Author: Robert Hyatt
Date: 20:11:12 03/15/03
Go up one level in this thread
On March 15, 2003 at 02:46:47, Tony Werten wrote: >On March 14, 2003 at 12:14:32, Robert Hyatt wrote: > >>I am attaching the SMP results for 1, 2 and 4 cpus, for the first 12 BS2830 >>positions. >>Vincent apparently wanted to pick and choose to produce a set of positions >>Crafty does >>badly on. I prefer to take a random set of positions, in this case the first 12 >>BS2830's >>without any attempt to "pick the best or the worst." >> >>Here's the results, including the edited log files. I deleted all output for >>searches >>less than 9 plies deep, and I deleted most of the statistics (EGTB probes, SMP >>stuff, >>etc) to make it as short as possible. It still turns out to be a _long_ post. >> >>Here we go: >> >>(this was also posted to rec.games.chess.computer since Vincent likes to make a >>fool of >>himself there as well as here).Here is the data produced by running the first 12 >>BS2830 >>positions thru Crafty, using one, two and four processors >>on a quad 550 xeon box. The first 12 lines are the raw times >>needed to search to a depth of ten plies for each of the 12 >>positions, using 1, 2 and 4 processors. >> >>Raw times: >> >> 1cpu 2cpus 4cpus >> 1. 19.9 15.5 11.9 >> 2. 171.0 132.0 79.0 >> 3. 90.0 51.4 28.3 >> 4. 3.5 2.1 1.4 >> 5. 78.0 42.8 21.0 >> 6. 0.4 0.2 0.2 >> 7. 0.4 0.2 0.2 >> 8. 39.0 19.0 16.2 >> 9. 18.2 9.9 5.5 >>10. 119.0 59.4 36.5 >>11. 29.2 15.4 8.2 >>12. 72.0 40.3 23.3 >> >>Speedup per position: >> >> 2cpu 4cpu >> 1. 1.3 1.7 >> 2. 1.3 2.2 >> 3. 1.8 3.2 >> 4. 1.6 2.6 >> 5. 1.8 3.7 >> 6. 2.1 2.3 >> 7. 1.8 1.6 >> 8. 2.0 2.4 >> 9. 1.8 3.3 >>10. 2.0 3.3 >>11. 1.9 3.5 >>12. 1.8 3.1 >> ----- ----- >> avg 1.8 2.7 >> >>The above average weighs every position equally. An alternative >>way is to add all the times in the columns, and then divide the >>one cpu total by the two cpu total (or the 4 cpu total) which >>weights the positions that took longer, higher. That gives the >>following speedup numbers. >> >> 2cpu 4cpu >> avg 1.6 2.8 > >This would suggest you have to adjust your speedup formula to 1+(cpu-1)*0.6 > >(IIRC correctly it was 0.7 ) > >Tony > Maybe or maybe not. You _do_ understand that the above is _one_ test run for ten positions? And I _have_ explained _many_ times, that SMP performance is neither (a) absolute nor (b) repeatable. IE for the above 10 positions, run once, your .6 may fit better, dependin on how you measure speedup. Notice that it _also_ produced 1.8 if you take each position independently. Which is how most _everybody_ does the computation. But a different set of positions will produce a different number. Which is why it is ridiculous to say that such a formula is either (a) absolutely accurate or (b) incorrect. There is _no_ formula that will work for all sets of test data. And I _do_ mean "no formula". The 1 + .7*(NCPUS-1) is an _estimate_. If you are looking for an absolute formula, you need to get away from SMP search altogether. It is _not_ that reproducible.
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.