Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2020-02-07
16:10:44 spatel I do have hw:cpu_policy=dedicated policy but its randomly mapping vCPU <--> pCPU
16:11:09 spatel sean-k-mooney: what do you suggest here?
16:11:43 spatel In short i want to make my VM fully NUMA map like bare metal has.
16:14:26 sean-k-mooney spatel: so it look like the erlang applciation is not numa aware
16:14:56 spatel sean-k-mooney: but if same erlang application i am running on bare metal it does work and result is good
16:14:59 sean-k-mooney meaning it is internally not numa affinitising its memory allocations
16:16:11 sean-k-mooney spatel: wehn you set hw:cpu_policy deicated it does a 1:1 mapping between cores on a virtual numa node to the a host numa node
16:16:24 spatel i thought if somehow openstack pin vcpu0 <--> pcpu0 , vcpu1 <--> pcpu1 so on... that would be great
16:16:30 sean-k-mooney so its not random it is numa aware
16:16:42 sean-k-mooney spatel: no it does not by design
16:17:05 spatel sean-k-mooney: look at this current mapping - http://paste.openstack.org/show/789225/
16:17:11 sean-k-mooney but it will map all fo the cores in one geust numa node to the a singel numa node on the host
16:17:45 sean-k-mooney thats takeing a long time to load for some reason
16:18:47 sean-k-mooney paste.openstack.org is down for me it seams
16:19:21 lyarwood https://review.opendev.org/#/c/701430/ - trying to get my head around the required api samples for a microversion that changes behaviour but nothing in the request or response, does anyone know what I would need to add here?
16:19:30 sean-k-mooney ok it finally loaded
16:20:10 sean-k-mooney spatel: how many cores to you have on the host
16:20:54 sean-k-mooney im assuming the are 16 core cpus and you have hypter treading enabled?
16:22:22 sean-k-mooney spatel: can you provide the host capablities xml and the full guest xml
16:22:42 spatel Give me few min.. i am on phone..
16:24:14 sean-k-mooney spatel: ok that xml fragment does not really look correct.
16:36:46 efried dansmith: I can push https://review.opendev.org/#/c/706331/ if you don't plan to reword that reno.
16:37:03 dansmith efried: it makes sense to me and nobody else has suggestions right?
16:38:21 efried dansmith: let me take a real swing at a reword. If I can't come up with something I like, I'll push it.
16:41:10 efried dansmith: do db migrations happen automatically as part of the upgrade process, or do users have to trigger them manually? (Or does it depend on the distro?)
16:41:24 dansmith efried: depends a lot
16:46:01 efried dansmith: done, see what you think.
16:46:29 efried dansmith: for me, the distinction of "apply the fix" versus "install a version that includes the fix" is where the confusion lies.
16:47:11 dansmith I totes don't get the substantive difference between your sentences and mine, but as such I'll be glad to change it.. just a sec
16:48:18 efried mriedem: if you're around -- does my suggestion improve anything IYO?
16:48:44 openstackgerrit Dan Smith proposed openstack/nova master: Fix instance.hidden migration and querying https://review.opendev.org/706331
16:48:46 efried It becomes hard to see anymore having discussed it to death
16:49:11 efried dansmith: coulda added the link while you were at it :P
16:49:19 dansmith guh, will fix
16:49:27 dansmith working on the patch above this so I wanted to get back to that
16:50:17 openstackgerrit Dan Smith proposed openstack/nova master: Fix instance.hidden migration and querying https://review.opendev.org/706331
16:50:26 dansmith efried: please check before I roll my tree forward again :)
16:51:42 efried dansmith: I'll build locally to make sure. You may need an extra newline between the paragraph and the link def. Also would be nice to link same from the `fixes` note (you don't need / can't include a second target def, they'll both use the same one)
16:52:20 mriedem you do
16:52:23 mriedem it blows up without that
16:52:28 openstackgerrit Dan Smith proposed openstack/nova master: Fix instance.hidden migration and querying https://review.opendev.org/706331
16:52:32 mriedem http://rst.ninjs.org/
16:52:33 dansmith like that?
16:53:07 efried dansmith: add trailing newline on the fixes one
16:53:15 efried s/newline/undescore/
16:53:21 dansmith "for all you people who have now corrupted their database, I hope you appreciate that we waited quite a while to make sure the bug link was perfect"
16:53:39 efried Yeah, these five minutes are really going to make the difference.
16:53:47 dansmith a trailing newline where
16:53:50 dansmith ?
16:53:50 efried If you want to go back to real work, I'll doodle with the formatting and fast approve it.
16:54:00 efried not a newline, the second bug link is backticked but not underscored.
16:54:02 dansmith no, it's very important to me that I get this right
16:54:10 mriedem .. _bug 1862205: https://launchpad.net/bugs/1862205
16:54:10 openstack Launchpad bug 1862205 in OpenStack Compute (nova) "Instances not visible when hidden=NULL" [Critical,In progress] - Assigned to Dan Smith (danms)
16:54:10 dansmith oh dammit
16:54:10 mriedem you need the ..
16:55:00 openstackgerrit Dan Smith proposed openstack/nova master: Fix instance.hidden migration and querying https://review.opendev.org/706331
16:55:20 efried this would be a great time for that fast-reno tool I've been meaning to write. The releasenotes target takes forEVER to build.
16:55:50 efried there, yeah, ps9 looks right to me.
16:56:14 dansmith ttfl
16:56:40 stephenfin efried: We need to stop 'rm -rf' ing the built docs
16:56:52 efried is that what does it?
16:57:08 efried I'll try that
16:57:26 mriedem efried: dansmith: so PS9 might work, but i've seen reno complain about links in weird ways between sections, so we'll see if the link defined in the first section is honored in the second
16:57:45 dansmith mriedem: efried said it was only needed once
16:57:58 mriedem i guess we'll know in a couple of hours :)
16:58:04 efried should be the case, but it depends how reno does its build.
16:58:07 stephenfin Sphinx builds doctrees and reuses those when possible. Unfortunately it occasionally breaks if you e.g. remove a document that previously existed (like by changing to an older branch)
16:58:13 stephenfin probably less on an issue for renos
17:01:19 efried stephenfin: In the specs repo I wrote a tool that would only build one directory at a time. It was a fiddly thing to figure out exactly how to word the sphinx-build command. In this case... if I could restrict it to building just $release.rst it would probably be equivalent.
17:09:42 efried dansmith, mriedem: local build is green, the link works in both spots as expected. Fast approving.
17:10:11 efried ...mriedem do you agree with the new wording? Or at least it's not worse?
17:12:06 efried dansmith: we shouldn't have to backport the fixture poison beyond Train, right? because we wouldn't be backporting any future fix that hasn't gone through that fixture in Train+
17:12:25 dansmith efried: we wouldn't backport that anyway
17:14:01 openstackgerrit Dan Smith proposed openstack/nova stable/train: Fix instance.hidden migration and querying https://review.opendev.org/706582
17:14:15 dansmith backport ^
17:15:50 efried +1
17:17:40 efried stephenfin: fwiw commenting out the 'rm' doesn't seem to be speeding anything up. It's still scanning the world, which seems to be the part that takes forever.
17:17:52 stephenfin darn
17:17:59 stephenfin I was thinking all that would be cached
17:18:24 efried stephenfin: considering a target that patches index.rst and (re)moves $not-this-release.rst before building.
17:18:30 gibi efried, dansmith: do we have re-schedule handling implemented in the cyborg series?
17:18:42 efried ...and then of course it would have to undo that after
17:18:58 dansmith gibi: unsure, I'm kinda depending on sean-k-mooney to poke at some of those things
17:19:10 efried gibi: How is that handled for bw? I would have thought it would be the same way.
17:19:33 stephenfin efried: Have we closed any more branches yet? https://github.com/openstack/nova/commit/857b5003ccc0b37f4642ed77d9f0d08f9ee28dfb
17:19:33 dansmith yeah, I think it should be the same for ports even, but actually showing that it works is different of course
17:19:55 stephenfin I guess with EM we no longer do that
17:19:58 efried stephenfin: yeah, was thinking about that
17:20:13 efried do we truly never EOL anything at this point?
17:20:21 gibi the arqs needs to be unbound from the failed dest host and bound to the alternate host. the placement handling shoudl work out of the box, but we at least needs some test coverage to prove
17:21:28 stephenfin Maybe we could just live with the lack of updated release notes for those branches?
17:21:38 dansmith gibi: yup
17:22:24 sean-k-mooney dansmith: how would you like me to test reschdule? i can set up a multi node deployment i guess and maybe kill libvirt?
17:22:31 efried gibi: agree. At a glance, it's weird that delete_arqs_for_instance is first introduced here https://review.opendev.org/#/c/673735/
17:24:06 dansmith sean-k-mooney: I dunno.. if that'd work then maybe that's easy, or just throw a raise into the code on one node?
17:24:43 gibi sean-k-mooney: I think if you kill libvirt then the compute service will go down and the scheduler will not select it as a target
17:24:52 efried stephenfin: I removed ocata-train rsts and from index, it's still taking bloody forever. It's the scanning that's slow.
17:25:28 stephenfin efried: dhellmann told me there was a way to configure earliest-version to previous each one going back forever
17:25:29 sean-k-mooney if i do a multinode deployment then maybe i could kill ovs or the cyborg agent instead

Earlier   Later