On Fri, Aug 31, 2012 at 04:27:05PM +0200, Marc Espie wrote: > On Fri, Aug 31, 2012 at 02:08:39PM +0000, Christian Weisgerber wrote: > > A couple of weeks ago, I ran a bunch of make builds (ncpu=4 amd64) > > with different malloc() options enabled. I don't want to spawn a > > discussion, but for anybody who's curious, and for the archives, > > here are the results: > > > > 36m17.68s real 47m33.50s user 26m49.97s system > > F 39m34.31s real 48m22.50s user 35m57.48s system > > G 38m20.87s real 48m40.61s user 31m36.31s system > > J 42m56.74s real 64m19.61s user 25m51.23s system > > FG 40m59.86s real 49m11.18s user 39m12.52s system > > FGJ 49m0.12s real 68m17.34s user 39m34.63s system > > S 55m14.16s real 61m29.02s user 71m24.59s system > > I kind of wonder about similar data for full bulk builds. > > Of course, I know a part of the answer: bulk with S will be shorter > than expected thanks to the fallout in large shit such as mono...
See? S saves time! BTW, if you are wondering why S is slower than FGJ: S is FGJ plus a page size cache of zero. That means than any unused page is unmapped immediately. In normal operation, malloc maintains a set of pages for later re-use. -Otto