Re: [PATCH 1/2] perf tools: Add reference timestamp to perf header

[Arnaldo Carvalho de Melo <acme@ghostprotocols.net>]

Em Mon, Dec 13, 2010 at 11:01:13AM -0700, David S. Ahern escreveu:
> On 12/13/10 10:57, Peter Zijlstra wrote:
> >> What about creating a PERF_RECORD_TIME and generate an event when the
> >> counter is opened? It contains a PERF_SAMPLE_TIME and say
> >> PERF_SAMPLE_TOD (time-of-day)? We're not sending rockets to saturn; we
> >> just need the timestamps to match other log files.

> > That's similar to the first thing I proposed. The problem is with long
> > record sessions your drift can become quite significant, then when you
> > merge sort your other log events stuff can get out of order. Which can
> > lead to some serious head-scratching..

> Gotcha. Missed that in the flury of emails.

> Arnaldo: Are you ok with this option? This should append mode as well.

What option, this one:

-----------------------------------------------------------------------
Peter Zijlstra wrote:

> In fact, the only sane way to do that is by creating a software counter
> that represents CLOCK_MONOTONIC and sample that say once a minute (or
> more often if you want smaller drift).
-----------------------------------------------------------------------

?

Yes, that looks the best option. I.e. no changes on headers, no new
user fake events, a new software event that may be useful for other
usecases.

We have:

[acme@mica linux]$ perf list | grep -- -clock
  cpu-clock                                  [Software event]
  task-clock                                 [Software event]
[acme@mica linux]$ 

So we would have a new one:

  monotonic-clock

Peter, agreed? I'll try to implement it now, good opportunity to learn a
bit more about soft pmus, I'd have to do that anyway for NIC stats, etc.

- Arnaldo