> Date: Mon, 26 Apr 2021 12:43:27 +0200 > From: Alexander Bluhm <alexander.bl...@gmx.net> > > On Thu, Apr 22, 2021 at 03:38:53PM +0200, Martin Pieuchot wrote: > > Diff below remove the KERNEL_LOCK()/UNLOCK() dance from uvm_fault() for > > both amd64 and sparc64. That means the kernel lock will only be taken > > for lower faults and some amap/anon code will now run without it. > > > > I'd be interested to have this tested and see how much does that impact > > the build time of packages. > > I did measure kernel build time make -j 4 on a 4 core machine. > System time goes down by 7.7%, real time stays the same.
That basically suggests that contention of the kernel lock isn't the major bottleneck for this workload. But less spinning is still good since it will reduce the power consumption of modern CPUs. > http://bluhm.genua.de/perform/results/2021-04-25T17%3A18%3A32Z/gnuplot/make.html > left column current, righ column with uvm unlock diff > > http://bluhm.genua.de/perform/results/2021-04-25T17%3A18%3A32Z/perform.html > see row time -lp make -CGENERIC.MP -j4 -s > > http://bluhm.genua.de/files/kstack-make.svg > http://bluhm.genua.de/files/kstack-make-uvm.svg > I cannot see much difference here. > > bluhm > >
