| Posted | Nick | Remark | |
|---|---|---|---|
| #openstack-nova - 2020-04-01 | |||
| 09:18:02 | brinzhang_ | gibi: ok, I will do by this way, thanks | |
| 09:21:17 | bauzas | gibi: around for a question about reshapes ? | |
| 09:21:27 | bauzas | and your vgpu func test ? | |
| 09:25:14 | gibi | bauzas: sure, hit me | |
| 09:27:17 | bauzas | gibi: okay, so I wonder why you mocked https://github.com/openstack/nova/blob/master/nova/tests/functional/libvirt/test_reshape.py#L96-L101 | |
| 09:27:40 | bauzas | gibi: you started two instances without having any supported types | |
| 09:27:53 | bauzas | gibi: so the instances wouldn't have any VGPU allocations | |
| 09:29:49 | bauzas | oh, but you created some VGPU inventories... | |
| 09:29:52 | gibi | bauzas: before that there is a placement update | |
| 09:29:55 | gibi | yeah | |
| 09:29:59 | gibi | that | |
| 09:30:37 | gibi | I guess this is a simplification | |
| 09:30:39 | bauzas | mmm ok | |
| 09:30:53 | gibi | in a good test the virt driver should report the new vgpu inventory | |
| 09:31:07 | gibi | I don't remember why I don't added a new fake driver that does that | |
| 09:31:13 | gibi | I guess there was complications | |
| 09:31:14 | bauzas | what I honestly wonder is whether your test is valid for a reshape | |
| 09:31:29 | bauzas | because when you restart nova-compute, you get a reshape, right? | |
| 09:31:50 | bauzas | in this case, when allocating an instance for this node, you get the nested RPs | |
| 09:32:45 | bauzas | I mean, we only have VGPU inventories on the root RP for nodes that were not restarted | |
| 09:32:57 | bauzas | (or <Stein) | |
| 09:34:55 | gibi | this test tries to create instances with old type of allocation then hit the reshape by restarting the compute | |
| 09:35:27 | bauzas | I see | |
| 09:35:33 | gibi | if the virt driver report the vgpu at fist start then with the current code reshape is triggered, so I had to simulate and old tree somehow | |
| 09:35:42 | gibi | then trigger a reshape | |
| 09:36:21 | openstackgerrit | Merged openstack/nova master: Add cyborg tempest job. https://review.opendev.org/670999 | |
| 09:36:21 | gibi | see https://github.com/openstack/nova/blob/master/nova/tests/functional/libvirt/test_reshape.py#L53 | |
| 09:36:28 | bauzas | gibi: but you agree with the fact it wouldn't be possible to create an instance having old types of allocations with Stein or Ussuri ? | |
| 09:37:07 | gibi | I agree. But to test reshap in >=Stein I have to simualte an old tree | |
| 09:37:11 | gibi | hence the hack in the test | |
| 09:37:18 | bauzas | I mean, the _allocate_mdev() code in libvirt that's eventually called when creating an instance would be different between Rocky and Stein | |
| 09:37:56 | gibi | that is a valid point, I haven't simulated the Rocky _allocate_mdev code | |
| 09:38:19 | bauzas | I mean, I wrote https://github.com/openstack/nova/blob/master/nova/virt/libvirt/driver.py#L7011-L7013 | |
| 09:38:25 | gibi | I guess creating a proper test is only possible a grenade job that upgrades from Rocky to Stein | |
| 09:39:06 | bauzas | but I think it wasn't needed | |
| 09:39:12 | bauzas | yeah maybe | |
| 09:39:27 | bauzas | anyway, I need to think about what do do | |
| 09:39:49 | bauzas | your test is nice and i don't want to remove it | |
| 09:40:24 | tobias-urdin | continue troubleshooting https://bugs.launchpad.net/oslo.config/+bug/1869929 seems like some issue with a nova dependency or import order issues, have a complete reproduce and a solution, but the solution is to remove the @db.select_db_reader_mode decorator from two functions in nova/objects/service.py | |
| 09:40:25 | gibi | yeah. feel free to remove the func test if it does not make sense any more. The hack there is definitely partially simulate Rocky state | |
| 09:40:26 | openstack | Launchpad bug 1869929 in oslo.config "RuntimeError: maximum recursion depth exceeded while calling a Python object" [Undecided,New] | |
| 09:41:00 | bauzas | gibi: well, I just feel we would just need to create a new instance without asking for VGPUs and then creating the VGPU allocations directly | |
| 09:41:04 | gibi | bauzas: if you remove the func test from master, we will still have it on stable branches where this test is more relevant | |
| 09:41:06 | bauzas | (for the func test) | |
| 09:41:22 | tobias-urdin | ^ if anybody knows import could potentially cause it, trying to track down changes to imports between the versions, seems like nova.cells.rpcapi changed | |
| 09:42:22 | gibi | bauzas: but then mdev assignments will be missing, isn't it? | |
| 09:42:42 | bauzas | ah shit you're right, indeed | |
| 09:42:47 | gibi | bauzas: so that will be partial from a different perspective | |
| 09:42:53 | bauzas | yeah | |
| 09:43:03 | bauzas | anyway, I'll think about it | |
| 09:43:11 | gibi | bauzas: cool. let me know if I can help | |
| 09:43:22 | bauzas | we still need to support the reshape for Ussuri like we said | |
| 09:43:54 | bauzas | so maybe a comment telling it's not really possible to have no parent UUID would be simple | |
| 09:50:14 | openstackgerrit | Johannes Kulik proposed openstack/nova master: Don't recompute weighers' minval/maxval attributes https://review.opendev.org/713863 | |
| 09:50:56 | jkulik | gibi, bug for ^ is here https://bugs.launchpad.net/nova/+bug/1870096 | |
| 09:50:57 | openstack | Launchpad bug 1870096 in OpenStack Compute (nova) "soft-affinity weight not normalized base on server group's maximum" [Undecided,In progress] - Assigned to Johannes Kulik (jkulik) | |
| 09:51:06 | gibi | jkulik: ack, I will check | |
| 09:51:09 | gibi | thanks | |
| 10:06:52 | openstackgerrit | Sylvain Bauza proposed openstack/nova master: Provide the parent pGPU when creating a new vGPU https://review.opendev.org/715489 | |
| 10:06:52 | openstackgerrit | Sylvain Bauza proposed openstack/nova master: WIP: Support different vGPU types per pGPU https://review.opendev.org/715490 | |
| 10:09:58 | bauzas | gibi: if you wanna look at what I wrote about what we discussed ^ | |
| 10:10:11 | bauzas | FWIW, this is a Ussuri target blueprint | |
| 10:10:16 | bauzas | targeted* | |
| 10:10:25 | bauzas | I have to finish the last bits tho | |
| 10:10:31 | bauzas | but that should be okay | |
| 10:29:47 | gibi | bauzas: ack, will check later | |
| 12:21:03 | openstackgerrit | Brin Zhang proposed openstack/python-novaclient master: Microversion 2.85: Change volume-update CLI https://review.opendev.org/712651 | |
| 13:35:05 | sean-k-mooney | brinzhang_: we should still have the info in the release note but im fine with also having a dedicated page in the api guide for this | |
| 13:35:26 | sean-k-mooney | this being cyborg | |
| 13:36:11 | openstackgerrit | Sylvain Bauza proposed openstack/nova master: Provide the parent pGPU when creating a new vGPU https://review.opendev.org/715489 | |
| 13:36:11 | openstackgerrit | Sylvain Bauza proposed openstack/nova master: WIP: Support different vGPU types per pGPU https://review.opendev.org/715490 | |
| 13:46:54 | openstackgerrit | Sylvain Bauza proposed openstack/nova master: Provide the parent pGPU when creating a new vGPU https://review.opendev.org/715489 | |
| 13:46:54 | openstackgerrit | Sylvain Bauza proposed openstack/nova master: WIP: Support different vGPU types per pGPU https://review.opendev.org/715490 | |
| 14:23:18 | kevinz | sean-k-mooney: Hi, when you have time, pls take a look at https://review.opendev.org/714311 and https://review.opendev.org/712607, already got +2 before :-) | |
| 14:29:39 | openstackgerrit | Stephen Finucane proposed openstack/nova master: libvirt: Always provide the size in bytes when calling virDomainBlockResize https://review.opendev.org/707590 | |
| 14:29:39 | openstackgerrit | Stephen Finucane proposed openstack/nova master: images: Remove Libvirt specific configurable use from qemu_img_info https://review.opendev.org/707591 | |
| 14:29:40 | openstackgerrit | Stephen Finucane proposed openstack/nova master: libvirt: Remove QEMU_VERSION_REQ_SHARED https://review.opendev.org/710239 | |
| 14:29:40 | openstackgerrit | Stephen Finucane proposed openstack/nova master: images: Make JSON the default output format of calls to qemu-img info https://review.opendev.org/711679 | |
| 14:30:02 | sean-k-mooney | kevinz: https://review.opendev.org/#/c/712607/16 looks good to me ill take alook at the other too | |
| 14:37:23 | sean-k-mooney | bauzas: are you free for 30 seconds to review two trivail patches for kevinz https://review.opendev.org/#/c/712607/16 and https://review.opendev.org/#/c/714311/3 both look good to me. | |
| 14:37:36 | bauzas | sean-k-mooney: lemme look | |
| 14:39:22 | kevinz | sean-k-mooney:bauzas: Thanks a lot | |
| 14:39:49 | bauzas | those are conflicting with my own changes, but i'll forget about this when reviewing :) | |
| 14:47:24 | sean-k-mooney | ah the allocation candiate changes | |
| 14:48:29 | sean-k-mooney | yes. i will need to take a look at those as i think i will need to use them if i want to support migration/resize with cyborg | |
| 14:50:33 | sean-k-mooney | dansmith gibi are either of ye around? | |
| 14:50:44 | gibi | I'm here | |
| 14:51:25 | sean-k-mooney | cool am for the cyborg evacuate support https://review.opendev.org/#/c/715326/ should i do the arq binding on the conductor like we do for spawn or on the compute node | |
| 14:51:45 | sean-k-mooney | i started doing it on the compute node but i think i should move it to the conductor | |
| 14:52:10 | sean-k-mooney | so that cyborg can start to program the device asyc of nova | |
| 14:52:19 | sean-k-mooney | gibi: any thoughts? | |
| 14:54:03 | dansmith | sean-k-mooney: condutor for sure unless there's some really compelling reason otherwise | |
| 14:55:24 | sean-k-mooney | dansmith: cool only reason not to is i start doing it the other way first which is easily solved via copy paste :) | |
| 14:55:40 | gibi | sean-k-mooney: yeah, qos stuff is also done there | |
| 14:55:51 | sean-k-mooney | dansmith: should i alther the rpc to pass the arqs | |
| 14:55:53 | gibi | I mean in the conducotr for evac | |
| 14:56:21 | dansmith | sean-k-mooney: we did for spawn because it's so much more efficient | |
| 14:56:46 | dansmith | sean-k-mooney: the early stuff landed without it because it was less change, but that's the only reason | |
| 14:57:24 | sean-k-mooney | so you would be in favor of just doing it for evac too | |
| 14:59:34 | sean-k-mooney | ok ill rework the change to bump the rpc and move the recreation of the arq and bindign to the conductor and do the same wait on the compute we do for spawn | |