[ovs-discuss] pmd-cpu-mask/distribution of rx queues not working on windows

Kevin Traynor ktraynor at redhat.com
Wed Nov 22 17:47:38 UTC 2017


On 11/14/2017 04:43 AM, Kevin Traynor wrote:
> On 11/14/2017 02:16 AM, aserdean at ovn.org wrote:
>>> Subject: Re: [ovs-discuss] pmd-cpu-mask/distribution of rx queues not
>>> working on windows
>>>
>>> On 10/19/2017 05:45 PM, Alin Gabriel Serdean wrote:
>>>> Hi,
>>>>
>>>>
>>>>
>>>> Currently the test “pmd-cpu-mask/distribution of rx queues” is failing
>>>> on Windows. I’m trying to figure out what we are missing on the
>>>> Windows environment. Any help is welcomed 😊.
>>>>
>>>>
>>>
>>> Hi Alin, the queues are sorted by measured rxq cycles since 79da1e411ba5. In
>>> this test case the rxq cycles are equal for all queues.
>>>
>>> On Linux, the sort result is 0,1,3,4,5,6,7,8,9 whereas you are seeing
>>> 1,3,4,5,6,7,8,9,0. This impacts the distribution to PMDs and that's causing the
>>> test fail.
>>>
>>> Currently the comparison function (rxq_cycle_sort) selects a winner when
>>> they are equal. I have submitted a patch which (amongst other things)
>>> changes that so it will just report they are equal. I've sent a v2,
>>> https://patchwork.ozlabs.org/patch/835385/ can you try that and see if it
>>> fixes the issue on Windows?
>>>
>>> thanks,
>>> Kevin.
>>>
>> Thanks a lot for the input Kevin.
>> I have applied the patch (https://patchwork.ozlabs.org/patch/835385/)
>> and reran the test.
>> Unfortunately, the same result:
>> "
>> +++ /c/ovs/tests/testsuite.dir/at-groups/1007/stdout  2017-11-14 04:07:44 +0200
>> @@ -1,7 +1,7 @@
>>  pmd thread numa_id <cleared> core_id <cleared>:
>>         isolated : false
>> -       port: p0        queue-id: <cleared>
>> +       port: p0        queue-id: 0 1 4 5
>>  pmd thread numa_id <cleared> core_id <cleared>:
>>         isolated : false
>> -       port: p0        queue-id: <cleared>
>> +       port: p0        queue-id: 2 3 6 7
>> "
>> Can I provide you any additional information from the logs?
>>
> 
> No need, it looks like the same failure. ok, in that case I'll send a
> patch to add a tiebreaker sort on port/queue, to happen when the cycles
> are the same. That *should* explicitly force the right order.
> 

Hi Alin, I've just sent an RFC that adds a port/queue tiebreaker for
comparisons between two queues that have the same amount of measured
cycles. https://patchwork.ozlabs.org/patch/840458/

If you could get a chance to test it, that would be great. It should
force the sort order - if it doesn't, we'll need to think about what
else can be done. Perhaps we could take a different approach and modify
the unit test based on OS.

thanks,
Kevin.

> Kevin.
> 
>> Thanks,
>> Alin.
>>
> 
> _______________________________________________
> discuss mailing list
> discuss at openvswitch.org
> https://mail.openvswitch.org/mailman/listinfo/ovs-discuss
> 



More information about the discuss mailing list