Re: Elisp native profiler

[Paul Eggert]

On 10/01/2012 01:29 PM, Stefan Monnier wrote:
> If it's common for 1ms to suffer from significant delays,
> then it's a good argument for setting the default higher.

I did a bit more investigation on this, on my host
(x86-64, AMD Phenom II X4 910e, Fedora 17).  I instrumented
trunk bzr 110342 to count total timer overruns, and did a
CPU benchmark that took about 15.6 seconds when not profiling.

With a 10 ms profiling interval, there were zero overruns and the
benchmark took 15.62 seconds, essentially full speed.

With a 1 ms interval, there were 65 overruns and the
benchmark took 51.52 seconds, a worse than 3x slowdown.

With an 0.1 ms interval, there were 466,644 overruns
and the benchmark took 52.12 seconds.  It appears that
the effective interval was closer to 1 ms than to 0.1 ms, as
most of the samples were overrun.

These numbers suggest that we should go with a 10 ms interval,
at least for this kind of host.