[ovs-dev] PMD threads starting/stopping strategy

[Mauricio Vásquez]

Thanks Ilya for your answer,
I'll try to give a look to your patch-set this week.

On Fri, May 20, 2016 at 11:55 AM, Ilya Maximets <i.maximets at samsung.com>
wrote:

> Hi, Mauricio.
>
> My thoughts about TX queue management described in patch-set
> "[PATCH RFC 0/6] dpif-netdev: Manual pinnig of RX queues + XPS."
> ( http://openvswitch.org/pipermail/dev/2016-May/070902.html ).
>
> Shortly:
> My solution is to allow user to set number of TX queues
> for each device (options:n_txq) just like it done for number of
> RX queues (options:n_rxq). PMD threads will choose appropriate
> TX queue dynamically using some kind of XPS logic.
>
> This solution, I think, is more convenient for users and it allows
> to solve various issues connected to static 'tx_qid' distribution.
>
> Manual setting of 'n_txq' will allow users to manage performance
> of OVS more accurately and also simplifies the configuration code.
>
> Best regards, Ilya Maximets.
>
> On 20.05.2016 12:23, Mauricio Vásquez wrote:
> > Hello,
> >
> > I noticed that pmd threads are created in a per NUMA node fashion,
> > it means, pmd threads are only created in a specific NUMA node when
> > there is a least one DPDK port on that node.
> >
> > My concern is that in some cases with this approach more threads than
> > needed are created, for example, lets suppose a system with two ports
> > with a single rx queue, in this case maximum 2 pmd threads would be
> > necessary, but if the user has set a pmd core mask including more than
> > two cores all of them would be created. (some cores would do "nothing",
> > just poll if they have to be restarted).
> >
> > Is there any particular reason for this behavior?
> > Would it make sense to consider starting pmd threads on demand?, I mean,
> > only when they are actually needed.
> >
> > Another related question is about the number of tx queues, I noticed
> > that it is set to n_cores + 1, where n_cores is the total number of
> > cores in the system. I think it is not common to have all cpu cores
> > are assigned to ovs, then many of those queues will be not used.
> >
> > Does it make sense to create the number of tx queues based on the
> > core_mask, in a way that all of the queues are actually used?
> >
> > Thank you very much for the attention,
> >
> > Mauricio V,
>
>