Re: Elisp native profiler

[Eli Zaretskii]

> From: Stefan Monnier <address@hidden>
> Date: Mon, 01 Oct 2012 10:25:28 -0400
> Cc: Tomohiro Matsuyama <address@hidden>, address@hidden
> 
> >> I generally consider that if it's fast enough for me, it'll be fast
> >> enough for the large majority of users.
> > OK, I changed the default back to 1 ms.
> 
> I'd much rather that people actually answer my questions about *why*
> they want to change the default, rather than have them change it
> back&forth as the wind blows.
> 
> > How about if we also change the sampling interval units
> > from millseconds to nanoseconds?  It's odd to have a knob
> > where the default is turned all the way to the left.
> 
> Whether 1ms is a good default (and leads to good/valid profiles) is
> still an open question.

Here's one data point: on a Windows XP system based on Core i7 CPU at
3.4GHz, running the profiler in an optimized build makes one of the 8
cores fully (90%) busy.  Emacs is fully usable during profiling, and I
see in the CPU utilization graph that another core does whatever Emacs
needs in the foreground thread.  This is with the default profiling
interval set by profiler.el, which causes the profiling timer be run
at 15 msec intervals, the minimal one supported by the emulated
itimer.

On fencepost.gnu.org, which appears to be a 7-node machine with 8
cores on each node, running the CPU profiler with the default sampling
interval barely takes 1% of the system CPU resources, i.e. again only
one core (at most) is busy.