From: Martin Pieuchot Subject: Re: UVM performance improvements To: tech@openbsd.org, robert@openbsd.org Date: Sat, 8 Mar 2025 11:30:18 +0100 On 07/03/25(Fri) 17:30, Kirill A. Korinsky wrote: > On Tue, 04 Mar 2025 14:25:44 +0100, > Martin Pieuchot wrote: > > > > Diff below includes all the changes I made during the last months to get > > rid of lock contention and enable parallel fault handling. > > > > With it I measured a 5% performance improvement and a reduction of %sys > > time of 25% on 24CPU amd64. On my 80CPU arm64 performances also improve > > by 5% and %sys time increase by ~5%. > > > > I'd be happy to hear about more tests, especially on bulk builds and on > > other architectures. > > > > I need to rebuild chromium, so I don't mind to do it twice to test the diff > on arm64 as well. > > Without your diff: > > 444m02.18s real 3014m28.87s user 1859m28.53s system > > With your diff: > > 427m18.89s real 3042m22.61s user 1687m28.27s system > > Differences: > - real time is decreased for about 3% > - user time is decreased for about 1% > - system time is decreased for about 10% > > I had built it as: make clean; make patch; time make MAKE_JOBS=1 > > No crash, deadlock or any issue during that test. Thanks for testing. What kind of machine are you running? Do you mind sharing a dmesg? If you set pfault to 1 (in uvm_fault.c) do you see improved performances?