Earlier  
Posted Nick Remark
#openstack-nova - 2019-03-20
18:59:42 sean-k-mooney mriedem: ya i was going to check the release notes for libvirt 1.2.17
18:59:55 openstack bugzilla.redhat.com bug 1203032 in libvirt "Enhancing block/disk migration in libvirt" [Low,Closed: nextrelease] - Assigned to pboldin
18:59:55 mriedem the commit message referenced https://bugzilla.redhat.com/show_bug.cgi?id=1203032
18:59:56 _pewp_ [ 1203032 – Enhancing block/disk migration in libvirt ] - bugzilla.redhat.com
19:00:22 eandersson sean-k-mooney, I see
19:00:42 eandersson I think this changed somehow during the upgrade which is why it's weird
19:00:48 eandersson I mean we went from.. Mitaka to Rocky :D
19:01:21 eandersson from my perspective it looks like base images are taken into account now, and weren't before.
19:01:30 sean-k-mooney eandersson: do you set allocation raitios in host aggreates?
19:01:42 eandersson nothing overcommited on these flavors
19:01:54 eandersson I tried adding disk_all = 1 but did nothing
19:04:05 _pewp_ [ OpenStack Docs: Configuration Options ] - docs.openstack.org
19:04:05 sean-k-mooney by disk_all do you mean https://docs.openstack.org/nova/rocky/configuration/config.html#DEFAULT.disk_allocation_ratio
19:04:18 eandersson sorry yea disk_allocation_ratio
19:04:38 eandersson we don't have it set in the aggregate for non-overprovisioned aggergates
19:05:33 sean-k-mooney eandersson: in ocata? i think we broke setting allocation ratios via host aggregates when we started using placment for cpu ram and disk
19:06:06 sean-k-mooney there is a mail from melwitt about butthis is proably not the issue.
19:06:39 sean-k-mooney http://lists.openstack.org/pipermail/openstack-dev/2018-January/126283.html
19:07:06 sean-k-mooney eandersson: in upgading form mitaka to rocky you went form that working to not
19:07:22 eandersson So I would expect that to break for instances that are over-commited
19:07:29 eandersson not for instances that are not overcommited right?
19:07:53 sean-k-mooney but also you nolonger need the AggregateCoreFilter,AggregateRamFilter,AggregateDiskFilter or the core/ram/disk filters
19:08:05 eandersson I think that would fix it for us
19:08:47 eandersson because placement does that now?
19:09:03 sean-k-mooney eandersson: yes this is handeled by placement now
19:09:21 sean-k-mooney or rather form ocata on
19:09:31 _pewp_ [ Gerrit Code Review ] - review.openstack.org
19:09:31 openstackgerrit Artom Lifshitz proposed openstack/nova master: [WIP] Revert resize: wait for external events in compute manager https://review.openstack.org/644881
19:09:50 artom _pewp_, what are you?
19:09:53 sean-k-mooney so you can safely disable the disk filter that is preventing the vm form working
19:10:14 sean-k-mooney artom: i think a bot of some kind
19:10:26 artom sean-k-mooney, heh, I got that part :)
19:10:31 eandersson Yea - I enabled debug
19:10:42 eandersson I gotta go, but will be back later. More meeting
19:13:24 sean-k-mooney artom: i think its https://github.com/hemna/sopel-tableflip
19:20:39 melwitt lbragstad, mriedem: well, the parts I thought were relevant to operators/users (and thus forum) are 1) are operators on board with the idea of fixing non-admin policy handling via scope types as a carrot to get operators to adopt different token scopes in their use of nova and 2) if the unified limit migration details in the spec sound OK to operators. 2) is maybe not something operators care too much about
19:23:13 _pewp_ [ Gerrit Code Review ] - review.openstack.org
19:23:13 efried dansmith: Sorry if I'm being a dolt here, but where does the request filter get required=COMPUTE_IMAGE_TYPE_FOO from originally? Is it image_meta? Is it translated from the image type somehow? This is the part I'm still missing in https://review.openstack.org/#/c/644625/
19:23:59 efried You've mentioned how the compute driver adds COMPUTE_IMAGE_TYPE_FOO into its capabilities
19:24:12 dansmith efried: I haven't worked it all out all the plumbing yet, but we know what image they asked for and we can know what format it's in
19:25:07 efried Okay. So like if there's an image.type field and we can do "COMPUTE_IMAGE_TYPE_SUPPORT_%s" % image.type kind of thing?
19:25:14 dansmith right
19:25:24 dansmith it's called format I think, but yeah
19:25:38 efried image.type.upper().replace("things that aren't legal", "_")
19:25:52 efried os_traits.normalize something
19:25:56 efried okay, cool. Thanks.
19:26:02 dansmith efried: | disk_format | qcow2 |
19:26:49 sean-k-mooney am so dumb question before i read your sepc. could we just tag the compute nodes with traits for the image typs they support and then read the image type form glance and add the trait to the placment request
19:26:58 dansmith efried: https://docs.openstack.org/glance/rocky/user/formats.html
19:26:59 _pewp_ [ OpenStack Docs: Disk and Container Formats ] - docs.openstack.org
19:27:06 dansmith efried: pretty constrained list there
19:27:15 sean-k-mooney im thinking of suppoted image types like compute capablites
19:27:18 efried dansmith: cool. Standard traits to be proposed I assume.
19:27:25 efried sean-k-mooney: That's what he's doing, yes.
19:27:33 dansmith efried: yep
19:27:41 sean-k-mooney efried: ok makes sense
19:28:09 dansmith efried: first item in work item list is "add to os-traits"
19:28:13 efried os_traits.normalize_name is only for custom traits. prolly going to want a map.
19:28:35 efried ++
19:28:56 sean-k-mooney efried: i was going to say we have some code in nova to normalise traits already but maybe we never merged that..
19:29:14 efried sean-k-mooney: Yeah, we merged it and then moved it to os.traits, but it's only for custom
19:29:53 sean-k-mooney i was thinking of the traits nomalisation code for the feature based nic schduling that we never finished.
19:29:56 efried It's probably okay to hardcode a map for this. Adding a new supported image type, that doesn't happen frequently, does it?
19:30:14 dansmith efried: correct
19:30:22 sean-k-mooney efried: not really no
19:30:35 efried if a new image type is added, does it require changes in nova anyway?
19:30:38 sean-k-mooney it happens about as often as adding a new hypervior
19:31:01 sean-k-mooney am not nessisarly
19:31:04 efried cause if not, we may want to consider making that translation generic rather than hardcoded
19:31:09 _pewp_ [ OpenStack Etherpad ] - etherpad.openstack.org
19:31:09 mriedem https://etherpad.openstack.org/p/nova-ptg-train
19:31:16 mriedem "New virt driver for rsd:"
19:31:19 sean-k-mooney we just pull the image form glance and give it to the hyperviro to use
19:31:21 dansmith efried: likely would.. libvirt can read a lot of stuff, but most of what is covered there
19:31:24 mriedem come to butthead https://github.com/openstack/rsd-virt-for-nova
19:31:31 dansmith efried: and if it wasn't just a "use this and load it" then it'd break us without nova support
19:31:53 efried anyway, I didn't want to get into all the impl details; I just wanted to understand where that bit was coming from.
19:32:03 dansmith efried: like, if it wasn't just a raw file we're copying around, libvirt would need tooling to deal with those files, even if it can read them
19:32:19 efried okay.
19:32:33 sean-k-mooney dansmith: we have a few other cases that are very similar
19:33:22 sean-k-mooney for example im thinkg of suport vif_models or gpu_models(not gpu/vgpu passthough but gpu emulation)
19:33:30 efried mriedem: I saw that, noticed they didn't have a whole lot of code in their virt driver
19:34:30 sean-k-mooney i was going to write a spec to enable virt-gpu support by extendign the enum we alredy have but do you think its woth doing somethign similar and reporting supprot vif_models or gpu_models as traits?
19:35:35 mriedem efried: i'm assuming you'll get all sorts of awesome internal pressure for that rsd driver now
19:36:16 sean-k-mooney mriedem: i have reviewd the code. its more limited then it used to be/could be
19:36:58 sean-k-mooney it currently relyi on create a 1:1 mapping between compute node and flavor
19:38:17 sean-k-mooney it previously used to to create 1 hypervior per compute node like the ironic on did so nova could act as a scudler. so while im happy to see the direver it has good way to go befoer it would be ready to move intree or be used in production
19:39:25 sean-k-mooney efried: have you looked at the rsd virt dirver at all
19:39:45 efried I glanced at what was there in github and it appeared to be empty. Beyond that, no.
19:40:20 efried Oh, I've been confusing RSD and RDT in my head. (can't imagine why)
19:40:30 efried so yeah, I've never heard of RSD, internally or otherwise.
19:40:48 sean-k-mooney efried: its ok you will :P
19:41:09 sean-k-mooney efried: https://github.com/openstack/rsd-virt-for-nova
19:41:52 efried ohh, there it is.
19:42:38 sean-k-mooney efried: ya it used to be a lot more powerful but things changed.
19:43:40 sean-k-mooney that said it still allows you to boot node with out ironic useig rsd via nova so thats kind of cool
19:46:17 sean-k-mooney but its currently missusing flavor in a way that will not scale. if you have 1000 RSD nodes it will create 1000 flavors...
19:49:12 cdent sounds like something I could use at home
19:53:00 sean-k-mooney cdent: the original plan was to create one hypervior per rsd node and 1 flavor per uniqute type of rsd node discovers whcih requried on allcoate of CUSTOM_RSD_NODE_TYPE_WHATEVER
19:53:36 sean-k-mooney so if you only had two types of rsd system you would get only 2 flavor and to customer resouce classes

Earlier   Later