[ovs-dev] [RFC PATCH 0/8] dpif-netdev: Refactor cycle count and rebased patches

Stokes, Ian ian.stokes at intel.com
Mon Jan 8 16:09:24 UTC 2018 

> Hi Ian,
> 
> That's why I brought up my original question before Christmas, but
> apparently too late ☹.

Apologies, Christmas was a bit hectic with the last push for output batching and finishing for the holidays so I missed following up on it.

> 
> Myself, was hoping that we could get all three changes: PMD performance
> metrics, time-based tx batching and Kevin's enhancement for the pmd-rxq-
> show command into 2.9.
> 

I think this is the ideal situation, but does require sign off from Ilya and Kevin as it will be their features it will impact on, I guess they can answer and give an idea on bandwidth to cooperate on something like this.

> PMD performance metrics are for sure around long enough to warrant that.
> And they are highly wanted too.

To my mind the detailed PMD logs patchset has a lot of content to work through and there are reviews in progress as well as re-work requests from previous reviews of the v4. I was waiting to see these completed before focusing on it myself as I haven't had the bandwidth to date to take it on.

> 
> Since all three are heavily affecting same parts of the code the order of
> merging matters a lot. If we want to make this happen in reasonable time
> we need to avoid constant manual re-basing for all of us.
> 
> That’s why I have taken the initiative to serialize them into one
> particular order for merging. That was a painful exercise and I am not
> looking forward to doing it again in a different order.

Agreed and thanks for taking the initiative. I do think this is important if all three are to make it in.

> 
> So I would greatly appreciate if we could agree on the proposed order and
> discuss how we review/test the resulting overall contribution with the
> ambition to get all parts into 2.9.

I'm open to others input here, based on the points you've raised I would suggest the following (only a suggestion, please feel free to counter):

1: Output Time batching (it's v9 and is well understood at this point, I would think would be little re-work needed if the cases suit the PMD balancing already in place).
2: Detailed PMD logs (from what I understand there is a review in press from Billy, and re-work requests from Aaron, we could roll these changes into the next rebase on top of the time output batching).
3: Kevin's PMD patches: probably the smallest of the set and lowest risk IMO as you seem familiar already with these Jan?

Does this seem acceptable?

It's probably a good topic for the community meeting Wednesday but I'd like to try and get agreement before then if we are to coordinate to get these upstreamed.

Ian
> 
> Thanks, Jan
> 
> > -----Original Message-----
> > From: ovs-dev-bounces at openvswitch.org
> > [mailto:ovs-dev-bounces at openvswitch.org] On Behalf Of Stokes, Ian
> > Sent: Monday, 08 January, 2018 16:04
> > To: Jan Scheurich <jan.scheurich at ericsson.com>; Kevin Traynor
> > <ktraynor at redhat.com>; Ilya Maximets <i.maximets at samsung.com>
> > Cc: dev at openvswitch.org
> > Subject: Re: [ovs-dev] [RFC PATCH 0/8] dpif-netdev: Refactor cycle
> > count and rebased patches
> >
> > > Hi guys,
> > >
> > > It would be great to get your feedback on the proposal for combining
> > > (and
> > > simplifying) our patches.
> > >
> > > Do you agree with a) the cycle counting refactoring and b) the
> > > simplification of rxq pmd load calculation?
> > >
> > > For actual merge I would suggest to re-order the changes and take
> > > the cycle counting refactoring before the time-based output
> > > batching. But that is perhaps just a matter of taste affecting some
> > > intermediate commits and not so important for the end-result...
> > >
> > > How should we proceed with including these to the dpdk_merge branch?
> > > - One patch series at a time in the order now laid out in my RFC
> > > series
> > > - One large series comprising all 4 contributions?
> > >
> >
> > Hi Jan,
> >
> > From my side I was hoping to review/test Ilya's v9 time based output
> > batching first with a view to upstreaming that feature for the 2.9
> release.
> >
> > My worry was that by attaching it to other feature changes it may not
> get up streamed.
> >
> > I've only had a cursory look at the Kevin's percentage pad patches.
> >
> > I was thinking of working to the following on dpdk_merge
> >
> > 1: Review/Upstream Ilya's time based output.
> > https://patchwork.ozlabs.org/project/openvswitch/list/?series=19865
> > 2: Review/Upstream Kevin's
> > https://patchwork.ozlabs.org/user/todo/openvswitch/?series=18301
> >
> > Just a question, in your mail below I see 'Detailed PMD performance
> > metrics and supervision' is included in [1], Billy O'Mahony is
> > currently reviewing the latest revision of this from what I'm aware of,
> this a pre-requisite for you before upstreaming Ilya and Kevin's patches?
> >
> > I'd like to hear from Ilya on Kevin on this also? If they are happy to
> > combine the work and review/validate as a group then it may be possible.
> >
> > Thanks
> > Ian
> >
> > > Regards, Jan
> > >
> > > > -----Original Message-----
> > > > From: jan.scheurich at web.de [mailto:jan.scheurich at web.de] On Behalf
> > > > Of Jan Scheurich
> > > > Sent: Thursday, 04 January, 2018 21:03
> > > > To: dev at openvswitch.org
> > > > Cc: Jan Scheurich <jan.scheurich at ericsson.com>
> > > > Subject: [RFC PATCH 0/8] dpif-netdev: Refactor cycle count and
> > > > rebased patches
> > > >
> > > > This RFC patch series contains three contributions:
> > > >
> > > > 1. A rebase of Ilya's series "[PATCH v9,0/5] Output packet
> > > > batching
> > > (Time-based)."
> > > > (http://patchwork.ozlabs.org/cover/852003/) on top of "[PATCH
> > > > v5,0/3]
> > > dpif-netdev:
> > > > Detailed PMD performance metrics and supervision"
> > > (http://patchwork.ozlabs.org/cover/855572/).
> > > >
> > > > 2. Refactoring and simplification of the PMD cycle counting id
> > > > dpif-
> > > netdev.c.
> > > >
> > > > 3. A rebase and simplification of Kevin's patches
> > > > (http://patchwork.ozlabs.org/patch/847972/)
> > > > to display PMD usage per queue in the ovs-appctl pmd-rxq-show
> command.
> > > >
> > > > The patches pass check_patch and "make check" and I have done some
> > > > basic tests of the simplified rxq cycle counting and the PMD usage
> > > reporting in pmd-rxq-show.
> > > >
> > > > This is my proposal how to combine the three existing patch series
> > > > together with the simplified cycle counting the into branch
> > > > dpdk_merge
> > > for release in OSV 2.9.
> > > >
> > > > Before merging the last two patches should probably be combined.
> > > >
> > > >
> > > > Ilya Maximets (5):
> > > >   dpif-netdev: Use microsecond granularity.
> > > >   dpif-netdev: Count cycles on per-rxq basis.
> > > >   dpif-netdev: Time based output batching.
> > > >   docs: Describe output packet batching in DPDK guide.
> > > >   NEWS: Mark output packet batching support.
> > > >
> > > > Jan Scheurich (2):
> > > >   dpif-netdev: Refactor cycle counting
> > > >   dpif-netdev: Add percentage of pmd/core used by each rxq.
> > > >
> > > > Kevin Traynor (1):
> > > >   dpif-netdev: Reset the rxq current cycle counter on reload.
> > > >
> > > >  Documentation/howto/dpdk.rst         |  12 ++
> > > >  Documentation/intro/install/dpdk.rst |  58 ++++++
> > > >  NEWS                                 |   3 +
> > > >  lib/dpif-netdev.c                    | 350 ++++++++++++++++++++++--
> ----
> > > -------
> > > >  tests/pmd.at                         |  51 +++--
> > > >  vswitchd/vswitch.xml                 |  16 ++
> > > >  6 files changed, 349 insertions(+), 141 deletions(-)
> > > >
> > > > --
> > > > 1.9.1
> >
> > _______________________________________________
> > dev mailing list
> > dev at openvswitch.org
> > https://mail.openvswitch.org/mailman/listinfo/ovs-dev

More information about the dev mailing list