From: Richard Maine Newsgroups: comp.lang.fortran,comp.parallel.pvm Subject: Re: superlinear speedup, (was: fortran77 programming) Date: 14 Dec 1998 10:48:25 -0800 Organization: NASA Dryden Message-Id: References: <3660D892.E1EDBF12@est.it> <3665044D.A0F57E59@kings.uq.edu.au> <3665E491.7993355@flash.net> <366BF6F3.1E8DC934@cic-mail.lanl.gov> Xref: ukc comp.lang.fortran:62162 comp.parallel.pvm:7870 rsilva@ime.usp.br (Paulo J. da Silva e Silva) writes: > Please, correct me if I am wrong. But I keep on thinking that SUPERLINEAR > speedup is not possible. As has been mentioned elsewhere in this thread, there seems to be a disconnect between those discussing abstract theory and those discussing practice. I've seen superliner speedup in real life. Not a lot, but enough to be measurable. Ran a set of jobs with a single-cpu system (or a multi-cpu system with all but 1 cpu disabled - twas long enough ago that I forget that detail). Then rerun it after adding (or enabling) the multiple cpus. Yes, 'tis true that when the extra cpus were added, they brought with them extra cache, registers, etc.. And that might have more to do with the speedup than with the cpu per se. But then, in practice I don't know how to separate these effects. I've never had handy a theoretical machine that had multiple cpus without also having changes like more registers. The explanations I heard were some of the same ones mentioned in this thread (fewer process context switches because the multiple cpus allow more process contexts), but I was more interested in the measurable result than in the explanation. Turned out the speedup wasn't enough to get very excited about (a few percent), in fact it was not a whole lot above the threshhold of repeatability, but it was by a little. (Not that it much matters, but the machine I recall measuring superlinear speedup on was an Elxsi). -- Richard Maine maine@altair.dfrc.nasa.gov