Author: Robert Hyatt
Date: 21:16:06 04/01/03
Go up one level in this thread
On April 01, 2003 at 21:29:29, Vincent Diepeveen wrote: >On April 01, 2003 at 16:45:53, Robert Hyatt wrote: > >>Three runs. First with one cpu, second with two threads, hyper-threading turned >>off. Final >>with hyperthreading on and using four threads: >> >>log.001: time=1:25 cpu=100% mat=0 n=85541803 fh=91% nps=1002k >>log.001: time=55.76 cpu=100% mat=0 n=62194368 fh=95% nps=1115k >>log.001: time=1:40 cpu=100% mat=-1 n=89351765 fh=94% nps=892k >>log.001: time=1:17 cpu=100% mat=0 n=82359320 fh=92% nps=1056k >>total time used = 317.76s >> >> >>log.002: time=58.39 cpu=198% mat=0 n=93417000 fh=91% nps=1599k >>log.002: time=32.12 cpu=199% mat=0 n=57789717 fh=95% nps=1799k >>log.002: time=1:01 cpu=195% mat=-1 n=88101174 fh=94% nps=1431k >>log.002: time=55.31 cpu=197% mat=0 n=92187038 fh=92% nps=1666k >>total time used = 206.82 1.54X SMP speedup (two threads/processors) > >I of course prefer logfiles instead of such outputs. > >diep@xs4all.nl > >Yet interesting to see is that you have a very poor hardware speedup in nodes a >second with 2 processes at the dual Xeon. > >Can you shed some light onto what explains that? OK, I just noticed what you are talking about. No, I have no idea. I booted the machine with SMT off, and ran two threads. I don't know why the NPS is so low, but I will play with this further by running two separate copies with SMT off to see if both copies run at normal speed or if both slow down due to some sort of interaction. > >I ask this because i remember some statement not so long ago which claimed 1.9x >speedup or something at the dual Xeon. That is correct with SMT on. 4 threads, vs 1 thread, which is not a real fair comparison but there is little to be done about it. > >Considering the hardware speedup it is of course amazing to see that the SMP >speedup is not much under that. I wasn't impressed with the SMP speedup, but in light of the huge NPS loss, it seems pretty good. Why the NPS is way down is unknown as I rarely run with SMT off for obvious reasons. And with SMT on, the NPS for 4 threads is right where it ought to be. Something seems fishy... > >so hardware speedup like 1.59 and SMP speedup like 1.54 that very interesting. >It would mean if your hardware speedup goes up, SMP speedup goes up a lot too. > probably just means one of the positions is "good" for SMP and produces a speedup > 2.0. I'll try this on my quads to see how the SMP stuff looks, and I'm going to do some testing to see what's up with the raw NPS with two cpus. >Best regards, >Vincent > >> >>log.003: time=53.90 cpu=397% mat=0 n=107889313 fh=91% >>nps=2001k >>log.003: time=25.70 cpu=395% mat=0 n=57514689 fh=95% nps=2237k >>log.003: time=1:03 cpu=396% mat=-1 n=113942925 fh=93% >>nps=1800k >>log.003: time=41.66 cpu=396% mat=0 n=86861862 fh=92% nps=2085k >>total time used = 184.26 1.73X SMP speedup (four threads, SMP on, two >>processors) >> >>I'm sure that if I run them multiple times, there will be some variability of >>sorts. However, >>none seem to be "badly behaved" positions so I don't think the variability will >>be wild. >> >>any more tests I should run??? Not that four positions is particularly >>interesting in >>terms of a test of anything. Or, as I expect, I'm sure something is wrong with >>these >>results too... >> >>Not particularly bad or good performance in my opinion, but that is just one >>small test >>set that is as meaningless as any other small test set...
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.