On 06/20/2011 06:16 PM, Philippe Gerum wrote:
> On Mon, 2011-06-20 at 16:25 +0100, Jakub Nowacki wrote:
>> On 06/20/2011 04:02 PM, Gilles Chanteperdrix wrote:
>>> To make proper benchmarks, you should:
>>> - let the latency run long
>>> - provide some load, you can use the dohell script in xenomai-head to
>>> generate the load.
>>>
>> I've generated a load using 'dd if=/dev/zero of=/dev/null' one for each
>> core (8 in total), which is roughly the way how xeno-test does it. The
>> test was running for 30 min, which is not very long but I should see the
>> long latencies by that time, I think at least.
>>
>>> That said, having some spikes is not abnormal. The real question is to
>>> know whether the latencies your get are sufficient for your application.
>>>
>> Well, I think it should be OK for peak latencies < 20 us, that is what
>> I'm getting now.
>>
>>>> ioctl(RTTST_RTIOC_SWTEST_CREATE_KTASK): Cannot allocate memory
>>>
>>> You have to increase CONFIG_XENO_OPT_SYS_STACKPOOLSZ.
>>>
>> It is 128 Kb now. Should it be more? Is there any (thumb) rule for this
>> value?
>
> Max # of concurrent kernel-based Xenomai threads x stack_size. 8k stack
> minimum is a safe bet for x86_64.
>
> Typically, RTDM (driver) tasks in kernel space are Xenomai ones, so you
> would need to reserve stack space for each of them in this amount.
>
> The switchtest driver is RTDM-based, and may run a number of RTDM tasks,
> depending on the test scenario.
The default scenario being to run 6 rtdm tasks by cpu (that would make
48 on an 8 cores processor).
--
Gilles.
_______________________________________________
Xenomai-help mailing list
[email protected]
https://mail.gna.org/listinfo/xenomai-help
