Author: Robert Hyatt
Date: 11:50:18 07/15/04
Go up one level in this thread
On July 15, 2004 at 11:20:42, Peter Berger wrote: >On July 15, 2004 at 11:11:40, Robert Hyatt wrote: > >>On July 15, 2004 at 03:45:46, Peter Berger wrote: >> >>>On July 14, 2004 at 22:19:12, Robert Hyatt wrote: >>> >>>>>I say "at least 4x", because each of those Opteron processors was faster than my >>>>>2.0GHz AMD 64. In fact I think it was quite more than 4x... >>>> >>>>Bad math. Here is the correct math. >>>> >>>>I had 4 X 2.4ghz. You had 1 X 2.0ghz. >>>> >>>>My SMP speedup is about 3.1X. that gives me 3.1 * 2.4ghz effective speed. IE >>>>7.44ghz. I also lose another 10% overall as that machine did not have all the >>>>working NUMA kernel stuff, so that drops to about 6.7ghz effective speed. >>>> >>>>The difference was really about 3.3x at best... >>>> >>>>Those with numbers that anyone can verify. >>> >>>There is some additional speedup (at least for Crafty, for Falcon I have no >>>idea) from running in 64bit mode. Falcon's computer was running Windows XP. >>> >>>For Crafty the speedup would have indeed been about 4 . >> >>I don't understand. The SMP search definitely is about 3.1x faster. I've >>posted the data to show this and its on my ftp box already. The raw NPS was off >>about 10%, which is the gain possible with the libNUMA things that I could not >>get working on such short notice. IE on a quad 2.4 I should get close to 4x the >>raw NPS. I did not. Each CPU ran about 10% slower on average than a single CPU >>2.4 would run, because of this. >> >>It doesn't really matter how it compares to another program. I actually have >>some logs from some 2.0ghz runs prior to the last CCT, where I ended up on a >>2.2ghz box. That was what I used to do the above computations... > >I assume you used Linux in 64 bit mode on the single-CPU-Opteron too, so you >measured only the real speedup from the additional processors. If Crafty had >used Windows XP on the single Opteron it would have been slower. No. It would be at least 10% faster. I have the actual comparison numbers from the same version, one run by Eugene using his compiler, one run by me on linux using 64 bit gcc. Eugene's compiler ate the gcc compiler for lunch. But you missed the other point as well. This is a NUMA box, but we ran a non-NUMA O/S and a non-NUMA version of Crafty. That cost about another 10% in performance due to non-optimal memory accessing. Crafty in CCT6 was faster if you look at raw NPS than Crafty in the WCCC, even though in CCT6 we ran on 2.2ghz processors. That's the NUMA loss. > Actually we >have no disagreement here - I simply compaired Omid's system as was with ours >and concluded a plausible speedup number for Crafty, as I can't do it for Falcon >for obvious reasons. > >Peter I have numbers for single CPU 2.2ghz vs quad 2.2ghz, but the quad numbers have NUMA support. The quad we used in WCCC did not, which made it a little bit slower than the 2.2ghz quad I used in the last CCT...
This page took 0.01 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.