hope for the rest of us??? >> >>I've never seen anyone that knows so much about everybody _else's_ programs. Or >>at least that _thinks_ he knows so much about everybody else's programs... >> >>Fortunately the reality of actually playing in events catches up with the >>fantasy-land claims and pronouncements... > > > >I guess it's normal for duals to have somewhat bigger speedup? I think that I am >getting about 1.8 speedup here on my dual Xeon (HT). > >Djordje First, the most important point to remember is that Crafty (and any other parallel search program) is nondeterministic. That means the speedup will vary. As a general rule speedup will be somewhat better with longer searches. This test was limited to 10 plies which took about 4 minutes over 24 positions, or about 10 seconds per position on one CPU. It is possible to find test positions where the speedup is 1.5X. It is possible to find positions where the 2-cpu speedup is way over 2X. That's why a reasonably-sized set of positions is important. Also, as a general rule, speedup varies on different test positions. Based on the DTS paper and results, speedup is slightly better in real games than when computed over random disconnected test positions such as Brato-Kopec. That's why I usually report results here by including what the results came from (in terms of the positions used). The "speedup" is more related to the positions than the search, when you talk about 1.7X vs 1.8X. I have tests where my quad has produced 1.7X for example (single positions). I have run different test sets and seen numbers from 2.5X to 3.5X overall. So there is no "single right answer" contrary to what "some" will suggest here... My oft-quoted speedup = 1 + (NCPUS - 1) * 0.7 is just a fairly reasonable approximation based on _lots_ of testing. It is unlikely that any single test will match that very closely.
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.