Re: [patch] voluntary-preempt-2.6.8.1-P0

[Lee Revell <rlrevell@joe-job.com>]

On Sun, 2004-08-15 at 20:25, Florian Schmidt wrote:
> On Sun, 15 Aug 2004 13:56:49 +0200
> Ingo Molnar <mingo@elte.hu> wrote:
> 
> > 
> > i've uploaded the -P0 patch:
> > 
> >  http://redhat.com/~mingo/voluntary-preempt/voluntary-preempt-2.6.8.1-P0
> 
> I haven't tried this patch yet, but i have a question regarding the
> mlockall issue:
> 
> Jackd also uses IPC mechnisms for remote procedure calls [i think,
> please correct me] and makes heavy use of shared memory. Might
> mlock(all) have influence of this? is jackd maybe producing xruns
> because some IPC stuff blocks when mlockall is used?

I reworked the jackd latency histogram to report the time elapsed
between entering and exiting the poll() of PCM fd's in alsa-driver.c. 
This is a different metric than the previous max_usecs histogram, which
I believe indirectly measured latency, this one measures it directly.

http://krustophenia.net/testresults.php?dataset=2.6.8.1-P0

The peaks on this graph should correspond directly to the length of the
non-preemptible critical section reported by Ingo's latency tracer.  I
think the large peak around 580-600usecs is caused by the
extract_entropy issue (which can be hit by regular processes and
ksoftirqd), and the large peak around 80-100 by the XFree86 unmap_vmas
issue, as the times match and these are by far the most common reported
in latency_trace.

There are a number of samples above 700us.  I am working with a period
time of 666 usecs, and since there are 2 periods per buffer, we would
have to hit two > 666 usec latencies in a row for an xrun - it appears
that there are many individual latencies above 666, certainly more than
there are xruns.  So, maybe the mlockall issue is not a result of
triggering a single large latency, but of increasing the frequency of
these higher latencies so that we are more likely to hit 2 in a row.

IIRC ksoftirqd will defer more work under load, and ksoftirqd is one of
the more common offenders to hit the extract_entropy latency.  Maybe
mlockall causes more softirqs to be deferred, thus increaing the change
that we will have to do more than 666 usecs worth of work on 2
successive wakeups.

Lee