Re: [RFC PATCH v4 2/6] perf stat: Fork and launch perf record when perf stat needs to get retire latency value for a metric.

[Andi Kleen]

On Wed, Mar 13, 2024 at 03:31:14PM +0000, Wang, Weilin wrote:
> 
> 
> > -----Original Message-----
> > From: Andi Kleen <ak@xxxxxxxxxxxxxxx>
> > Sent: Tuesday, March 12, 2024 5:56 PM
> > To: Wang, Weilin <weilin.wang@xxxxxxxxx>
> > Cc: Namhyung Kim <namhyung@xxxxxxxxxx>; Ian Rogers
> > <irogers@xxxxxxxxxx>; Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>; Peter
> > Zijlstra <peterz@xxxxxxxxxxxxx>; Ingo Molnar <mingo@xxxxxxxxxx>;
> > Alexander Shishkin <alexander.shishkin@xxxxxxxxxxxxxxx>; Jiri Olsa
> > <jolsa@xxxxxxxxxx>; Hunter, Adrian <adrian.hunter@xxxxxxxxx>; Kan Liang
> > <kan.liang@xxxxxxxxxxxxxxx>; linux-perf-users@xxxxxxxxxxxxxxx; linux-
> > kernel@xxxxxxxxxxxxxxx; Taylor, Perry <perry.taylor@xxxxxxxxx>; Alt, Samantha
> > <samantha.alt@xxxxxxxxx>; Biggers, Caleb <caleb.biggers@xxxxxxxxx>
> > Subject: Re: [RFC PATCH v4 2/6] perf stat: Fork and launch perf record when
> > perf stat needs to get retire latency value for a metric.
> > 
> > "Wang, Weilin" <weilin.wang@xxxxxxxxx> writes:
> > 
> > >> -----Original Message-----
> > >> From: Andi Kleen <ak@xxxxxxxxxxxxxxx>
> > >> Sent: Tuesday, March 12, 2024 5:03 PM
> > >> To: Wang, Weilin <weilin.wang@xxxxxxxxx>
> > >> Cc: Namhyung Kim <namhyung@xxxxxxxxxx>; Ian Rogers
> > >> <irogers@xxxxxxxxxx>; Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>;
> > Peter
> > >> Zijlstra <peterz@xxxxxxxxxxxxx>; Ingo Molnar <mingo@xxxxxxxxxx>;
> > >> Alexander Shishkin <alexander.shishkin@xxxxxxxxxxxxxxx>; Jiri Olsa
> > >> <jolsa@xxxxxxxxxx>; Hunter, Adrian <adrian.hunter@xxxxxxxxx>; Kan Liang
> > >> <kan.liang@xxxxxxxxxxxxxxx>; linux-perf-users@xxxxxxxxxxxxxxx; linux-
> > >> kernel@xxxxxxxxxxxxxxx; Taylor, Perry <perry.taylor@xxxxxxxxx>; Alt,
> > Samantha
> > >> <samantha.alt@xxxxxxxxx>; Biggers, Caleb <caleb.biggers@xxxxxxxxx>
> > >> Subject: Re: [RFC PATCH v4 2/6] perf stat: Fork and launch perf record
> > when
> > >> perf stat needs to get retire latency value for a metric.
> > >>
> > >> weilin.wang@xxxxxxxxx writes:
> > >>
> > >> > From: Weilin Wang <weilin.wang@xxxxxxxxx>
> > >> >
> > >> > When retire_latency value is used in a metric formula, perf stat would fork
> > a
> > >> > perf record process with "-e" and "-W" options. Perf record will collect
> > >> > required retire_latency values in parallel while perf stat is collecting
> > >> > counting values.
> > >>
> > >> How does that work when the workload is specified on the command line?
> > >> The workload would run twice? That is very inefficient and may not
> > >> work if it's a large workload.
> > >>
> > >> The perf tool infrastructure is imho not up to the task of such
> > >> parallel collection.
> > >>
> > >> Also it won't work for very long collections because you will get a
> > >> very large perf.data. Better to use a pipeline.
> > >>
> > >> I think it would be better if you made it a separate operation that can
> > >> generate a file that is then consumed by perf stat. This is also more efficient
> > >> because often the calibration is only needed once. And it's all under
> > >> user control so no nasty surprises.
> > >>
> > >
> > > Workload runs only once with perf stat. Perf record is forked by perf stat and
> > run
> > > in parallel with perf stat. Perf stat will send perf record a signal to terminate
> > after
> > > perf stat stops collecting count value.
> > 
> > I don't understand how the perf record filters on the workload created by
> > the perf stat. At a minimum you would need -p to connect to the pid
> > of the parent, but IIRC -p doesnt follow children, so if it forked
> > it wouldn't work.
> > 
> > I think your approach may only work with -a, but perhaps I'm missing
> > something (-a is often not usable due to restrictions)
> > 
> > Also if perf stat runs in interval mode and you only get the data
> > at the end how would that work?
> > 
> > iirc i wrestled with all these questions for toplev (which has a
> > similar feature) and in the end i concluded doing it automatically
> > has far too many problems.
> > 
> 
> Yes, you are completely right that there are limitation that we can only support -a, -C 
> and not support on -I now. I'm wondering if we could support "-I" in next step by 
> processing sampled data on the go.

-I is very tricky in a separate process. How do you align the two
intervals on a long runs without drift. I don't know of a reliable
way to do it in the general case only using time.

Also just the non support for forking workloads without -a is fatal imho. That's 
likely one of the most common cases.

Separate is a far better model imho:

- It is under full user control and no surprises
- No uncontrolled multiplexing
- Often it is fine to measure once and cache the data

It cannot deal with -I properly either (short of some form of
phase detection), but at least it doesn't give false promises
to that effect.

The way to do it is to have defaults in a json file
and the user can override them with a calibration step.
There is a JSON format that is used by some other tools.

This is my implementation:
https://github.com/andikleen/pmu-tools/blob/master/genretlat.py
https://github.com/andikleen/pmu-tools/blob/89861055b53e57ba0b7c6348745b2fbe6615c068/toplev.py#L1031


-Andi