From: Jeremie Courreges-Anglas Subject: Re: improve spinning in mtx_enter To: Mark Kettenis Cc: mpi@openbsd.org, alexander.bluhm@gmx.net, mjguzik@gmail.com, tech@openbsd.org Date: Sun, 24 Mar 2024 13:23:39 +0100 Le Sun, Mar 24, 2024 at 12:39:28PM +0100, Mark Kettenis a écrit : > > Date: Sun, 24 Mar 2024 11:53:14 +0100 > > From: Mark Kettenis > > > > > Date: Sun, 24 Mar 2024 11:09:12 +0100 > > > From: Martin Pieuchot > > > > Alexander, your mail still (again?) falls into the spam trap. So I > > didn't see your mail. > > > > > On 21/03/24(Thu) 13:22, Alexander Bluhm wrote: > > > > On Wed, Mar 20, 2024 at 01:43:18PM +0100, Mateusz Guzik wrote: > > > > > few years back I noted that migration from MD mutex code to MI > > > > > implementation happened to regress performance at least on amd64. > > > > > > > > > > The MI implementation fails to check if the lock is free before issuing > > > > > CAS, and only spins once before attempts. This avoidably reduces > > > > > performance. > > > > > > > > > > While more can be done here, bare minimun the code needs to NULL check, > > > > > which I implemented below. > > > > > > > > > > Results on 7.5 (taken from snapshots) timing make -ss -j 16 in the > > > > > kernel dir: > > > > > > > > > > before: 521.37s user 524.69s system 1080% cpu 1:36.79 total > > > > > after: 522.76s user 486.87s system 1088% cpu 1:32.79 total > > > > > > > > > > That is about 4% reduction in total real time for a rather trivial > > > > > change. > > > > > > > > I refactored the diff a bit to keep the while loop. > > > > > > > > With that I get 5% reduction in sys time during kernel build on a > > > > 8 CPU machine. It has two sockets with 4 cores each. > > > > > > > > Flame graphs of kernel build are here: > > > > http://bluhm.genua.de/perform/results/2024-03-19T23:41:35Z/2024-03-19T00%3A00%3A00Z/btrace/time_-lp_make_-CGENERIC.MP_-j8_-s-btrace-kstack.0.svg > > > > http://bluhm.genua.de/perform/results/2024-03-19T23:41:35Z/patch-sys-mutex-spin.1/btrace/time_-lp_make_-CGENERIC.MP_-j8_-s-btrace-kstack.0.svg > > > > > > > > If you zoom into kernel and search for mtx_enter, you see that > > > > spinning time goes down from 13.1% to 10.7% > > > > I assume that's an amd64 machine? Can we test this on some other > > architectures too before we commit it? I don't expect a regression, > > but it would be interesting to see what the effect is on an > > architectures that has better atomic operations than amd64. > > FWIW, I see no measureable difference on a kernel build on my 10-core > M2 Pro Mac mini with this diff. (And that is fine). Same on my single-socket 8 cores amd64 builder. -- jca