Re: [Xenomai-help] AT91SAM9260 latency

[Juan Antonio Garcia Redondo <juan-antonio.garcia@domain.hid>]

On 24/01/08 11:02, Gilles Chanteperdrix wrote:
> On Jan 24, 2008 10:41 AM, Juan Antonio Garcia Redondo
> <juan-antonio.garcia@domain.hid> wrote:
> >
> > On 23/01/08 14:15, Gilles Chanteperdrix wrote:
> > > On Jan 23, 2008 11:04 AM, Gilles Chanteperdrix
> > > <gilles.chanteperdrix@xenomai.org> wrote:
> > > > On Jan 23, 2008 7:52 AM, Juan Antonio Garcia Redondo
> > > >
> > > > <juan-antonio.garcia@domain.hid> wrote:
> > > > > I see everything OK except for the first samples of cyclictests. Any comments ?
> > > >
> > > > The load you apply does not load the cache, which is a source of
> > > > jitter. You should run the cache calibrator (I do not find the cache
> > > > calibrator URL, but it is somewhere in Xenomai distribution or wiki).
> > >
> > > It is in the TROUBLESHOOTING guide, question "How do I adequately stress test".
> > >
> > > --
> > >                                                Gilles Chanteperdrix
> >
> > Thanks Gilles, I've done more tests using the cache calibrator from
> > http://www.cwi.nl/~manegold/Calibrator. The latency numbers are very
> > similar althought I've found an strange behaviour related to telnet
> > sessions.
> 
> Are you kidding ? In the first results you posted, the latency was
> around 80us whereas now, you get a latency around 130us. And from I
> read, you did not run the tests for long period. If you want reliable
> results, you should let the test run, under load, for hours.

Well, after several tests, (one of them 4 hours long), I can't see
latencies above 100 us. Anyway I'll do more tests this weekend. The
latency around 130us occurs with telnet activity.
> 
> >
> > Environment:
> >         o Tests running from console over atmel serial port.
> >         o A telnet session over on-chip ethernet.
> >         o System without load.
> >
> > ./latency -p 500 -t0
> > == All results in microseconds
> > warming up...
> > RTT|  00:00:01  (periodic user-mode task, 500 us period, priority 99)
> > RTH|-RTH----lat min|-----lat avg|-----lat max|-overrun|----lat best|---lat
> > worst
> > RTD|      49.613|      52.190|      62.822|       0|      49.613| 62.822
> > RTD|      42.203|      52.512|      66.365|       0|      42.203| 66.365
> >
> >
> > Now If hit a key on the telnet session :
> >
> > RTD|      36.726|      57.989|     109.536|       0|      31.572| 109.536  <-------- Here I've hit the key.
> > RTD|      36.404|      51.868|      69.587|       0|      31.572| 109.536
> > RTD|      35.760|      51.868|      73.775|       0|      31.572| 109.536
> >
> > Now, I launch an script which executes four instances of cache
> > calibrator.
> >
> > RTD|      45.103|      57.667|      75.708|       0|      32.538| 122.422
> > RTD|      45.425|      57.023|      76.030|       0|      32.538| 122.422
> > RTD|      46.069|      57.023|      75.708|       0|      32.538| 122.422
> >
> > Now, I can hit a key on the telnet session without effects over latency
> > numbers:
> >
> > RTD|      44.136|      57.989|      75.386|       0|      27.384| 128.221
> > RTD|      46.713|      57.345|      76.353|       0|      27.384| 128.221
> > RTD|      44.780|      57.345|      76.675|       0|      27.384| 128.221
> > RTD|      43.492|      56.701|      76.997|       0|      27.384| 128.221
> >
> > Now I stop the calibrator process and launch 'ping -f -s2048 192.168.2.82' from an external
> > machine.
> >
> > RTD|      40.270|      68.621|      90.850|       0|      27.384| 128.221
> > RTD|      36.082|      68.621|      88.273|       0|      27.384| 128.221
> > RTD|      40.592|      67.976|      91.494|       0|      27.384| 128.221
> > RTD|      41.237|      68.298|      89.239|       0|      27.384| 128.221
> >
> >
> > Now If hit a key on the telnet session :
> >
> > RTD|      42.203|      67.976|      88.273|       0|      27.384| 128.221
> > RTD|      32.216|      93.427|     128.543|       0|      27.384| 128.543 <---------- Here I've hit the key.
> > RTD|      42.203|      68.298|      87.628|       0|      27.384| 128.543
> >
> > And again the calibrator execution results on eliminate the strange
> > behaviour whith the telnet session.
> >
> > Any clues ?
> 
> No mystery: hitting a key on a telnet session causes an interrupt
> masking section of 110us, you see it as the maximum if you never
> observed longer masking sections, but it is not the maximum if you
> observed longer masking sections.

OK, but why the masking section on linux side affects to xenomai side ?
Another thing I don't understand is why when the system has load (above
I'm talking about calibrator but the same occurs with dd if=/dev/zero
of=/dev/null), the effect seems to dissapear.  

> 
> >
> > BTW, if finally the bad numbers on ARM are user-context switches related,
> > are you considering the ipipe upgrading to 2.6.23 ?
> 
> No comment. I have already answered this question.

Sorry, I missed the last entries on "High latencies on ARM" from
xenomai-core list.

Regards,
	Juan Antonio