profiling (was: Re: [vdr] How to speed up vdr start ?)

[Jouni Karvo]

hi,

Marko Mäkelä writes:
 > I've used OProfile on Intel Celeron, Intel Pentium M and AMD Opteron,
 > maybe even the AMD K6-2, on 2.4 and 2.6 series kernels.  Be sure to
 > enable the local APIC, or otherwise the NMI based performance counter
 > interrupts won't be available.  On my vdr box, I have tuned the
 > softdevice plugin based on OProfile measurements.
 > 

I tried to check where all the cycles in the VDR box go:

CPU: P4 / Xeon, speed 2398.74 MHz (estimated)
Counted GLOBAL_POWER_EVENTS events (time during which processor is not
stopped) with a unit mask of 0x01 (mandatory) count 100000
samples  cum. samples  %        cum. %     image name	app name                 symbol name
47908471 47908471      27.5110  27.5110    vmlinux	vmlinux                  get_offset_pmtmr
13718848 61627319       7.8779  35.3889    libc-2.3.2.so libc-2.3.2.so            (no symbols)
13160494 74787813       7.5573  42.9462    xineplug_post_tvtime.so xineplug_post_tvtime.so  DeinterlaceGreedy2Frame_SSE
8346769  83134582       4.7930  47.7392    vmlinux      vmlinux                  sysenter_past_esp
5992342  89126924       3.4410  51.1802    libpthread-0.60.so libpthread-0.60.so       __pthread_mutex_unlock_usercnt
4959342  94086266       2.8479  54.0281    libxine.so.1.12.0 libxine.so.1.12.0        yv12_to_yuy2_mmxext

I wonder why 27.5 % of samples go to get_offset_pmtmr...  It seems to
be found in /usr/src/linux-2.6.9/arch/i386/kernel/timers/timer_pm.c

Would this be normal, or what should I trim?

yours,
		Jouni