On Wed, 16 Jun 2004, Robert Watson wrote: > > On Wed, 16 Jun 2004, Robert Watson wrote: > > > On Wed, 16 Jun 2004, Robert Watson wrote: > > > > > Netperf-UP-giant 4902.41 14.3 > > > Netperf-SMP-giant 2566.18 16.83 > > > Netperf-UP-mpsafe 4799.35 22.04 > > > Netperf-SMP-mpsafe 3022.51 18.06 > > > > FYI, when I add ADAPTIVE_MUTEXES on this box, the mean of 3000 q/s goes > > to 4000 q/s on the netperf+smp+mpsafenet number. > > > > I'm off to bed now, but it seems like (at least in the HTT > > configuration), it makes a big difference. I'll run the remainder of > > that set with ADAPTIVE_MUTEXES tomorrow as well. > > And switching to 4BSD from ULE takes the mean to 6426 q/s when combined > with netperf, smp, debug.mpsafenet=1, and adaptive mutexes. > > (Keeping in mind that this mysql benchmark is basically an > inter-thread/process kernel IPC benchmark using UNIX domain sockets, since > the workload is fairly minimal in userspace, this makes reasonable sense). > > Now I'm really going to bed. One thing that was happenning at one time, was that when a process is made runnable on one processor, If another processor is idle, it remains idle until the next scheduling tick. I proposed patches about a year ago that fixed this but "as far as I know" it is still true. The solution is to generate an IPI when you make a thread runnable and you see an idle processor.. To see if this is the problem a quick test would be to move HZ to 5000 or something and see what happens. > > Robert N M Watson FreeBSD Core Team, TrustedBSD Projects > robert_at_fledge.watson.org Senior Research Scientist, McAfee Research > > >Received on Wed Jun 16 2004 - 05:49:07 UTC
This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:37:57 UTC