CAMUS Benoit wrote:
> Hi,
>
> I'm using Xenomai 2.3.1 with a 2.6.19.2 kernel (from ELDK) on a custom
> board based on Xscale PXA processor and with smc91c111 ethernet chip.
> I did modifications on Linux kernel to port it on board and then just
> applied Adeos's patch following Xenomai's proceedings.
> Http and telnet servers are running on and i notice that latencies get
> really worse while i access any of them (see traces below).
>
> I understood that Adeos might avoid that kind of behavior due to its
> architecture. So i infer that the ethernet driver might access the
> hardware by-passing Adeos ?
Despite Adeos architecture, latency with Xenomai is known to depend on
the non real-time workload, because there are some low-level ressources
such as busses and processor cache that the real-time and non real-time
processes need to share.
Nevertheless, the latencies you show us are a bit high, this may have
several reasons:
- maybe you enabled some debugging/statistics options, they have an
impact on performance;
- the latency of your system may not be properly calibrated, try:
echo 25000 > /proc/xenomai/latency
- 10ms is a period a bit high, this makes plenty of time for Linux to
refill the cache, with a lower period you should observe a lower latency.
> In this way and before modificate the driver, i decided to try RT driver
> thanks to Rtnet project and get an oops error at device probing due to
> memory mapping wich i still don't resolve.
> I don't have experience with Xenomai and would apreciate any advice on
> that phenomenum.
Note that using a real-time driver should rather increase the latency
than decrease it, since there will be more unpreemptible critical sections.
> I'm a bit surprised too about gap between best and worst latencies so
> i'm asking question to myself about my porting, does this gap seam
> strange for you too?
--
Gilles Chanteperdrix
_______________________________________________
Xenomai-help mailing list
[email protected]
https://mail.gna.org/listinfo/xenomai-help
