On Mon, Nov 29, 2021 at 10:50:32PM +0100, Martin Pieuchot wrote: > On 24/11/21(Wed) 11:16, Martin Pieuchot wrote: > > Diff below unlock the bottom part of the UVM fault handler. I'm > > interested in squashing the remaining bugs. Please test with your usual > > setup & report back. > > Thanks to all the testers, here's a new version that includes a bug fix.
I removed the witness part and ran it through my perfomamnce tests. 4 core machine http://bluhm.genua.de/perform/results/2021-12-01T23:10:12Z/perform.html 8 core, 2 socket machine http://bluhm.genua.de/perform/results/2021-12-01T09:53:27Z/perform.html The best improvement is make -j8 bsd on the 8 core machine, 22% system time, 5% real time decrease. I see slowdown in network performance, but that may be within the variation I always see when kernel layout changes. Here you can see best how kernel lock goes away in usertrap. http://bluhm.genua.de/perform/results/2021-12-01T09:53:27Z/2021-12-01T00%3A00%3A00Z/btrace/time_-lp_make_-CGENERIC.MP_-j8_-s-btrace-kstack.0.svg http://bluhm.genua.de/perform/results/2021-12-01T09:53:27Z/patch-mpi-uvm-unlock-nowitness.0/btrace/time_-lp_make_-CGENERIC.MP_-j8_-s-btrace-kstack.0.svg On 4 core machine system time during kernel build is reduced by 6%. bluhm
