Re: [Jackit-devel] Re: Statistical methods for latency profiling

[Lee Revell]

On Sun, 2004-08-01 at 07:21, Ingo Molnar wrote:
> On Sun, 1 Aug 2004, Lee Revell wrote:
> 
> > So stressing the filesystem moves the center to the right a bit, from
> > 6-7 to 9-10, and *drastically* lengthens the 'tail'.
> 
> basically each codepath has a typical latency distribution, and when a
> workload uses multiple codepaths then the latencies get intermixed almost
> linearly.
> 

I noticed several distinct spikes with 1M samples, that blend into
smooth Erlang/gamma type distribution at 5M.  I posted some more results
to jackit-dev.  It seems like each of these would represent a common
code path out of a non-preemptible region.  I suspect the spike at 70-80
usecs is a bug in my code, from updating the histogram every 1024
cycles.  I will start posting results on the web soon, it's getting big.

> > These numbers suggest to me that a lot of the latencies from 47 usecs
> > and up are caused by one code path, because they are so uniformly
> > distributed over the upper part of the histogram.  The prime suspect of
> > course being the ide io completions.  I tested this theory by lowering
> > max_sectors_kb from 64 to 32:
> 
> > These numbers all point to the ide sg completion code as the only thing
> > on the system generating latencies over ~42 usecs.
> 
> yep, that's a fair assumption. Once the IO-APIC irq-redirection problems
> are solved i'll try to further thread the IDE completion IRQ to remove
> that ~100 usecs latency.
> 

It would be interesting to identify the code paths corresponding to the
other peaks.  It occurred to me that if you suspect a peak in the
histogram is related to a certain code path, you could stick a udelay in
there, and see if the spike moves up by the same amount.

Lee


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/