On Fri, Mar 25, 2011 at 11:21:27PM +0000, Mark Morgan Lloyd wrote: > Everybody knows that benchmarks are odious, but I've got a handy > "torture test" that I've run on a variety of systems from a 16-way Sun > down to a "Slug". Usefully, this can be used both to get a time to > completion of the standard job and a Watt-minutes figure, both of which > are significant when choosing systems:
Me too. But I use FPC's "make all" usually. The sheevaplug does about 25minutes doing that (compared to sub 2 minutes on modern Core2 systems). It would be interesting to see how the efika fares. (also 25min I'd expect, since it is at 600MHz half the speed, but twice the DMIPS. But I think it will only do so if it really gets specialized code generated. So we'll probably have to compile something with gcc to benchmark too, assuming that gcc can already generated optimized code) I looked up Henry's comments, and I can also only find the "(up to) 2 DMIPS)" remark, but no actual benchmarks. (only subjective "feels faster" remarks) The fun part of using FPC as benchmark is that it is not a synthetic benchmark, and not purely number crunching, and as it doesn't use highly optimized libraries, it gives a good impression of how memory intensive business code will perform. So I actually use it at work to get an impression of new HW. In the past, the linux kernel compile was used as a benchmark figure a lot, and I'm somewhat sad that the benchmarking community dropped that. -- _______________________________________________ Lazarus mailing list Lazarus@lists.lazarus.freepascal.org http://lists.lazarus.freepascal.org/mailman/listinfo/lazarus