Messages in this thread | | | Date | Thu, 28 May 2009 17:04:47 +0200 | From | Martin Schwidefsky <> | Subject | [patch 0/2] NOHZ vs. profile/oprofile |
| |
Greetings, Rob pointed me to a deficiency with the current profile/oprofile code together with NOHZ. For us this problem crept in with the conversion of s390 to generic clock events, git commit 5a62b192 If the system is running with the HZ-tick disabled and the cpu spents time in idle we see skewed percentages e.g. with the oprofile output. On an I/O bound system the number of idle ticks is way to small. The reason is that the generic clock events code reports either zero or one tick to profile/oprofile on wakeup from idle even if the cpu has slept much longer. I've tried to fix that with the two patches in this series and another pure s390 specific fix (profile_tick is called from clock_comparator_work which is nonsense). These three do correct the oprofile output on s390. The best idea I had to get oprofile in good shape again is to let the system tick at the HZ rate while oprofile is working. Better ideas welcome.
-- blue skies, Martin.
"Reality continues to ruin my life." - Calvin.
| |