[ovs-dev] [RFC 0/5] Quicker pmd threads reloads

David Marchand david.marchand at redhat.com
Tue May 7 15:07:19 UTC 2019


Hello Ilya,

Thanks for looking at this series.

On Mon, May 6, 2019 at 5:22 PM Ilya Maximets <i.maximets at samsung.com> wrote:

> On 30.04.2019 15:17, David Marchand wrote:
> > We have been testing the rebalance code in different situations while
> > having traffic going through OVS.
> > Those tests have shown that part of the observed packets losses is due to
> > some time wasted in signaling/waiting for the pmd threads to reload their
> > polling configurations.
> >
> > This RFC series is an attempt at getting pmd threads reloads quicker and
> > more deterministic.
> >
>
> Do you have some performance data to share?
>

During our testing of rebalance, we were tracking packets losses with
traffic running during a rebalance.

I focused on the cycles spent in the transition between two polling
configurations.
I triggered 1000 rebalances on each patch of this series, with rte_rdtsc()
probes in reconfigure_datapath() / pmd_thread_main().
Between the moment we stop polling in a pmd for a given configuration and
the moment we start polling again in the new configuration:
- before the patches, a pmd would spend 140k/330k/3 000k cycles
(minimum/average/maximum).
- after the patches, a pmd would spend 13k/20k/43k cycles.

Originally, the numbers are highly volatile: I got a 20 000k cycles in a
previous test run (which translated to 28k lost packets with the ofrules I
had).


I had a look at your comments, I will come with the fixes later this week
(tomorrow is off in France).

-- 
David Marchand


More information about the dev mailing list