Re: profiling aufs

From: Robert Collins <robertc@dont-contact.us>
Date: 07 Oct 2002 20:59:16 +1000

On Mon, 2002-10-07 at 20:34, Henrik Nordstrom wrote:
> On Monday 07 October 2002 11.00, Robert Collins wrote:
>
> > This isn't true. IFF the worker threads do not use -any- common
> > profiled calls, then it works, otherwise we may get xmalloc timings
> > trashed (for example).
>
> xmalloc isn't exacly threads safe today if any of the
> trace/debug/leakcheck xmalloc options is enabled..

Yeah, and this is a minor problem.
 
> And I agree that having "threads safe" profiling is a good idea and
> needs to be investigated. I think a good start would be to allow for
> separate profiling timers. Having a full profiling block per therad
> is most likely overkill.

Probably, but the overhead is relatively small, and it seemed simpler to
simply use a consistent struct across the board.

> A interesting question is how to perform any
> meaningful statisticts on per-thread timers.. the actual timer itself
> is only a small part, you also need a readout of the timer.

I believe I already addressed this. Unless we are interested in the
performance within each thread (I don't think we are), we can simply
collate the statistics each profile event.

Rob

Received on Mon Oct 07 2002 - 04:59:20 MDT

This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:16:53 MST