Comment 25 for bug 1539775

Hi,
thanks for your reply - it improved my expectation a bit, but not in all regards.
I'll slim down the content of the discussion a bit and try to focus on the important pieces.

> 1. Start with Ubuntu 16.10 Tech Preview and support only Mellanox adapter (requires MLNX_OFED)

As I explained before - the mlx pmd drivers in dpdk can't be built.
This can't be fixed in DPDK (well it might be), but the right solution is getting Mellanox to properly upstream their bits into libverbs and such.

> 2. In Ubuntu 17.04 expand the Tech Preview to include additional adapters
> 3. Lift the Tech preview in 17.10

> That should hopefully erase some of the concerns and ease acceptance. We are committed to supporting DPDK on Power.

I like the step by step approach in general and I trust you in your commitment - thanks!

> One of the things we discovered during the course of our testing is that package ovs-dpdk has not been enabled on Power in the ppa.

I can add that to the ppa, but this is just one more example of this likely isn't ready.
Openvswitch has no enablement for dpdk on ppc yet. I can add it for you, and provide it in the ppa but I highly doubt we want to make that change as well so lat in the cycle.

> I understand some of your concerns. It is precisely to facilitate thorough testing that we are requesting a Tech Preview. The request is to release note it with a "Tech Preview" label only for the Mellanox devices. In this way, a customer will not attempt to use other devices.

I was at IBM long enough to know the Tech Preview business - I just haven't found a Ubuntu way of doing this for "parts of a package".
I mean it is easy to have let's say a package "ppc-special-tool-X" as preview.
But dpdk and more so openvswitch are in main and important - and I can't easily make different support statements per arch afaik.

>> - take over responsibility the power related issues for the 16.10 support
>> time

> Yes, we do understand the commitment required from IBM and previous responses should provide the assurance you are seeking.

Thank you, that is one thing I wanted to hear.

> Given the previous explanations, can you move these to the main archive? That would be our preference. This will also enable additional run time outside of IBM. Please let us know what you think, and what are the next steps moving forward.

Yes I think working together we could put it in the archive, but not in 16.10 with all you wanted.
The following can't be met:
- dpdk mlx pmd drivers which can't be made available
- I doubt you'll get the openvswitch dpdk enablement on ppc64el that late either

That said you have two options:
A)
- enable ppc64el in dpdk in 16.10
  - without MLX drivers
  - without changing openvswitch
  - the impact to other architectures should be zero, but one can never be sure
  - drive a FFE very late in the cycle
  - on such a short runway I expect issue to show up which will be unnecessary work and bad publicity for both of us

B)
- enable ppc64el in dpdk in 16.10 via a ppa
  - without MLX drivers initially
    - but if you later on can provide a DPDK patch that lets it build with upstream libibverb I surely can add them to the ppa
  - also provide a dpdk ppc64el enabled openvswitch to the same ppa
  - the impact to other architectures now IS zero
  - the ppa is what makes it a "tech preview" IMO
    - here it is perfectly fine to add custom patches/fixes if needed
    - do not consider ppa a bad thing, many projects have "official ppa's" and many people use packages from ppa's on a daily base.

For both options we would then go on:
- Right when 17.04 opens up we will
  - enable dpdk on ppc64el
  - enable dpdk support on ppc64el in openvswitch
  - If we get things reasonably stable 17.04 will be all fine and released with it
  - if things are bad, we identify roadblocks, anybody looses funding or anything like it we can disable them again before 17.04 is released

I clearly prefer B). Because A) will leave you without any dpdk exploiters (ovs) and without the device drivers you wanted in 16.10.
That might only be worth if that was on someones scorechart, but not for any real use by a customer or tester.
B clearly is better for everyone as it has all you need in 16.10 as a ppa and we can work on 17.04 together (earlier than after feature freeze this time) to get it in proper shape to be in the release itself.

If you, despite all tries to explain why B is better than A still want A - could you please tell me what the benefit of A would be over B? I really would like to see that, but I don't so far.

You asked for next steps:
- Next step if you choose A is that you add a https://wiki.ubuntu.com/FreezeExceptionProcess here in the bug to convince the release Team.
- Next step if you chose B (and I hope you do) is give me a few days to enable dpdk/ppc in openvswitch for you as well, I'll modify the bug here accordingly for all further things needed towards 17.04.