| Posted | Nick | Remark | |
|---|---|---|---|
| #openstack-nova - 2019-09-24 | |||
| 14:26:10 | sean-k-mooney | it defualt to stict in numatune because if you said you want to tune the numa memory it makes sense to default to strict | |
| 14:26:24 | sean-k-mooney | given by default numa tune is not used and no tuning is used | |
| 14:26:50 | sean-k-mooney | so you are already opting into affinty by generating the numatune elements | |
| 14:26:58 | KeithMnemonic | mriedem yes there is an issue? or yes i can try and backport it ? | |
| 14:27:15 | mriedem | KeithMnemonic: you basically asked if it's ok to backport something from train that removed something that was deprecated in stein, and has nothing to do with rocky | |
| 14:27:32 | mriedem | i know suse is super in love with rocky | |
| 14:27:35 | mriedem | but... | |
| 14:28:01 | kashyap | sean-k-mooney: I'm NUMA-unaware, so afraid, I still can't parse _why_ libvirt defaults to 'strict' | |
| 14:28:17 | kashyap | (Or weakly-aware) | |
| 14:28:33 | KeithMnemonic | oh i see, filters were deprecated in stein | |
| 14:28:39 | kashyap | sean-k-mooney: Please rephrase, if you can... | |
| 14:28:42 | KeithMnemonic | https://review.opendev.org/#/c/596502/ | |
| 14:28:57 | mriedem | correct | |
| 14:28:59 | sean-k-mooney | kashyap: by default you do not specify numatune elements. so by defualt libvirt does not enforce strict affinity | |
| 14:29:01 | KeithMnemonic | so at a minimum it would make sense to put this to stein as well if someone needed it? | |
| 14:29:09 | kashyap | sean-k-mooney: Aah, like that! | |
| 14:29:37 | mriedem | KeithMnemonic: put what? https://review.opendev.org/#/c/551026/ in stein? | |
| 14:29:40 | sean-k-mooney | kashyap: when you therefor add the numatune element it makes sense for it to default to strcit since you ar opting out of the default behavior of no tuneing | |
| 14:30:07 | sean-k-mooney | kashyap: does that make sense? | |
| 14:30:19 | mriedem | KeithMnemonic: https://review.opendev.org/#/c/551026/ depends on https://review.opendev.org/#/c/672065/ which you can't backport to stein | |
| 14:30:20 | kashyap | sean-k-mooney: Yes, that's clearer. Thank you | |
| 14:30:27 | KeithMnemonic | got it | |
| 14:30:28 | mriedem | you can't say "x is deprecated and also removed in the same release" | |
| 14:30:31 | mriedem | maybe suse can | |
| 14:30:32 | KeithMnemonic | thanks for cleating it up | |
| 14:30:36 | mriedem | but that's a pretty bad precedent | |
| 14:30:41 | KeithMnemonic | no we dont | |
| 14:31:21 | KeithMnemonic | just a customer complaining about the root issue https://bugs.launchpad.net/nova/+bug/1469179 and see what options there are | |
| 14:31:21 | openstack | Launchpad bug 1469179 in OpenStack Compute (nova) "instance.root_gb should be 0 for volume-backed instances" [Medium,Fix released] - Assigned to Dan Smith (danms) | |
| 14:32:03 | sean-k-mooney | KeithMnemonic: thy can change there flavor to have 0 root disk | |
| 14:32:44 | mriedem | KeithMnemonic: the main fix was in rocky https://review.opendev.org/#/c/580720/ | |
| 14:32:47 | KeithMnemonic | yup, that is the workaround they have in place | |
| 14:33:00 | mriedem | the DiskFilter should be disabled (since pike i think) | |
| 14:33:17 | sean-k-mooney | mriedem: yes it should | |
| 14:33:36 | KeithMnemonic | but some sales guy saw the launchpad bug and is asking why it was not fully fixed in rocky. and thanks to mriedem i can answer now ;-) | |
| 14:34:52 | mriedem | it is fixed in rocky | |
| 14:35:27 | sean-k-mooney | did we add the policy rule in rocky too? i think we did but its deisable by defualt? | |
| 14:35:54 | mriedem | yes https://docs.openstack.org/nova/rocky/configuration/policy.html | |
| 14:35:58 | mriedem | os_compute_api:servers:create:zero_disk_flavor | |
| 14:39:08 | sean-k-mooney | so the rocky fix is 1.) disable ram|cpu|disk filters 2.) create bfv 0 disk flavors, 3.) chage the os_compute_api:servers:create:zero_disk_flavor rule to rule:admin_api | |
| 14:40:15 | mriedem | it depends on what you care about, | |
| 14:40:48 | mriedem | the immediate issue is failing scheduling b/c of the DiskFilter or claim (by having DiskFilter enabled) and booting a volume-backed server - that is fixed with dan's fix above and disabling the DiskFilter since placement won't filter on DISK_GB | |
| 14:41:10 | mriedem | the resource tracker should also not be counting volume-backed server usage against the compute node (hypervisors API) since rocky | |
| 14:41:27 | mriedem | so i'm not sure what is *not* fixed in rocky | |
| 14:42:34 | sean-k-mooney | right. the only thing i can think of is that we dont default to safe regarding os_compute_api:servers:create:zero_disk_flavor | |
| 14:42:46 | sean-k-mooney | but that is for backwards comaptablity in rocky | |
| 14:56:01 | openstackgerrit | Merged openstack/os-resource-classes master: Update bug link in docs to point to storyboard https://review.opendev.org/681464 | |
| 15:02:09 | openstackgerrit | Artom Lifshitz proposed openstack/nova master: Move pre-3.44 Cinder post live migration test to test_compute_mgr https://review.opendev.org/683597 | |
| 15:02:24 | openstackgerrit | Merged openstack/os-resource-classes master: Build pdf docs https://review.opendev.org/681463 | |
| 15:13:38 | openstackgerrit | Kashyap Chamarthy proposed openstack/nova master: libvirt: Use the "memory" mode 'preferred' for a guest NUMA node https://review.opendev.org/684375 | |
| 15:14:13 | kashyap | sean-k-mooney: ^ In your "copious spare time" | |
| 15:14:34 | kashyap | sean-k-mooney: Shouldn't we consider migration here: if an instance from a Compute node with 'strict' is being migrated to a host with 'preferred' - possible explosions? | |
| 15:15:01 | kashyap | sean-k-mooney: Err, that main commit summary is nonsense, let me fix that | |
| 15:15:06 | sean-k-mooney | yes this would need to be taken into account | |
| 15:15:38 | sean-k-mooney | but if we hardcore preferred then it not an issue excapt on upgrde | |
| 15:15:49 | sean-k-mooney | it would be fixed by a hard reboot | |
| 15:16:54 | sean-k-mooney | i would not nessisarly require this to be handeled in live migration given that numa live migration was not really correct until train | |
| 15:19:18 | openstackgerrit | Kashyap Chamarthy proposed openstack/nova master: libvirt: Use the `preferred` memory allocation mode for NUMA https://review.opendev.org/684375 | |
| 15:19:31 | aspiers | kashyap: "pfalsh" typo in https://libvirt.org/formatdomaincaps.html#elementsOSBIOS | |
| 15:19:47 | kashyap | aspiers: Ah, you saw the libvirt patch I Cced you on? | |
| 15:19:50 | aspiers | yes | |
| 15:20:04 | kashyap | aspiers: Will fix; it's merged already :D | |
| 15:20:11 | kashyap | Thanks for spotting; feel free to reply there | |
| 15:20:25 | aspiers | kashyap: So I saw. Thanks - don't think I'm subscribed to that list | |
| 15:20:46 | kashyap | sean-k-mooney: Ah-ha; good point on "NUMA + live migration not really being a 'thing' until Train" | |
| 15:20:47 | aspiers | kashyap: While you're at it, you could add some helpful hyperlinks | |
| 15:21:13 | sean-k-mooney | kashyap: well it was a thing but it had a load of edgecases | |
| 15:21:15 | kashyap | aspiers: You don't need to be subscribed; just reply they moderators will (usually, quickly) take care | |
| 15:21:27 | sean-k-mooney | so i dont really considre this edgecase to be important | |
| 15:21:40 | sean-k-mooney | at least not relitive to the other ones that were not handeled | |
| 15:21:45 | kashyap | sean-k-mooney: So, an upgrade note is required, too, along with the 'fixes' note I added -- isn't it? | |
| 15:21:56 | aspiers | kashyap: I'd rather not get involved - too much other stuff to deal with :) | |
| 15:21:58 | kashyap | aspiers: Mind pointing out which ones? | |
| 15:22:03 | kashyap | aspiers: Heh, I hear ya | |
| 15:22:16 | aspiers | kashyap: Here are some suggested edits | |
| 15:22:18 | aspiers | "For the loader element, the following can occur:" | |
| 15:22:28 | aspiers | "the following elements can occur:" | |
| 15:22:36 | kashyap | (Oh, I wanted to fix that, too! Realized it too late) | |
| 15:22:43 | aspiers | but then you list "value" alongside "type" etc. which is incorrect | |
| 15:22:55 | aspiers | type/readonly/secure are values of the name attribute of <enum> | |
| 15:22:56 | sean-k-mooney | we could update the numa migration code in train to also update it to prefered | |
| 15:22:59 | aspiers | but you don't even mention <enum> | |
| 15:23:03 | sean-k-mooney | but i would do that as a seperate patch | |
| 15:23:44 | aspiers | and the confusion over the point of <enum> was the whole reason we started looking at this in the first place | |
| 15:24:15 | aspiers | kashyap: ^^^ | |
| 15:25:00 | kashyap | aspiers: Okay, my brain is completely exhausted for the moment; will make a note to get back to this and submit a patch | |
| 15:25:03 | kashyap | Thanks! | |
| 15:25:05 | aspiers | kashyap: "For cases like that, os element exposes ..." is ambiguous about whether it's <os> in domCaps or caps | |
| 15:25:12 | aspiers | which again was the crux of the confusion | |
| 15:25:28 | aspiers | These docs need to be crystal clear about when they are referring to elements of domCaps vs. caps | |
| 15:25:50 | aspiers | so that the parallel between the two is understood | |
| 15:26:33 | kashyap | aspiers: I didn't touch those bits; but you're right, they _do_ need "less careless" documentation. I'll get back to this tomm / this week | |
| 15:26:45 | aspiers | kashyap: thanks! one more pointer | |
| 15:26:49 | kashyap | Non-null? | |
| 15:26:53 | kashyap | :D | |
| 15:27:06 | aspiers | kashyap: I'm suggesting to hyperlink text like "the <loader/> element in the domain XML" | |
| 15:27:20 | aspiers | should link to https://libvirt.org/formatdomain.html#elementLoader | |
| 15:27:29 | aspiers | in both directions, ideally | |
| 15:27:39 | aspiers | kashyap: feel free to send me something to review | |
| 15:27:43 | aspiers | when you have a chance | |