Indeed, there's a huge difference regarding minimum and average latencies now that FCSE is enabled and HRT and CONFIG_PREEMPT are disabled. I get around 13.5 microseconds average latencies.
Thanks a lot. On Sun, Nov 7, 2010 at 11:15 AM, Gilles Chanteperdrix < [email protected]> wrote: > at91_enthus wrote: > > Yes, it solved the issue. > > > > There is one thing that bothers me, though. When I run "xeno-test", the > > average and minimal latencies are constant all the time. > > > > The unpatched version of the xeno kernel seems more consistent when it > comes > > to minimum, maximum and average latencies. > > > > I attached a couple of snippets of xeno-test outputs . > > Do not trust xeno-test too much. Are you sure the only difference > between the two runs is the patch, for instance, have you not played > with the CONFIG_ARM_FCSE option? Also, good benchmarks are done with > running a separate load, and not with xeno-test, in fact. > > Also, using the high-res timers configuration has a high overhead on > such low-end configurations as AT91s, so has CONFIG_PREEMPT. So, if you > are looking for a configuration with the lowest overhead possible, I > would recommend disabling CONFIG_HIGH_RES_TIMERS and using > CONFIG_PREEMPT_NONE or CONFIG_PREEMPT_VOLUNTARY, and relying on Xenomai > for the low latency and high resolution timer features. (You can measure > the overhead of kernel options with the hackbench test for instance). > > On the AT91 I have (AT91SAM9263), the minimum and average latencies > oscillate much less than for your unpatched kernel, either with or > without the patch so I wonder if you would not have another issue. Do > you have FCSE enabled? > > To give you an idea of the figures here, with FCSE enabled, the minimal > latency is around 15us and the average latency around 50us. And they > vary in a 5us range, they do not make big jumps. > > > -- > Gilles. >
_______________________________________________ Xenomai-help mailing list [email protected] https://mail.gna.org/listinfo/xenomai-help
