| Posted | Nick | Remark | |
|---|---|---|---|
| #openstack-nova - 2017-08-29 | |||
| 14:54:13 | jaypipes | edleafe: which is soon :) | |
| 14:54:17 | edleafe | jaypipes: thx | |
| 14:54:21 | stephenfin | sean-k-mooney: I think efried and claudiub|2 opened some blueprints | |
| 14:54:37 | stephenfin | https://blueprints.launchpad.net/nova/+spec/allow-pci-alias-choose-subset-of-devices and https://blueprints.launchpad.net/nova/+spec/devices-as-resources | |
| 14:54:59 | sean-k-mooney | edleafe: ovo for scheuler info from placement make sense or at least object instead of dictionaries. | |
| 14:55:46 | sean-k-mooney | stephenfin: there are two other spec that related to sriov bonding that would be good to group with those | |
| 14:56:05 | edleafe | sean-k-mooney: placement is http, so ovo isn't applicable. | |
| 14:56:26 | edleafe | sean-k-mooney: but yeah, from scheduler->conductor, ovo will help | |
| 14:56:48 | sean-k-mooney | edleafe: it is if you call object_to_primitive and then jsonutils.dump on it first | |
| 14:57:26 | sean-k-mooney | i would be really happy if we could start passing json serialised ovo across our apis in the future | |
| 14:57:33 | openstackgerrit | Vladyslav Drok proposed openstack/nova master: Fix _delete_inventory log message in report client https://review.openstack.org/498833 | |
| 14:58:59 | mriedem | edleafe: jaypipes: dansmith: we should return compute node uuid back in the dict from select_destinations too, that would make lookups for providers on the client side quicker, so the client doesn't need to lookup a compute node by host/node just to get the uuid | |
| 14:59:23 | edleafe | mriedem: agreed. I already did that in my first crack at these objects | |
| 14:59:24 | mriedem | the dict that's returned isn't even versioned today | |
| 14:59:32 | dansmith | I wasn't really paying attention to that conversation so I'll have to catch up when I look at that | |
| 14:59:40 | mriedem | this is unrelated | |
| 14:59:51 | mriedem | but to fix a bug with cleaning up allocations during live migration, | |
| 14:59:57 | jaypipes | mriedem: yup, agreed. good idea. | |
| 15:00:07 | mriedem | i need the compute node uuid to remove the allocations and only have the dict from select_destinations which has the host/nodename | |
| 15:00:10 | edleafe | mriedem: we don't really need that dict per se, just the data in it | |
| 15:00:29 | edleafe | and the object will version that | |
| 15:00:36 | mriedem | so let's say i added a uuid key to the dict that's returned today, what would we change for the version? | |
| 15:00:39 | mriedem | the client rpc? | |
| 15:00:45 | mriedem | *scheduler client rpc? or manager? | |
| 15:01:21 | edleafe | probably the former | |
| 15:01:30 | mriedem | well it's 4.5 either way i think | |
| 15:01:47 | edleafe | but there's no point if we're gonna be changing the interface to use objects | |
| 15:02:24 | mriedem | plus i couldn't backport this, and we need to backport this fix... | |
| 15:02:26 | mriedem | TODO in the code it is | |
| 15:03:24 | beagles | sean-k-mooney, thanks for the info btw! | |
| 15:03:25 | edleafe | mriedem: you mean you don't have any reference to the compute node in the migration? | |
| 15:03:27 | sean-k-mooney | mriedem: well for now you would have to do a condtional check to see if the uuid is present in the dictionary and look it up if not | |
| 15:03:39 | mriedem | edleafe: not when we're still in the conductor task | |
| 15:03:45 | mriedem | i have the dict returned from select_destinations | |
| 15:03:55 | mriedem | which has the host/nodename, which i can use to lookup the ComputeNode to get the UUID | |
| 15:03:56 | sean-k-mooney | beagles: no worries i was off since thursday so just go back | |
| 15:03:57 | mriedem | it's just a hassle | |
| 15:04:05 | artom | claudiub|2, fips have always been associated to ports IIRC | |
| 15:04:19 | edleafe | mriedem: I see | |
| 15:04:22 | mriedem | sean-k-mooney: i'd rather not write in dead code | |
| 15:05:13 | sean-k-mooney | mriedem: agreed but it may be needed for the backport. if we go to ovo defintions and and a object based interfaces then we can do it correctly in that version for queens on | |
| 15:05:48 | claudiub|2 | artom: yeah, you're right. anyways. will let you know what i'll find out | |
| 15:06:32 | artom | claudiub|2, sure, thanks | |
| 15:06:36 | cdent | edleafe: i lied, apparently i had something to say now | |
| 15:07:32 | edleafe | cdent: typical | |
| 15:08:05 | mriedem | sean-k-mooney: i'm just going to lookup the compute node using the host and node strings that i already have | |
| 15:08:07 | mriedem | old school | |
| 15:08:18 | mriedem | we can clean it up in queens with a uuid we get back later if that happens | |
| 15:08:20 | mriedem | hence the TODO | |
| 15:08:53 | cdent | edleafe: well you know, I was taunted, I couldn’t help myself, I’m an easy target for bait | |
| 15:11:25 | sean-k-mooney | speaking of ptg does anyone have the link to nova etherpads? | |
| 15:12:00 | mriedem | sean-k-mooney: https://etherpad.openstack.org/p/nova-ptg-queens | |
| 15:12:15 | sean-k-mooney | mriedem: thanks :) | |
| 15:12:42 | sean-k-mooney | i just got approval at the weekend to travel so not i need to book all the things | |
| 15:13:08 | mriedem | sean-k-mooney: and start coming up with 10 different scheduler filters you want to talk about | |
| 15:13:10 | mriedem | like in ATL | |
| 15:14:22 | sean-k-mooney | haha well perhaps. honestly alot of what i want to talk about on that front will hopefully be covered in the existing scheduler and placement sessiosn | |
| 15:18:58 | sean-k-mooney | that said some topics i want to ensure we can model with placement and traists are, verifed boot(not the same as trusted boot), fpgas for tenants, power management of guests, and bandwith based scheduling(maily for sriov but maybe others in the future). | |
| 15:19:56 | mriedem | i think the huawei product team has a requirement for sriov bandwidth based scheduling | |
| 15:19:59 | mriedem | but i don't really have details yet | |
| 15:20:11 | sean-k-mooney | im also interested in the pci overhal topics and live migration both across vif types and with sriov | |
| 15:20:47 | sean-k-mooney | mriedem: cool we started working on that last year i think rodolfo had some patches up at one point | |
| 15:21:08 | sean-k-mooney | we should be able to model that with nested resource providers. | |
| 15:21:13 | gibi | sean-k-mooney, mriedem: ericsson also interested in the bandwith based scheduling, both for sriov and for ovs port as well | |
| 15:21:25 | sean-k-mooney | each pf would have a pool of bandwidth that could be consumed by the vf | |
| 15:21:45 | sean-k-mooney | gibi: ovs does not support enforcing the bandwidth | |
| 15:21:58 | sean-k-mooney | actully it has bandwidth limits | |
| 15:22:15 | sean-k-mooney | it just does not support mimium bandwidth guarentees | |
| 15:22:40 | gibi | sean-k-mooney: true, but as a first step we can trust the guest to only send/receive the requested bandwidth | |
| 15:22:50 | mriedem | sean-k-mooney: gibi: has anyone started a spec or etherpad for ideas or ML discussion, anything? | |
| 15:23:14 | gibi | mriedem: not from ericsson side | |
| 15:23:21 | sean-k-mooney | we had neutron or nova spec for this for pike | |
| 15:23:47 | gibi | sean-k-mooney: I think there was a neutron spec | |
| 15:24:11 | edleafe | cdent: taunted right back | |
| 15:24:12 | sean-k-mooney | ya ill see about geting a nova one created for queens and add it to the etherpad | |
| 15:24:21 | cdent | edleafe: gracias | |
| 15:24:29 | gibi | sean-k-mooney, mriedem: here is the neutron spec https://review.openstack.org/#/c/396297/ | |
| 15:25:14 | mriedem | cool thanks, i'll pass this along | |
| 15:26:29 | sean-k-mooney | mriedem: we have already landed part of this i think or atleas implemented it. what was missing on the nova side was a way to ensure we dont over subscribe which needs placemnet to do cleanly | |
| 15:27:10 | sean-k-mooney | mriedem: ya the sriov enforcement is implemented here https://review.openstack.org/#/c/401254/25 | |
| 15:28:22 | cfriesen | gibi: is it reasonable to trust the guests to do that? That seems like a private-cloud-only scenario. | |
| 15:28:53 | sean-k-mooney | cfriesen: we dont have to turst them in all cases | |
| 15:29:12 | sean-k-mooney | with sriov the hardwer can enforce the minium bandwith gureentee | |
| 15:29:46 | sean-k-mooney | with vpp we may be able to do that in software and we could also extend ovs in the future to do it but that is a lot more work | |
| 15:30:11 | gibi | cfriesen: this is a temporary solution until OVS can be used to actually limit the bandwidth | |
| 15:30:47 | gibi | cfriesen: but you are correct that this is for private cloud only | |
| 15:31:33 | cfriesen | sean-k-mooney: gibi: we'd probably make use of the SRIOV case | |
| 15:32:43 | sean-k-mooney | for the sriov case the hardware will enforce the minimum bandwidth. provided we dont oversubscribe. eg. create 11 vf with 1G minium bandwith on a 10G nice | |
| 15:39:15 | openstackgerrit | Lucian Petrut proposed openstack/nova master: [wip] Fix nova assisted volume snapshots https://review.openstack.org/498845 | |
| 15:44:16 | cdent | edleafe: back at ya | |
| 15:46:22 | edleafe | cdent: gee thanks! | |
| 15:46:44 | cdent | it’s a pleasure collaborating with you sir | |
| 15:46:48 | mriedem | lpetrut: mostly right, but some comments inline | |
| 15:46:59 | edleafe | cdent: wish I could say the same! :-P | |
| 15:47:09 | cdent | \o/ | |
| 15:48:46 | lpetrut_ | mriedem: great, thanks for the quick review | |
| 15:49:57 | lpetrut_ | mriedem: about the context manager: didn't use that so that the targeted cell remains set | |
| 15:51:32 | mriedem | oh hrm | |
| 15:51:37 | mriedem | lpetrut: yeah good point | |
| 15:51:47 | mriedem | maybe leave a comment in there then, in the code i mean | |