On Thu, Apr 25, 2013 at 11:28 AM, Andre Oppermann <andre_at_freebsd.org> wrote: > > > Again one has to be really careful drawing any firm conclusions from this > as it was measured on a Pentium4 and UP kernel (GENERIC would add WITNESS > and INVARIANT overhead as well). > > The Pentium4 is about the worst micro-architecture when it comes to locks > and easily regresses. At the same time modern Intel Core i[3-7] and AMD64 > may actually improve with these changes. Unless more recent micro-archs > have been shown to exhibit the same regression we can't claim this change > was bad (other than for Pentium4). OK, here are the results of the same bench on another server (HP ProLiant DL320 G5): - Dual Core: Intel Xeon CPU 3050 2.13GHz (2133.45-MHz K8-class CPU) - NIC changed to dual 82571EB Graph: http://gugus69.free.fr/freebsd/benchs/current2/current-pps.png gnuplot data: http://gugus69.free.fr/freebsd/benchs/current2/plot/ ministat data: http://gugus69.free.fr/freebsd/benchs/current2/ministat/ raw data: http://gugus69.free.fr/freebsd/benchs/current2/raw/ Notice the Glebius' explanation regarding a unique one-flow test and the new pf-smp behavior: http://lists.freebsd.org/pipermail/freebsd-net/2013-April/035417.html Regards, OlivierReceived on Sun Apr 28 2013 - 13:57:01 UTC
This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:40:37 UTC
