On Mon, Feb 13, 2017 at 10:00:25AM -0800, Bryan O'Sullivan wrote: > On Mon, Feb 13, 2017 at 9:29 AM, Simon Farnsworth <simon...@fb.com> wrote: > > > The perfwrite microbenchmark shifts to: > > Linux: > > ! wall 3.316087 comb 0.900000 user 0.810000 sys 0.090000 (best of 3) > > Mac: > > ! wall 0.939282 comb 0.580000 user 0.470000 sys 0.110000 (best of 8) > > > > If I open-code timings in ui.write instead of using the context manager, I > > see: > > Linux: > > ! wall 3.478070 comb 0.500000 user 0.420000 sys 0.080000 (best of 3) > > Mac: > > ! wall 0.218112 comb 0.220000 user 0.150000 sys 0.070000 (best of 15) > > > > > Thanks for adding this data. (For future reference, it's good to include an > analysis of the difference instead of just the plain numbers, as I have to > flip back and forth between two emails and memorize the numbers to do the > comparison in my head.) > > These numbers suggest to me that open-coding the performance measurement is > the right thing to do for ui.write and ui.flush.
+1 on open-coding for perf reasons, maybe make a comment to that effect so future optimizers don't get overeager. > > Here's why. > > First, let's discount the wall time, as this number is so noisy as to make > no sense. For example, if you compare the uninstrumented code with the > numbers above, you'll see that wall time has *dropped* by 30% or so on OSX, > even though we're doing more work. This is a good object lesson in why > contrib/perf.py is very far from a trustworthy benchmarking tool :-) > > If you look instead at the somewhat less suspect "comb" time (user+system > combined), you'll see that it more than doubles under the context manager, > vs a smallish bump with open coding of the measurements. > > With its fancy templating and colour support, Mercurial is already > surprisingly slow at simply writing to stdout, an activity that is > obviously on the critical path for many uses. I believe it's worth the few > extra lines of code to keep the lowest-level part of this as fast as > possible. > _______________________________________________ > Mercurial-devel mailing list > Mercurial-devel@mercurial-scm.org > https://www.mercurial-scm.org/mailman/listinfo/mercurial-devel _______________________________________________ Mercurial-devel mailing list Mercurial-devel@mercurial-scm.org https://www.mercurial-scm.org/mailman/listinfo/mercurial-devel
