On Fri, Aug 21, 2020 at 11:59:58AM -0700, Junio C Hamano wrote: > > After writing a response elsewhere in the thread, it occurred to me that > > a good candidate for explaining this may be that our modern sha1dc > > implementation is way slower than what we were using in 2012 (which > > would have been either block-sha1, or the even-faster openssl > > implementation). And since a good chunk of index-pack's time is going to > > computing sha1 hashes on the resulting objects, that means that since > > 2012, we're spending relatively more time in the hash computation (which > > parallelizes per-object) and less time in the other parts that happen > > under a lock. > > Believable conjecture that is. You could benchmark again with > block-sha1 on today's hardware, but because the performance profile > with sha1dc is what matters in the real world anyway... Yeah, I agree on the "real world" part, but I'm the curious sort, so here are numbers compiled against openssl (which is generally even faster than block-sha1, and would thus emphasize the results of our hypothesis): 5302.3: index-pack 0 threads 108.78(106.39+2.31) 5302.4: index-pack 1 threads 110.65(108.08+2.49) 5302.5: index-pack 2 threads 67.57(110.83+2.75) 5302.6: index-pack 4 threads 48.18(123.82+3.02) 5302.7: index-pack 8 threads 39.07(153.45+4.13) 5302.8: index-pack 16 threads 38.38(265.78+7.71) 5302.9: index-pack default number of threads 54.64(117.35+2.73) So it's actually pretty similar. Things continue getting faster as we go past 3 threads. Though our total improvement is less (29% better with 8 threads compared to 3, versus 42% better when using sha1dc). So I think it's _part_ of the reason, but not all of it. -Peff
