Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2019-06-11
15:07:19 sean-k-mooney kashyap: my -1 is mainly for the extraspec names an di have one question regarding if post copy and auto converge can be used togeter at teh qemu/libvirt layer
15:07:21 kashyap sean-k-mooney: Thanks. Didn't read everything yet; need a fresher mind.
15:07:38 kashyap sean-k-mooney: I asked Dave (PostCopy author) to also comment, FWIW.
15:07:51 sean-k-mooney kashyap: cool
15:07:53 kashyap sean-k-mooney: Another thing is, how does one prioritize between PostCopy and AutoConverge
15:08:24 sean-k-mooney kashyap: currently its states its an error to use both
15:09:05 sean-k-mooney aparently nova perfers post copy over auto converge
15:09:45 sean-k-mooney although if i rememebr correctly danpb prefered the idea of having auto converge enabled by default in nova and was less confindent in post copy
15:09:57 kashyap 17:08 < kashyap> dgilbert: Can you use PostCopy and AutoConverge at the _same_ time, Dave?
15:09:57 kashyap sean-k-mooney: That's the response:
15:10:00 kashyap 17:08 < dgilbert> hmmmm
15:10:03 kashyap 17:09 < dgilbert> if you can I don't think we've ever tested it
15:10:40 kashyap sean-k-mooney: The thing is, some workloads will be "surprized" in a bad way to realize AutoConverge is auto-enabled
15:10:43 cfriesen it doesn't make sense to use postcopy and autoconverge at the same time
15:11:08 kashyap cfriesen: Yeah, it doesn't. One needs to know WTF they're doing before they start trying things
15:11:09 sean-k-mooney kashyap: ya i know things like watcher
15:11:11 cfriesen once you've enabled postcopy it'll naturally slow down as it pulls pages over the network, no need to artificially slow it down to allow to converge
15:11:49 sean-k-mooney cfriesen: well autoconverge only slow it down if it failed to copy less the half of the outstanding pages in the last interval
15:11:51 kashyap cfriesen: Yeah, the unsatisfying answer, as usual is, "it depends on the workload"
15:12:07 sean-k-mooney and it does not slow down the workload intially untill a certin time has passed
15:12:27 cfriesen sean-k-mooney: details. :)
15:12:48 stephenfin printf-debugging++
15:12:55 sean-k-mooney if you are dirtiying memory fast enough i dont know if post-copy is required to ever finish or if write to remote memory will always be paged to local memory instead
15:13:15 cfriesen sean-k-mooney: once you've written to all local pages, then post-copy is done. :)
15:13:19 sean-k-mooney if post-copy works by treating the remote meory as read only then its fine
15:13:25 cfriesen sean-k-mooney: it does
15:13:36 sean-k-mooney ok then ya no need to use both
15:13:44 cfriesen sean-k-mooney: but if you lose the network while in post-copy, you're hosed
15:13:53 sean-k-mooney yep i know
15:14:19 sean-k-mooney but if you looes netwroking with shared storage your equally screwed
15:16:01 artom We're allowed methods in neutronv2/api.py that aren't in network/api.py, right?
15:16:16 sean-k-mooney private ones yes
15:16:16 artom Or is this like virt drivers, they all need to match?
15:16:24 artom Public ones
15:16:30 sean-k-mooney what methods are you adding
15:16:43 artom The get_events stuff, see convo with dansmith above
15:16:43 sean-k-mooney that moduel shoudl really only have methods that call neutron
15:17:01 sean-k-mooney its does not interact with the api
15:17:08 sean-k-mooney so why would it be in that module
15:17:30 artom TBH, because dansmith said so, and it makes sense to me as well
15:17:37 artom But mostly because he said so ;)
15:17:50 artom This is not something I have the energy to debate
15:17:59 dansmith artom: don
15:18:00 dansmith artom: don't hedge :)
15:18:44 efried stephenfin: I'm stumped. Are you getting anywhere? If not, I'm going to call time and just disable the test for now.
15:18:54 sean-k-mooney dansmith: why in the api module and not here https://github.com/openstack/nova/blob/master/nova/network/model.py#L460
15:18:58 stephenfin efried: Yeah, I'm down in pycodestle
15:19:02 stephenfin *pycodestyle
15:19:17 dansmith sean-k-mooney: I said nova/network/something would be fine too
15:19:47 artom I wasn't aware of the "must call neutron API" rule
15:20:04 efried stephenfin: Was that "yeah I'm getting somewhere" or "yeah disable the test"?
15:20:14 dansmith artom: I'm aware of no such rule
15:20:25 sean-k-mooney artom: im not sure it a rule but that module is nova.network.neutronv2.api
15:20:37 stephenfin efried: Disable it. I'll keep looking
15:20:37 artom sean-k-mooney, actually yeah, with the new abstraction suggested by dansmith (plug-time vs bind-time), it fits in model as well
15:20:43 stephenfin Here's our issue though https://github.com/PyCQA/pycodestyle/blob/2.0.0/pycodestyle.py#L1639
15:20:57 stephenfin We shouldn't be getting there but I don't know yet what's changed
15:20:59 sean-k-mooney pretty much everything it does interacts with the neutron api in one way or another
15:21:30 dansmith sean-k-mooney: we have constants defined in that module, and this detail is completely related to "what to expect from the neutron api" so I have no problem saying it should (or can) be in there
15:21:34 stephenfin Things like this are why I was trying to get us onto a newer hacking version. Things invariably break
15:21:49 sean-k-mooney artom: plugtime and bind time is more or less what i was thinking about in the neutron spec too so that makes sense to me
15:21:49 efried stephenfin: aren't we on a newer hacking version at this point?
15:22:03 stephenfin not really. Hacking is relying on a really old version of flake8
15:22:08 artom kawn sen suss!
15:22:11 stephenfin Like, super old
15:23:09 stephenfin and I can't update because they broke all the things in flake8 3.x :( https://gitlab.com/pycqa/flake8/issues/545
15:23:46 sean-k-mooney ok well its not where i would instintivly look for them but i can live with it.
15:25:05 sean-k-mooney artom: to your orginal question https://github.com/openstack/nova/blob/master/nova/network/api.py isthe old nova networks api and the neton one extends it
15:25:16 sean-k-mooney i think you can ignore https://github.com/openstack/nova/blob/master/nova/network/api.py at this point
15:25:48 artom sean-k-mooney, ack, thanks for looking :)
15:26:04 artom (Though tbf my original question was rendered moot when we agreed to put it in model)
15:27:31 sean-k-mooney artom: :) by the way i proposed https://review.opendev.org/#/c/664442/ to test this
15:28:13 artom sean-k-mooney, yep, thanks for that
15:30:17 openstackgerrit Eric Fried proposed openstack/nova master: Skip test_check_doubled_words hacking check UT https://review.opendev.org/664622
15:30:20 efried stephenfin: ^
15:35:24 efried dansmith: mind unblocking the gate ^
15:45:26 openstackgerrit Adam Spiers proposed openstack/nova master: Add extra spec parameter and image property for memory encryption https://review.opendev.org/664420
15:45:39 aspiers efried: vastly improved patchset based on your feedback ^^^
15:47:31 openstackgerrit Sylvain Bauza proposed openstack/nova-specs master: Proposes NUMA affinity for vGPUs https://review.opendev.org/650963
15:48:26 efried aspiers: looking.
15:49:51 stephenfin efried, cdent: Here's what's happening
15:50:12 stephenfin Because of https://bugs.python.org/issue33899, we now insert a NEWLINE token at the end of all lines
15:50:26 stephenfin that manifests itself in pycodestyle
15:50:45 stephenfin which is interating through all the tokens on the test line
15:50:46 stephenfin https://github.com/PyCQA/pycodestyle/blob/2.0.0/pycodestyle.py#L1616
15:51:45 stephenfin previously, when we got to the end of the line we were falling through to this case https://github.com/PyCQA/pycodestyle/blob/2.0.0/pycodestyle.py#L1650-L1652
15:52:09 stephenfin but now, because we see this newline token, we hit this case instead https://github.com/PyCQA/pycodestyle/blob/2.0.0/pycodestyle.py#L1632-L1635
15:52:26 stephenfin note that that case doesn't include a call to 'check_physical'
15:52:33 stephenfin so our test never gets run
15:53:06 stephenfin I don't know how to fix that. It feels like this is now a bug in pycodestyle
15:54:08 stephenfin I note that the latest version of pycodestyle does pretty much the same thing too https://github.com/PyCQA/pycodestyle/blob/2.5.0/pycodestyle.py#L2104-L2107
15:54:33 cdent stephenfin: sorry stephenfin, I thought most of what you said above was already understood: and it being a bug on pycodestyle was known
15:54:49 cdent or rather: pycodestyle's continued behavior is the root
15:54:56 stephenfin Um, whut? :)
15:55:02 stephenfin Clearly I didn't read far enough
15:55:37 cdent Or maybe I just reached that conclusion based on the discussion on https://bugs.python.org/issue33899 which if I recall right suggests the same thing
15:56:37 stephenfin Yup, now that I search for pycodestyle I see the line of course. Dang it
15:56:47 stephenfin Their fix is rather odd though https://github.com/PyCQA/pycodestyle/pull/801/files
15:58:06 openstackgerrit Adam Spiers proposed openstack/nova master: Pass extra_specs to flavor in vif tests https://review.opendev.org/662556
16:01:45 aspiers efried: What's supposed to happen when hw:mem_encryption=True but trait:HW_CPU_X86_AMD_SEV=required is missing?

Earlier   Later