Re: [Xenomai-help] Known-to-work motherboards?

[Jan Kiszka <jan.kiszka@domain.hid>]

[-- Attachment #1: Type: text/plain, Size: 1474 bytes --]

Gilles Chanteperdrix wrote:
> Jan Kiszka wrote:
>> Gilles Chanteperdrix wrote:
>>> Jan Kiszka wrote:
>>>> Gilles Chanteperdrix wrote:
>>>>> Stuart O Anderson wrote:
>>>>>> If there isn't a list online that I've missed, is anyone running
>>>>>> xenomai on a relatively recent machine with 8 or more cores without
>>>>>> problems?  If so, can you provide hardware specs?
>>>>> Xenomai will have scalability problems when running on an 8 cores system
>>>>> (if you intend to run real-time tasks on each of the 8 cores).
>>>>>
>>>> We are running Xenomai on up to 4x4 boxes, but only with 1 or 2 cores
>>>> used by RT tasks. I've some patch hanging around that restricts Xenomai
>>>> (specifically its host timer hook) to a CPU subset in order to mitigate
>>>> the scalability issues - guess I should finally get it in shape and post it.
>>> The host timer already runs on only one CPU.
>> Not on x86.
> 
> Some time ago, when using the APIC (which was always the case on x86
> SMP), the host timer was not even started, and the Linux APIC irqs were
> triggered by a custom 8254 irq handler. Well, it seems things have
> changed since then...

...around the Linux clocksource/clockevent rework. And it also makes
more sense with the number of cores increasing and NUMA becoming
standard. CPU-local timers are simply more efficient. That's why the
nklock hurts that any host tick must pass even if the core is not used
by Xenomai otherwise.

Jan


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 257 bytes --]