lkml.org 
[lkml]   [2010]   [Mar]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Subject[patch 0/12] sched: fastpath cycle recovery
From
Date
Hi Peter,

The following patchlets take a pinned pipe-test context switch frequency
in tip from 663KHZ to 694KHZ, and an unpinned instance from 450KHz to
540KHz. With these applied to tip.today, I have zero 31-12->today
regressions, and even some modest progressions.

The biggest difference is made by the first patch. We have a problem
with nohz when waking cross-cpu, which given select_idle_sibling(), we
do quite a bit. In testing netperf TCP_RR, hitting nohz code on every
micro-idle was eating ~10% of throughput, making cross-cpu wakeup a
loser. These patchlets combined turned netperf TCP_RR cross-cpu vs
affine from big loser into a winner.

All of these are trivial, mostly axe murder, but cycles add up.

-Mike



\
 
 \ /
  Last update: 2010-03-11 10:51    [W:0.232 / U:0.216 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site