perf sched record with -p option

perf sched record with -p option

[Vuille, Martin (Martin)]

The data files collected by 'perf sched record' are very large and our embedded
system target has limited storage, so I thought of using the '-p PID' option on
'perf sched record' to focus only on a single PID and its threads/children and
thus reduce the amount of data.

However, when I review the results with 'perf sched latency', the results from
data collected with '-p PID' are significantly different (e.g., ten-times longer
latencies) that those from data collected without the '-p PID' option. Workloads
are the same.

Is there some reason why using '-p PID' option on 'perf sched record' wouldn't
make sense?

MV

Re: perf sched record with -p option

[Namhyung Kim]

Hello,

On Sun, Feb 04, 2018 at 05:31:07PM +0000, Vuille, Martin (Martin) wrote:
> The data files collected by 'perf sched record' are very large and our embedded
> system target has limited storage, so I thought of using the '-p PID' option on
> 'perf sched record' to focus only on a single PID and its threads/children and
> thus reduce the amount of data.
> 
> However, when I review the results with 'perf sched latency', the results from
> data collected with '-p PID' are significantly different (e.g., ten-times longer
> latencies) that those from data collected without the '-p PID' option. Workloads
> are the same.
> 
> Is there some reason why using '-p PID' option on 'perf sched record' wouldn't
> make sense?

I think it's because perf missed sched-in and wakeup events for the PID
since they are from different tasks.

Thanks,
Namhyung

RE: perf sched record with -p option

[Vuille, Martin (Martin)]

> On Sun, Feb 04, 2018 at 05:31:07PM +0000, Vuille, Martin (Martin) wrote:
> > The data files collected by 'perf sched record' are very large and our
> > embedded system target has limited storage, so I thought of using the
> > '-p PID' option on 'perf sched record' to focus only on a single PID
> > and its threads/children and thus reduce the amount of data.
> >
> > However, when I review the results with 'perf sched latency', the
> > results from data collected with '-p PID' are significantly different
> > (e.g., ten-times longer
> > latencies) that those from data collected without the '-p PID' option.
> > Workloads are the same.
> >
> > Is there some reason why using '-p PID' option on 'perf sched record'
> > wouldn't make sense?
> 
> I think it's because perf missed sched-in and wakeup events for the PID since
> they are from different tasks.

Thanks.

I am checking the collected events and comparing with and without -p option.

I am collecting data for 60 seconds (passing command 'sleep 60' to perf.)

I have found something very, very odd. When I include the -p option, for the
first 50 or so seconds of the capture, there are only sched_wakeup events
recorded. After that, I start seeing all the events, sched_wakeup, sched_switch,
etc.

Any thoughts?

Regards,
MV