Gilles Chanteperdrix wrote: > Dmitry Adamushko wrote: > > On 20/03/07, Gilles Chanteperdrix <[EMAIL PROTECTED]> wrote: > > > Muruganandam Ganapathy wrote: > > > > The board is based on the Fujitsu SOC which has the ARM926EJ processor > > core. > > > > > > > > This board has SPI, I2C and 10/100 ethernet interfaces and it can > > support > > > > 16/32MB SDRAM > > > > and 4/8MB flash memory. > > > > > > You still do not tell us the name of the board, but it is probably not > > > supported. > > > > > > > > > > > > > > > In addition, I would like to know the interrupt reponse mesaured with > > > > Xenomai in ARM9 > > > > processor based platforms, if any. > > > > The interrupt response expected is around 40 -50 microseconds in our > > case. > > > > The interrupt response I mean, it is the time between the generation > of > > the > > > > interrupt and the actual ISR invocation. > > > > > > > > > > > > Whether use of Xenomai will enable us meet this timing reqirement. > > > > > > The latencies I have observed so far on ARM are usually larger than 150 > > > microseconds, but these are userspace dispatch latencies. > > > > > > So, you could improve situation if you stay in interrupt handlers. > > > > > > Another way to improve the situation a bit more is to use ucLinux, if > > > your platform is supported. > > > > > > Still, IMHO, 40-50 microseconds is too ambitious. > > > > Gilles, as I understood the question was about the interrupt latency, not > > the scheduling one. > > > > I guess, the vitually tagged cache is an additional component of high > > scheduling latencies on ARM. > > > > The interrupt latency should be ok though. > > Interrupt handler reside in cached memory as well. And a source of high > latencies on ARM is the fact that the TLB is flushed with interrupts > off. So, even if the interrupt handler was in TCM to avoid latencies > induced by the cache, interrupts off sections would still cause high > latencies. So, the only chance of success is to use uCLinux with the > interrupt handler in TCM.
I think implementing coloured caches (with reservations for RT processes) could be an option as well. Once RT context switches no longer require full cache flushes, those for non-RT processes could be made interruptible. But all this would require heavy Linux hacking, I'm afraid. Jan
signature.asc
Description: OpenPGP digital signature
_______________________________________________ Xenomai-core mailing list Xenomai-core@gna.org https://mail.gna.org/listinfo/xenomai-core