On Thu, 9 Feb 2017 13:19:09 +0100 Julian Taylor <[email protected]> wrote:
> These numbers are very surprising to me. I do agree with Julian ... Once upon a time, I have been a computational chemist and believed I could get this 2-6x speed up by recompiling the whole scientific stack with the best compiler (intel icc & ifc) on the very specific computer. After spending a week or two of my PhD on that, I compared the performances with the one from the stock debian build and it turned out to be within 1% my costly build. Very disappointing to me. Good job you guys! Comparing figures with and without OpenMP is one thing, with and without an optimization flag is another and mixing the two is a nonsense. Last point, as a computational chemist, we often do convergence of self-consistent systems, which means loops of loops of calculation. By tuning some compiler option (-O3 with icc), the precision of calculation were traded for speed which made the inner loop go faster (x2 sometimes) but the outer loop became completely chaotic and sometimes failed in converging, resulting in much slower executable. I believe debian should simplify the re-build of packages, if one wants or needs. But I don't believe the figures out there. Site specialized in benchmarking also confirm the speed-ups are in the range of % and not times: https://openbenchmarking.org/result/1701168-PTS-GCC7JANU87 Cheers, -- Jérôme Kieffer tel +33 476 882 445
