[Ecls-list] Release candidate?

[Matthew Mondor]

On Thu, 31 May 2012 09:57:38 +0200
Juan Jose Garcia-Ripoll <juanjose.garciaripoll at googlemail.com> wrote:

> You do not, by any chance, have the commit ID which last proved stable for
> you. It would help me list the changes that might have had an impact.

I checked my logs and I believe that it might have been
fbd3587b1af74a2c5efcdb70671ec9314802fd2e that I merged with then.
That's slightly far back, but it's plausible, as I then ran the HTTPd
for 30 or so days, after which I reported here the uptime on May 16.

So I did another short test run using -c16 -n1000 and even with 1000,
if I do several runs one after the other rather quickly, some runs take
a lot longer to finish than others.  Example:

Requests per second:    2816.90 [#/sec] (mean)
Requests per second:    1046.03 [#/sec] (mean)
Requests per second:    76.03 [#/sec] (mean)
Requests per second:    252.84 [#/sec] (mean)
Requests per second:    2564.10 [#/sec] (mean)

Yet I remember using successfully up to -n10000 on that instance that I
let run 30 days, with good constant performance.  Also, in the slow
cases such as the 76.03 above, the threads seem to be in CPU-busy
loops.  However, if I check the WCHAN of the various threads in top(1),
I see those threads alternate between CPU/RUN/pause.  If I use
ktruss(1) on the process, I see a lot of sched_yield() calls.

I should probably try to revert to that older merge and try again, to
really confirm that I don't experience that issue with it.  I'll try to
do this later on today.

Thanks,
-- 
Matt