This is an option - to set the affinity -of the main thread to another
core(other than the first core) - that is a datapath core.
But that means to update the cunnit common and set the affinity before the call
to odp_init_local. The calling sequence will be
Something like this:
odph_linux_cpumask_default(&cpumask, 0);
cpu = odp_cpumask_first(&cpumask);
cpu = odp_cpumask_next(&cpumask, cpu);
odp_cpumask_zero(&cpumask);
odp_cpumask_set(&cpumask, cpu);
...
if (0 != odp_init_local()) {
But I didn't look further to see if this will affect other validation tests
that use the cunnit_common.
As I said, according to your suggestion, from the classification tests
perspective this minimal change will fix the schedule approach.
Regards,
Radu
-----Original Message-----
From: Balasubramanian Manoharan [mailto:[email protected]]
Sent: Wednesday, February 11, 2015 2:42 AM
To: Bulie Radu-Andrei-B37577
Cc: [email protected]
Subject: Re: [lng-odp] classification tests scheduled queues
On 10/02/15 11:28 pm, Balasubramanian Manoharan wrote:
>
> On 10/02/15 5:37 pm, Radu-Andrei Bulie wrote:
>> Hi,
>>
>> Regarding my comments, I want to know how it will be proceeded. Will
>> be there any patches that will update the current classification
>> tests to support the poll and updated scheduled queues version?
> We will update the validation suite to test for both schedule and poll
> based queues.
>
> Regards,
> Bala
Hi,
In case of schedule based queues can you set the core affinity on the
application to run only on data path cores.This should solve the issue which
you had described earlier in this mail.
Regards,
Bala
>>
>> Regards,
>>
>> Radu
>>
>> -----Original Message-----
>> From: Bulie Radu-Andrei-B37577
>> Sent: Tuesday, February 03, 2015 3:59 PM
>> To: 'Jerin Jacob'
>> Cc: [email protected]; '[email protected]';
>> 'Balasubramanian Manoharan'
>> Subject: RE: [lng-odp] classification tests scheduled queues
>>
>> Hi,
>>
>> I think it's a good approach. One could see the functional test
>> behavior using either poll or schedule on the classification path.
>>
>>
>>
>> -----Original Message-----
>> From: Jerin Jacob [mailto:[email protected]]
>> Sent: Tuesday, February 03, 2015 3:29 PM
>> To: Bulie Radu-Andrei-B37577
>> Cc: [email protected]
>> Subject: Re: [lng-odp] classification tests scheduled queues
>>
>> On Tue, Feb 03, 2015 at 01:07:15PM +0000, Radu-Andrei Bulie wrote:
>>> As I said in my comment we could use two approaches.(poll or data
>>> path thread).
>> How about the scheme(#define IPSEC_POLL_QUEUES) followed in existing
>> example/ipsec to abstract polled vs schedule mode ?
>>
>>> The one you mentioned reflects the model of the reference applications.
>>> Classification test itself, for the present moment, is a functional
>>> test and does not target performance. So I don't see a major
>>> difference in using poll or a data path thread.
>>>
>>>
>>> Regards,
>>>
>>> Radu
>>>
>>>
>>>
>>> -----Original Message-----
>>> From: Ola Liljedahl [mailto:[email protected]]
>>> Sent: Tuesday, February 03, 2015 2:59 PM
>>> To: Bulie Radu-Andrei-B37577
>>> Cc: [email protected]
>>> Subject: Re: [lng-odp] classification tests scheduled queues
>>>
>>> Wouldn't it be better to modify the validation program to make sure
>>> scheduling is always performed on a data path thread?
>>> We want to promote usage of the scheduler, HW-accelerated
>>> classification and scheduling are some of the differentiators of ODP.
>>>
>>> On 3 February 2015 at 13:53, Radu-Andrei Bulie
>>> <[email protected]> wrote:
>>>> Hi,
>>>>
>>>>
>>>>
>>>> I have a comment regarding the scheduling approach in the
>>>> classification validation tests (this also can be extended to other
>>>> tests which use the same pattern).
>>>>
>>>> The schedule function should be called in context of a data path
>>>> thread that is bound to a known core (the same model as in pktio
>>>> application for instance). Otherwise
>>>>
>>>> (as in the mentioned test) the main process (in this case the cunit
>>>> test) can be scheduled by Linux on any core – e.g core 0 – that is
>>>> not in the data path. In this situation no dequeue will occur,
>>>> because the scheduling does not take place on a data path thread
>>>> (unlike the case for pktio application where there is a cpu
>>>> dedicated to control path).
>>>>
>>>> From the linux-generic perspective, there is no apparent issue in
>>>> using the schedule function in this context. But as it is given in
>>>> the reference application, the purpose of the scheduling is to
>>>> function on the data path and thus to provide the advantages given
>>>> by different SoCs acceleration implementations. (there will always
>>>> be a control core and some data path cores).
>>>>
>>>> Thus, being in accordance with the reference applications from odp,
>>>> I suggest replacing the scheduled queues with poll queues or create
>>>> a separate thread which receives the packet, otherwise the test
>>>> will function only on linux generic implementation.
>>>>
>>>>
>>>>
>>>> PS: I could send the patch (using poll queues approach) if we reach
>>>> a consensus.
>>>>
>>>>
>>>>
>>>>
>>>>
>>>> Regards,
>>>>
>>>>
>>>>
>>>> Radu
>>>>
>>>>
>>>> _______________________________________________
>>>> lng-odp mailing list
>>>> [email protected]
>>>> http://lists.linaro.org/mailman/listinfo/lng-odp
>>>>
>>> _______________________________________________
>>> lng-odp mailing list
>>> [email protected]
>>> http://lists.linaro.org/mailman/listinfo/lng-odp
>
_______________________________________________
lng-odp mailing list
[email protected]
http://lists.linaro.org/mailman/listinfo/lng-odp
