Author: Robert Hyatt
Date: 18:58:08 09/05/02
Go up one level in this thread
On September 05, 2002 at 21:18:42, Slater Wold wrote: >There has been a fairly large debate lately over 'speedup'. > >The problem: > >1 CPU: 800k NPS & 60 seconds to solve position >2 CPUs: 1600k NPS & 45 seconds to solve position > >What's the "speedup"? > This was another "vincent" debate. Speedup is always 1cputime/ncputime. The "NPS" stuff was something he pulled out of his hat, claiming that most users expect about 2x the NPS when they run a program on a dual, but that Crafty suffered a "huge loss" (his words) and couldn't get near that. I said "can too" and there we went. To make matters worse, Eugene has posted some Intel and AMD numbers and the Intel looked pretty good, but the AMD numbers were bad. And, of course, Vincent runs everything on AMD. So I simply asked a few to run one and two cpu tests and provide the raw NPS number. It had nothing to do with the speedup in the context most of us use here... But clearly if a program can't search 1.5X the nps, then it can't search 1.5x faster either... so it is an interesting number as nps is certainly an upper bound on the expected actual speedup, which will probably be less in reality. > >As far as I can tell, most super-computer manufacturers and parallel design >"specialists" define speedup as "the ratio of the >serial run time of a parallel application for solving a problem on a >single processor, to the time taken by the same parallel application >to solve the same problem on n processors". Correct. That is the definition I use also... > >RSn = T1/Tn > > >Today I ran the WAC suite using Crafty 18.15 (SMP) from Hyatt's FTP site. One >using just 1 CPU, and the other using both. Using these 300 positions, I plan >to determine if their is any link between an increase in NPS and an increase in >solve time. > >I will try a few different compiles of Crafty. The first one yielded a >*horrible* NPS speedup in the first 100 positions (1.23). Probably because they are solved so very quickly... but that is just a guess. > > >[D]1k6/5RP1/1P6/1K6/6r1/8/8/8 w - - >(WAC 41) > >Average NPS using 1 CPU: 1503k NPS >Average NPS using 2 CPUs: 1786k NPS > >Time to 11th ply using 1 CPU: 48.36 seconds >Time to 11th ply using 2 CPUs: 2.99 seconds > >[D]r1b1qrk1/2p2ppp/pb1pnn2/1p2pNB1/3PP3/1BP5/PP2QPPP/RN1R2K1 w - - >(WAC 76) > >Average NPS using 1 CPU: 804k NPS >Average NPS using 2 CPUs: 1306k NPS > >Time to 11th ply using 1 CPU: 19.39 seconds >Time to 11th ply using 2 CPUs: 13.42 seconds > > >These 2 logs are *full* of these kinds of things. > > >Any comments/thoughts/ideas/suggestions welcome.
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.