Earlier  
Posted Nick Remark
#openstack-nova - 2018-01-16
17:59:45 openstackgerrit Eric Fried proposed openstack/nova master: WIP: Use update_provider_tree from resource tracker https://review.openstack.org/520246
17:59:45 openstackgerrit Eric Fried proposed openstack/nova master: Fix nits in update_provider_tree series https://review.openstack.org/531260
17:59:48 efried There, ffs. mgoddard Sorry dude
18:00:39 lyarwood mdbooth: sure, thanks again
18:00:47 melwitt jaypipes: scheduler question for you ... when the filter scheduler does the compute node prune based on placement before going through the scheduler filters, does it take into consideration overcommit values set by aggregate?
18:01:28 jaypipes melwitt: no
18:01:53 jaypipes melwitt: placement has no concept of "aggregate overcommit"
18:03:22 melwitt jaypipes: k, that's what I was thinking
18:05:02 melwitt jaypipes: we're seeing issues where all of the overcommit was set via aggregates, so placement is filtering out very many nodes that can actually serve requests given overcommit
18:05:31 melwitt as a workaround I was thinking, have to set overcommit per compute node in nova.conf to get around this
18:06:51 jaypipes melwitt: I'll be honest, I didn't know about setting overcommit via aggregate.
18:07:04 jaypipes melwitt: sounds like a half (quarter?)-baked feature
18:07:27 melwitt heh. well, it's the AggregateCoreFilter and friends
18:07:48 jaypipes melwitt: and as we've discussed before, aggregates and Ironic don't mix. at all...
18:08:06 jaypipes melwitt: or is this not an ironic thing?
18:08:13 melwitt this is not an ironic thing
18:08:24 jaypipes kk
18:08:47 melwitt since they're doing overcommit via aggregate, placement is returning only like 6 nodes when 20+ can fulfill given overcommit
18:08:47 jaypipes melwitt: still sounds like a half-baked "feature" to me...
18:09:20 melwitt meaning, you expect setting overcommit per compute node should be the fully baked way?
18:09:44 melwitt just trying to understand how they should proceed going forward
18:09:46 jaypipes melwitt: yes. it's the *only* way.
18:11:47 melwitt k. I think the concern there is it hurts usability of aggregates. setting them per aggregate (a few times) vs setting them per compute node (a lot of times)
18:11:48 cdent overcommit by aggregate. interesting. yet another thing I've never heard of before. I can see the appeal
18:12:19 melwitt since aggregate means "bucket of like compute nodes" I can see how one would want to set overcommit same per bucket
18:12:33 melwitt *since aggregate usually means
18:12:56 openstackgerrit Jay Pipes proposed openstack/nova master: add _has_provider_trees() utility function https://review.openstack.org/531474
18:12:56 openstackgerrit Jay Pipes proposed openstack/nova master: placement: _get_trees_matching_all() https://review.openstack.org/531512
18:12:57 openstackgerrit Jay Pipes proposed openstack/nova master: add tests for _get_trees_matching_all() with trait https://review.openstack.org/531899
18:12:57 openstackgerrit Jay Pipes proposed openstack/nova master: add test for scenario with sum of child resources https://review.openstack.org/534339
18:13:31 jaypipes melwitt: an aggregate in placement doesn't have any metadata associated with it at all. it's a pure grouping mechanism.
18:14:30 efried stephenfin Not sure which two comments were the ones holding you back. Let us know if we can clarify anything further. Thanks for the review!
18:14:31 melwitt jaypipes: yeah, I understand. just pointing out the high level effect on the usability in this case
18:14:40 jaypipes melwitt: the problem with the aggregate overcommit thing in Nova is what happens when the compute node's allocation ratio is != the aggregate's allocation ratio? also, what happens when a compute node is in multiple aggregates? which allocation ratio should be used? all these things are not issues in placement because we don't store metadata for aggregates.
18:15:11 melwitt jaypipes: good points
18:15:24 melwitt especially the multiple aggregates
18:16:38 jaypipes melwitt: and I'm not trying to be argumentative. just pointing out that there are definitely some things we won't be porting over to placement land...
18:17:15 melwitt jaypipes: understood. I'm not trying to be argumentative either. just looking to understand and be able to explain it to ppl :)
18:17:21 melwitt so thanks
18:18:55 jaypipes melwitt: when discussing a similar thing with operators (having to set traits on compute node resource provider records individually instead of relying on agg metadata) I had suggested a simple CLI tool that would essentially apply a particular trait to all providers associated with an aggregate. they seemed responsive to that idea.
18:26:14 efried jaypipes edleafe cdent gibi For all the hubbub, we may still want to implement an accessor for provider generation. Given the way I'm using it e.g. in https://review.openstack.org/#/c/532564/9/nova/scheduler/client/report.py@1046
18:26:19 melwitt jaypipes: k, I'll mention that to them. we'll be looking at how to change over from the Aggregate<resource> filters and configure the deployment to set it per compute node
18:27:09 melwitt s/it/overcommit/
18:27:56 edleafe efried: what's the reason you can't just use generation = self._provider_tree.generation?
18:28:21 efried edleafe ProviderTree doesn't have a generation. A single provider does.
18:28:43 efried edleafe Back to the thing where individual providers are hidden in the ProviderTree.
18:29:03 edleafe efried: ok, I thought there was something I was missing
18:29:26 efried ProviderTree.snapshot(rp_uuid).generation *works* - it's just wildly inefficient.
18:30:25 edleafe efried: are there other atts besides generation that you have a need to access?
18:31:14 openstackgerrit Lee Yarwood proposed openstack/nova master: rbd: flatten images when creating/unshelving an instance https://review.openstack.org/457886
18:31:43 efried edleafe Yes, the complex ones (aggs, traits, inventory) which we *must* copy out. Primitives, though... maybe parent_uuid? Eventually name?
18:32:16 efried edleafe More will shake out once drivers start implementing update_provider_tree.
18:33:11 edleafe efried: so are you thinking something like: ptree.getattr(rp_uuid, attname) ?
18:34:42 efried edleafe Initially I was just thinking about ptree.generation_for(rp_uuid), qua https://review.openstack.org/532922
18:35:16 edleafe efried: well, that's why I asked if you will need this for more than just generation
18:35:39 efried edleafe Surely. Even so, the idea of N accessors was discussed on Monday and jaypipes expressed a preference for doing it this way.
18:35:49 edleafe efried: it would suck to have ptree.name_for(rp_uuid), ptree.parent_uuid_for(rp_uuid), etc.
18:35:55 efried agreeeeed.
18:36:10 efried edleafe I think what I'm actually coming around to is that we shouldn't do the deep copy of the nested children in .snapshot() - at least not by default.
18:36:46 efried edleafe Then at least we're only throwing away one provider's worth of stuff we copied for no reason.
18:38:11 sean-k-mooney stephenfin: jaypipes i ran all the tempest smoke test + senario test + tox -e py35,py27,functional,pep8,docs locally on https://review.openstack.org/#/c/534307/2 here http://paste.openstack.org/show/645744/ all passed so just waiting on the gate to do the same.
18:38:15 edleafe or have snapshot() take an optional rp_uuid
18:38:32 edleafe and only return that provider w/o children
18:38:54 efried edleafe It already does take rp_uuid; that's how you tell it which provider you want to snapshot.
18:41:23 esberglu stephenfin: Asked around and it turns out that nova-networking isn't supported. You okay with leaving bit in for now and removing it in a follow up?
18:41:34 esberglu Trying to avoid respinning the series if possible
18:42:02 sean-k-mooney esberglu: nova-networking has not been removed yet. it will be in early rocky
18:42:28 esberglu sean-k-mooney: This is in regards to powervm virt driver support
18:42:52 esberglu It only supports neutron
18:43:27 sean-k-mooney esberglu: oh well if povervm never supported nova-net before or you have already dropped(or never supported) cells v1 then you dont need nova-networks support
18:44:02 efried esberglu fyi stephenfin is on UK time, so may be gone for the evening.
18:44:22 esberglu efried: ahh thanks for the info
18:45:28 efried esberglu And gibi is also euro. So you might as well respin the series and they should be able to nail 'em quick in the morning (like 4am for us).
18:46:53 sean-k-mooney efried: are you east coast or west?
18:47:12 efried sean-k-mooney Central (Texas)
18:47:27 sean-k-mooney ah ok
18:47:31 efried I think esberglu is either Central or East
18:47:47 esberglu central
18:48:26 openstackgerrit Eric Berglund proposed openstack/nova master: PowerVM Driver: vSCSI volume driver https://review.openstack.org/526094
18:49:10 sean-k-mooney ya unlike me stephen starts early and finishes early. i generally stay later but start later to give more us overlap in my day. also i hate have 2 7 oclocks in my day unless the frist on is 7 pm
18:50:10 efried sean-k-mooney Sounds like you coordinate way more than we do :)
18:51:14 efried As far as meetings go, I consider a successful week to be one with three IRC meetings and zero in-person or phone meetings.
18:52:28 efried (at least, zero in-person meetings that don't involve choking the persons involved)
18:52:52 openstackgerrit Claudiu Belu proposed openstack/nova master: tests: fixes mock autospec usage https://review.openstack.org/447505
18:53:18 sean-k-mooney efried: haha i would love that. lately i have not had much internal meeting but in the past i had. i like beeing on irc when most of the nova,neutron and kolla core team are about too so haveing a slightly later day helps
18:53:41 efried Mm. I believe cdent is of similar philosophy.
18:54:03 edleafe efried: I have 3 phone meetings *today*
18:54:28 cdent ... that's more side-effect than goal
18:58:37 openstackgerrit Eric Berglund proposed openstack/nova master: PowerVM driver: ovs vif https://review.openstack.org/422512
18:58:37 openstackgerrit Eric Berglund proposed openstack/nova master: PowerVM Driver: SEA https://review.openstack.org/523216
18:58:38 openstackgerrit Eric Berglund proposed openstack/nova master: PowerVM Driver: vSCSI volume driver https://review.openstack.org/526094
19:01:10 sean-k-mooney efried_nomnom: esberglu is there a reason you are not using and os-vif plugin intead of https://review.openstack.org/#/c/422512/40/nova/virt/powervm/vif.py
19:02:17 efried_nomnom sean-k-mooney The main reason is probably because os-vif didn't exist when we wrote this stuff, or even when we first started porting it in tree.
19:02:51 sean-k-mooney efried_nomnom: os-vif has been around for 3 maybe 4 cycles now
19:03:11 sean-k-mooney efried_nomnom: it would be nice to convert in rocky
19:03:23 efried_nomnom Duly noted. esberglu ^ for the to-do list?
19:03:59 efried_nomnom Also sean-k-mooney we wouldn't know how to do that offhand. But since you've volunteered to guide us.... :)
19:04:06 cfriesen I'm trying to figure out if a bug is in nova or libvirt/qemu. Do we have any functional/tempest tests of live migration with a config drive and an attached cinder volume?
19:04:45 Roamer` hmm, if mriedem is really on vacation, could I interest any of you in taking a quick look at https://review.openstack.org/140733/ - the StorPool libvirt volume attachment driver? :) The Cinder and os-brick support is in, and Nova's requirements.txt has a dependency on the new version of os-brick now.
19:05:31 Roamer` of course, "leave it till he gets back" might also be kind of reasonable... although it may be cutting it a bit too close to the January 25th deadline

Earlier   Later