Author: Gordon Rattray
Date: 09:44:08 01/15/04
Go up one level in this thread
On January 15, 2004 at 11:39:28, Robert Hyatt wrote: [snip] >>So, supposing I have two PCs, with PC1 being faster than PC2 (both single CPU). >>Assume everying else is the same between them (engine, book, etc.). What's the >>best time control for PC1's faster hardware to show it's superiority? > >Superiority in what? Blitz games or long time controls? If you are >interested in blitz, I'd play blitz. If you are interested in 40/2hr, I'd >play 40 moves in 2 hours. I'm interested in PC1 winning as many games as possible over PC2. :-) I'm thinking blitz is the answer (for my simplified conditions). >> >>I think that a faster time control is required, as this has the best chance of >>PC1 gaining a deeper search depth than PC2. The fact that both are using the >>same engine is crucial to my thinking here. I'm also using the belief that >>deeper searches have diminishing returns, although I'm not sure of the latest >>thoughts/data on this. > >Using the same program on both is one thing. I was discussing two different >programs playing a match... Sure. I was highlighting this in order to compare with my main question... the SMP case. I guessed it was easier to simplify the engine factor before varying the hardware. >> >>But now let's complicate things a bit... >> >>Suppose PC1 is a dual CPU. PC2 is a single CPU. PC1 still generally faster >>while using both CPUs, but slower than PC1 if only using one CPU - assume 2 in >>use. Engine still the same, and SMP capable. >> >>Could this change matters? Could the fact that PC1 is running SMP - and hence >>the search is undeterministic - mean that PC1's search is more "hit and miss" >>and that luck starts to play more of a part? Or will the SMP search be lucky >>and unlucky in such a manner that it balances itself out and doesn't matter >>overall? > > >Here is my answer. Run the program on machine 1 (single cpu machine) and >look at the NPS. Run the program on machine 2 using 2 cpus, but here take >the NPS and divide by 2 to convert to 1 cpu number, then multiply by 1.7 >as that will be the rough approximation to the parallel speedup. Compare >that number to the single-cpu machine's NPS. The faster NPS should win, >the wider the margin in speed, the wider the margin in wins. Ok. But if the two machines (same engine) play 10000 games at a "all in 1 minute" control, and then another 10000 at "all in 10 min" control can we expect the win ratio to be the same (allowing for statistical error margins)? As in the "two PCs, both single CPU" example, we'd expect a faster time control to result in a bigger win ratio for the faster PC. But does the inclusion of a SMP machine complicate this? Could an ultra-fast time control result in more "unstable" searches and more "luck" being introduced? >> >>As a side note, I've recently run through some of the Nolot test suite with a >>SMP engine and the range of times (for the same test case) was greater than I >>thought. >> >>Gordon > > >You haven't been reading my posts here very long then. :) I have pointed this >out hundreds of times and given amazing examples of how wildly SMP results can >vary. Even though one or two claim a variance of < 1% for _their_ program... >This makes testing and evaluation of changes _very_ complicated, as you can >see. Maybe I missed it. But it's more likely the figures just didn't hit home with me till I seen it with my own eyes. :-) Gordon
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.