On Tue, 04 Sep 2012 10:51:06 +0300, Avi Kivity <avi@xxxxxxxxxx> wrote: > On 09/04/2012 04:30 AM, Nikunj A Dadhania wrote: > > On Mon, 03 Sep 2012 17:33:46 +0300, Avi Kivity <avi@xxxxxxxxxx> wrote: > >> On 08/21/2012 02:25 PM, Nikunj A. Dadhania wrote: > >> > > >> > kernbench(lower is better) > >> > ========================== > >> > base pvflushv4 %improvement > >> > 1VM 48.5800 46.8513 3.55846 > >> > 2VM 108.1823 104.6410 3.27346 > >> > 3VM 183.2733 163.3547 10.86825 > >> > > >> > ebizzy(higher is better) > >> > ======================== > >> > base pvflushv4 %improvement > >> > 1VM 2414.5000 2089.8750 -13.44481 > >> > 2VM 2167.6250 2371.7500 9.41699 > >> > 3VM 1600.1111 2102.5556 31.40060 > >> > > >> > >> The regression is worrying. We're improving the contended case at the > >> cost of the non-contended case, this is usually the wrong thing to do. > >> Do we have any clear idea of the cause of the regression? > >> > > Previous perf numbers suggest that in 1VM scenario flush_tlb_others_ipi > > is around 2%, while for contented case its around 10%. That is what is > > helping contended case. > > But what is causing the regression for the uncontended case? > Haven't been able to nail that, any clue on how to profile would help. Regards Nikunj -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
