11.07 oprofile on panda busted?
f-turgis at ti.com
Tue Aug 30 16:48:29 UTC 2011
I could suggest to run 10 times the test in a row ;-) but oprofile clearly relies on a configurable timer and not the scheduler tick itself so there is no reason for this limitation, as this is not a rewrite.
For example x86 has a CONFIG_OPROFILE_EVENT_MULTIPLEX feature for which developers added /dev/oprofile/time_slice for tuning so no religion here. Probably inherited from the past where platforms were less powerful.
FYI, oprofile "file operations" are set in arch/arm/oprofile/common.c, leveraging ARM Performance Monitoring Unit. But if oprofile.timer is set, oprofile functions are redefined to leverage generic hrtimer.
PMU can lose interrupt when register overflows (ARM errata), current potential work-around is to use several counters for same event so that at least 1 of them catches the info :-( not very reliable => so oprofile.timer = 1 shall be the default.
OMAP Platform Business Unit - OMAP System Engineering - Platform Enablement
Texas Instruments France SA, 821 Avenue Jack Kilby, 06270 Villeneuve Loubet. 036 420 040 R.C.S Antibes. Capital de EUR 753.920
> From: Tom Gall [mailto:tom.gall at linaro.org]
> Sent: Tuesday, August 30, 2011 6:03 PM
> To: Turgis, Frederic
> Cc: Clark, Rob; Linaro Dev
> Subject: Re: 11.07 oprofile on panda busted?
> On Tue, Aug 30, 2011 at 10:53 AM, Turgis, Frederic
> <f-turgis at ti.com> wrote:
> > Hi,
> > As indicated by Siamashka, we are expected to operate at
> scheduler tick rate, i.e. 128Hz
> (drivers/oprofile/timer_int.c). From your test, we have 84
> ticks, i.e. 656ms of test.
> > Is that correct ? I checked that also in our Android kernel
> and it was the same.
> Makes sense, but it brings me back to my earlier lament, that
> a user tunable option to up the rate for the purpose of
> collecting more samples would be quite desirable.
> > The only impacts we have seen in the past are related to
> idle time. For same duration, we had less ticks if more idle
> time. But Android guys, who used it more than us, never
> really complained on this or timer granularity. Probably this
> was more orented to multimedia use cases so 10s of s.
> > Regards
> > Fred
> > Frederic Turgis
> > OMAP Platform Business Unit - OMAP System Engineering - Platform
> > Enablement
> > Texas Instruments France SA, 821 Avenue Jack Kilby, 06270
> > Loubet. 036 420 040 R.C.S Antibes. Capital de EUR 753.920
> "We want great men who, when fortune frowns will not be discouraged."
> - Colonel Henry Knox
> Linaro.org ¦ Open source software for ARM SoCs
> w) tom.gall att linaro.org
> w) tom_gall att vnet.ibm.com
> h) tom_gall att mac.com
More information about the linaro-dev