| Posted | Nick | Remark | |
|---|---|---|---|
| #openstack-nova - 2019-09-25 | |||
| 15:56:46 | mlavalle | yeah same thing.... | |
| 15:57:44 | efried | mlavalle: I'm toying with ideas for ways to dramatically constrain the scope of what we commit to in ussuri, on the theory that stakeholders (such as your employer, whoever that may be) would rather get a hard "no" up front than have a bp/spec approved and have NO IDEA whether it's going to land, based on our track record of landing a random 60-ish percent of the things we approve, with no relation to the supposed prioritization | |
| 15:58:15 | mriedem | mlavalle works for verizon media with jroll and penick and the gang | |
| 15:58:46 | mlavalle | LOL, I didn't know we are the gang | |
| 15:58:47 | mriedem | the routed networks stuff in that spec back in newton was a lot more grandiose with refactoring all of the port creation stuff in conductor - a lot of stuff johnthetubaguy worked on while at OSIC | |
| 15:59:09 | mriedem | what functionally landed on the nova side was actually a really small change to allow deferred fix ip allocation | |
| 15:59:27 | mriedem | for which https://review.opendev.org/#/c/641670/ is a follow up | |
| 15:59:44 | mriedem | the actual routed networks thing in nova is a lot more complicated, | |
| 15:59:53 | efried | Beyond that, I have no objection to this being a thing in general. But I haven't yet figured out a reasonable way to propose and effect such a constrained scope. | |
| 16:00:33 | mriedem | with the routed networks stuff, i'd say showing it working in a multinode ci job would be table stakes since there is a lot of orchestration and setup involved between nova, neutron and placement | |
| 16:00:48 | mriedem | but lots of people have wanted that feature for a long time (godaddy also) | |
| 16:01:02 | mriedem | so that you can scale out neutron, | |
| 16:01:06 | mriedem | efried: think of it like cells v2 for neutron | |
| 16:01:47 | mriedem | constrain a network to a host aggregate and when creating a server on that network, make nova smart enough to only ask the scheduler for hosts from that network | |
| 16:02:00 | mriedem | rather than randomly pick some host in another aggregate that can't talk to that network | |
| 16:02:29 | mriedem | i wouldn't be surprised if people do that with customer filters and scheduler hints today | |
| 16:02:44 | mlavalle | Neutron already creates the data in placement. what we need now is for Nova to use that data | |
| 16:02:57 | mlavalle | during the scheduling process | |
| 16:03:00 | mriedem | yup, that's what my patch started | |
| 16:03:16 | mriedem | link the requested network(s) to a resource provider aggregate and pre-filter the placement results using that | |
| 16:03:29 | efried | Above is explaining "what" and "why", which is great. Right now though I'm only focused on "how much". The "why" will become relevant when arguing whether this thing should bump some other thing out of scope for ussuri. | |
| 16:05:12 | mriedem | if verizon can somehow say they'll use vpmems would that sweeten the deal? | |
| 16:05:48 | mriedem | i kid | |
| 16:06:00 | mriedem | cyborg is probably the biggest ticket thing that should be pushed forward in ussuri | |
| 16:06:07 | mriedem | given the years of talk | |
| 16:06:29 | mriedem | granted, routed networks was from newton but who's keeping track | |
| 16:09:35 | mlavalle | All I can add to this is that we really, really need this | |
| 16:10:26 | mlavalle | how can we get this topic in the "how much" consideration? | |
| 16:10:35 | cdent | people | |
| 16:11:00 | mlavalle | well, I already said we would push it forward | |
| 16:11:12 | mlavalle | that's how the conversation started | |
| 16:12:00 | mriedem | mlavalle: how hard would it be to get a 2-node CI job setup which sets up a host aggregate for one node and a host aggregate for another and a separate network for each aggregate so the neutron+placement stuff happens? | |
| 16:12:32 | cdent | mlavalle: sorry I was joining in late and didn't mean to sound snarky (if that's the way it sounded) | |
| 16:12:37 | mriedem | i think work could be done toward the eventual goal that needs to happen anyway | |
| 16:13:20 | mlavalle | cdent: oh, no I didn't take it as "snarky".... I was just catching up ;-) | |
| 16:13:26 | mriedem | maybe it's just a matter of a tempest test against a 2 node job? tempest would create the per-node network and aggregate right? | |
| 16:13:42 | mriedem | the neutron would wire up the placement resource provider aggregate? | |
| 16:13:46 | mriedem | *then | |
| 16:13:50 | mlavalle | correct | |
| 16:14:07 | mriedem | after that it's just a matter of writing a tempest test that makes sure the server gets created on the correct host | |
| 16:14:12 | efried | mlavalle: honestly, our poor track record has more to do with reviewer resource than developer resource | |
| 16:14:16 | efried | bbiab | |
| 16:14:33 | mriedem | having functioning CI goes a long way in reviewer confidence | |
| 16:14:43 | mriedem | numa live migration wouldn't have made it this cycle (again) without that | |
| 16:15:01 | mlavalle | mriedem: this is a great point | |
| 16:15:13 | mlavalle | we can start there | |
| 16:15:37 | mriedem | a deterministic tempest test for routed networks would be, i think, trying to create a server with a network in aggregate 1 but requesting a host in aggregate 2 and seeing it blow up during scheduling | |
| 16:15:47 | mriedem | the test would assert the server fails due to NoValidHost | |
| 16:15:56 | mlavalle | yeap | |
| 16:16:31 | mriedem | with 2.74 in train that's pretty easy https://docs.openstack.org/nova/latest/reference/api-microversion-history.html#id66 | |
| 16:17:26 | mriedem | so you (or whoever) could get started with the tempest test which would actually assert that the server gets created on the wrong host for the network (port binding might actually fail i'd guess) until nova supports routed networks during scheduling | |
| 16:17:30 | mlavalle | yeah, thanks for pointing that out | |
| 16:18:23 | mriedem | david bingham on https://review.opendev.org/#/c/656885/ is from godaddy and he and another guy were at the last ptg asking about this as well - it was the only thing they talked about while at the ptg | |
| 16:18:26 | mriedem | coincidentally | |
| 16:18:37 | mlavalle | efried_rollin: developer resource is a start, isn't it? | |
| 16:18:37 | mriedem | so, maybe verizon and godaddy SIG UP on this | |
| 16:18:45 | mriedem | oh wait, | |
| 16:18:47 | mriedem | not SIG | |
| 16:18:49 | mriedem | POP UP TEAM UP! | |
| 16:19:06 | mriedem | cdent: as TC emeritus you can correct me if i'm wrong | |
| 16:19:32 | cdent | for routed networks pop up team would be the team-type of the the day | |
| 16:19:40 | mriedem | du jour?! | |
| 16:19:46 | cdent | so sorry, yes | |
| 16:20:01 | cdent | i used up all my french recently | |
| 16:20:12 | mriedem | TC santioned collaborama du jour | |
| 16:20:16 | mriedem | is what i'm going to go with | |
| 16:20:31 | cdent | i believe pop up came about in response to encrypted volume efforts failing to get traction | |
| 16:20:44 | cdent | I fail to get traction understanding why a different name helps | |
| 16:20:45 | mriedem | s/volume/image/ but yeah | |
| 16:20:53 | mriedem | because SIGs were too formal dude | |
| 16:22:19 | cdent | I have really mixed feelings on all this stuff. | |
| 16:22:30 | cdent | On the one hand it is wrong that nova gatekeeps so much | |
| 16:22:39 | cdent | but on the other hand, if nova doesn't gate keep things go sideways | |
| 16:24:22 | mriedem | mlavalle: i added you to https://review.opendev.org/#/c/656885/3 and left a comment there with link to this irc conversation if that helps | |
| 16:24:54 | mlavalle | mriedem: it definitely does. Thanks you very much! | |
| 16:29:14 | gregwork | is it terribly difficult to to fiddle with libvirt <features> and <cpu> in nova? | |
| 16:29:30 | gregwork | i need to figure out how to set these to fool windows into not being terrible at everything because it detects its in kvm | |
| 16:29:31 | gregwork | https://pastebin.com/jw8Duq9S | |
| 16:29:51 | gregwork | it actively checks to see if its in kvm and turns off stuff | |
| 16:30:02 | gregwork | thats how i work around it using regular libvirt | |
| 16:30:47 | mriedem | gregwork: gpu? | |
| 16:31:04 | mriedem | gregwork: see https://review.opendev.org/#/c/579897/ | |
| 16:31:42 | gregwork | so not an nvidia badness, i want to enable HyperV server role in my guest so I can run the cloud-base.it image generation git project | |
| 16:31:49 | gregwork | and it wont let you do that becuase HyperV detects KVM | |
| 16:31:53 | gregwork | and says "newp!" | |
| 16:32:18 | gregwork | this thing: https://cloudbase.it/windows-server-2016-openstack-images/ | |
| 16:32:46 | mriedem | well i think you're looking for the img_hide_hypervisor_id image property or the hide_hypervisor_id flavor extra spe | |
| 16:32:48 | mriedem | *spec | |
| 16:33:03 | mriedem | the image property isn't in queens but the flavor extra spec might be | |
| 16:33:19 | mriedem | nope https://blueprints.launchpad.net/nova/+spec/hide-hypervisor-id-flavor-extra-spec was rocky | |
| 16:33:39 | gregwork | and there is no way to pass libvirt domain customizations | |
| 16:33:52 | gregwork | even kludgy ones :) | |
| 16:33:57 | gregwork | or even qemu execution lines | |
| 16:34:10 | mriedem | not through the compute api no | |
| 16:34:29 | mriedem | that's not really...cloud | |
| 16:35:29 | artom | Don't we have image props or something to make Windows happy? | |
| 16:35:54 | mriedem | artom: read scrollback | |
| 16:36:03 | gregwork | ever since they started planning to go to core based licensing instead of socket they have implemented some draconian things to prevent the potential for nested virtualization | |
| 16:36:11 | openstackgerrit | Merged openstack/nova master: docs: Scrub available quotas https://review.opendev.org/670125 | |
| 16:36:31 | gregwork | especially since you can do performant nested virt with kvm on intel xeon e3-v4 processors using vmcs shadowing and device passthrough of network/storage | |