| Posted | Nick | Remark | |
|---|---|---|---|
| #openstack-nova - 2018-09-20 | |||
| 13:57:48 | bauzas | just throw me tomatoes now if I regressed :) | |
| 13:57:59 | dansmith | bauzas: you did regress | |
| 13:58:03 | dansmith | but I will fix | |
| 13:58:11 | efried | stephenfin: Can you please see if I used the correct arguments here? https://review.openstack.org/#/c/603993/ | |
| 13:58:28 | bauzas | dansmith: you're the hero and I'm the villain then ;) | |
| 13:58:38 | dansmith | per usual | |
| 13:58:58 | bauzas | I can snip fingers, don't tempt me | |
| 13:59:01 | bauzas | snap* | |
| 14:37:41 | openstackgerrit | Dan Smith proposed openstack/nova master: Filter deleted computes from get_all_by_uuids() https://review.openstack.org/604108 | |
| 14:45:07 | mriedem | mdbooth: did you see the evacuate integration tests are passing? | |
| 14:45:53 | mdbooth | mriedem: I didn't, but I will as soon as poss :) | |
| 14:46:39 | mriedem | not much to see - we evacuate with the down compute so it fails, and then bring libvirtd back up and evacuate and it passes | |
| 14:46:54 | mriedem | only question is if it should be horked into the live migration job which is convenient but sorta confusing | |
| 14:47:23 | mdbooth | mriedem: Ideally we'd have it in a few places: would be good to cover both shared and non-shared. | |
| 14:47:46 | mdbooth | Perfectly we'd cover 'all 4' cases, but I don't think we have all 4 in CI. | |
| 14:48:24 | mriedem | all 4 is what again? volume-backed, local disk not shared, local disk shared | |
| 14:48:26 | mriedem | what's the other? | |
| 14:48:40 | mdbooth | 2 axes are: shared instance, shared disk. | |
| 14:48:41 | mriedem | right now the patch is running on ceph local shared | |
| 14:49:11 | mriedem | the script could be updated to create a volume-backed server and evacuate again | |
| 14:49:42 | mnaser | is there a reason why running debug=True doesn't show the http requests done by placement client? | |
| 14:49:50 | openstackgerrit | Eric Fried proposed openstack/nova master: WIP: libvirt: Turn off max_unit https://review.openstack.org/604110 | |
| 14:50:02 | mnaser | rather than reconstructing the http request, it'd be nice to know what nova-scheduler is asking placement | |
| 14:50:27 | mriedem | we call placement over ksa, | |
| 14:50:41 | mriedem | so our debug=True logging is likely not carrying over to ksa | |
| 14:51:19 | mnaser | because before debug=True would show you the filters running through things | |
| 14:51:27 | mnaser | and you'd be able to debug scheduler issues | |
| 14:51:43 | mriedem | you just want to see like the GET /allocation_candidates request right? | |
| 14:51:45 | mnaser | now it just starts filtering a subset and you don't know where that came from so yeah | |
| 14:51:59 | mnaser | stuff like that, yes, so that i know why this live migration is not getting scheduled on one of the hosts | |
| 14:52:07 | mnaser | there's empty capacity but it's not adding it | |
| 14:52:21 | mnaser | probably something along the lines of placement being inconsistent | |
| 14:52:57 | mriedem | are you passing the specific target host for the live migration to the compute API? | |
| 14:53:05 | mriedem | because i know of another bug related to that... | |
| 14:53:15 | mnaser | nope im not passing one | |
| 14:53:30 | mriedem | ok, you just expect a certain host to be picked b/c of weights and capacity etc? | |
| 14:53:54 | mnaser | yeah, it's not even hitting weight/capacity, placement is not returning a host that has capacity | |
| 14:54:08 | mriedem | i think we'd have to add something like this https://github.com/openstack/nova/blob/master/nova/config.py#L39 | |
| 14:54:09 | efried | Does the placement log not show the full command line? | |
| 14:54:11 | mnaser | which tells me there's probably incorrect allocations in placement that make that host look 'overloaded' | |
| 14:54:22 | mnaser | ah that's a good alternative, checking uwsgi logs on placement side | |
| 14:54:30 | mriedem | there should be the same request id on both sides | |
| 14:54:46 | mriedem | that's your workaround for now | |
| 14:54:53 | efried | The placement logs by default won't show the payloads (request or response) but the queryparams are there I thought. | |
| 14:55:01 | mriedem | yes they should be | |
| 14:55:09 | mnaser | i should be able to rerun that http request | |
| 14:55:24 | mriedem | http://logs.openstack.org/75/470775/38/check/tempest-full/a97793a/controller/logs/screen-placement-api.txt.gz#_Sep_06_13_02_32_181234 | |
| 14:55:25 | efried | sorry, "command line" was a bogus phrase | |
| 14:55:29 | mriedem | Sep 06 13:02:32.181234 ubuntu-xenial-ovh-bhs1-0001824138 devstack@placement-api.service[6991]: DEBUG nova.api.openstack.placement.requestlog [req-68f9280e-61b2-446e-8622-a64e27e7c617 req-df696f1f-3435-4b6c-8a3b-c6dc1c58c1ba service placement] Starting request: 158.69.79.172 "GET /placement/allocation_candidates?limit=1000&resources=MEMORY_MB%3A64%2CVCPU%3A1" {{(pid=6993) __call__ /opt/stack/nova/nova/api/openstack/placement/r | |
| 14:55:29 | mriedem | stlog.py:38}} | |
| 14:55:54 | mriedem | must be a bfv instance ^ because there is no DISK_GB requested | |
| 14:56:35 | mnaser | yup /allocation_candidates?limit=1000&resources=DISK_GB%3A10%2CMEMORY_MB%3A2048%2CVCPU%3A4 | |
| 14:56:36 | mnaser | lets see | |
| 14:58:20 | mnaser | openstack allocation candidate list --resource DISK_GB=10 --resource MEMORY_MB=2048 --resource VCPU=4 => Operation or argument is not supported with version 1.0 | |
| 14:58:25 | mnaser | mriedem: do i need some special env variable? | |
| 14:58:49 | mnaser | sigh | |
| 14:58:49 | mnaser | --os-placement-api-version 1.10 | |
| 14:58:52 | mnaser | ill read docs from now on | |
| 14:59:47 | gibi | FYI, I've added the bp use-nested-allocation-candidates to the 3rd runway slot | |
| 14:59:51 | gibi | First part of https://blueprints.launchpad.net/nova/+spec/use-nested-allocation-candidates (gibi) [END: 2018-10-04] | |
| 15:00:01 | cdent | i guess I need to learn osc-placement so I can parse this stuff. I'm still back on raw curl which is nice and comfy and explicit | |
| 15:00:05 | mnaser | yeah it's returning one host (the host itself) which gets filtered out in a live migration, i think i'll have to write up the clean up / audit stuff | |
| 15:00:36 | mnaser | cdent: it's nice to not have to write it all out esp to grab a token and then build out a curl request | |
| 15:01:26 | mriedem | mnaser: yeah osc is opt-in per request for microversions, unlike nova cli | |
| 15:01:36 | mriedem | you could just set an env var if you want the latest, | |
| 15:01:49 | mriedem | export OS_PLACEMENT_API_VERSION=<whatever you have that is the latest> | |
| 15:03:06 | cdent | mnaser: yeah, totally agree that if you haven't built the curl-based habits the osc way is good. It's just that since I've been curl-ing against placement for such a long time I've got habits (including the token getting and using parts) that are rather built in. I'm complaining about myself here, not other people. | |
| 15:03:22 | mnaser | cdent: yeah i have a lot of terrible old habits | |
| 15:03:32 | mnaser | cdent: like using LOG.debug() instead of an actual debugger | |
| 15:03:50 | cdent | dood! LOG.debug() is too fancy. sys.stderr.write() 4evah | |
| 15:04:07 | mnaser | System.out.println() | |
| 15:04:11 | mnaser | at least we're not in that darkest timeline | |
| 15:06:40 | tobias-urdin | when live migrating is the vcpu_pin_set option a 1:1 mapping for source and dest host? I start at the same lowest possible core but some machines have different amount of vcpus available, if I migrate an instances that has vcpu 79 but source host only has 60 vcpu will it reallocate where to place that or error out? | |
| 15:07:37 | mnaser | i think that's a libvirt decision | |
| 15:08:33 | tobias-urdin | you are probably correct, will trace some more | |
| 15:08:54 | openstackgerrit | Brin Zhang proposed openstack/python-novaclient master: Add support changes-before for microversion 2.66 https://review.openstack.org/603549 | |
| 15:09:25 | mriedem | i didn't think live migration + cpu pinning was supported yet? | |
| 15:09:26 | mriedem | cfriesen: ^? | |
| 15:12:10 | cfriesen | mriedem: tobias-urdin: live migration doesn't currently recalculate CPUs on the dest node. This is one of the things covered under the "numa-aware live migration" blueprint. | |
| 15:12:52 | tobias-urdin | cfriesen: roger that, thanks :) | |
| 15:13:49 | cfriesen | tobias-urdin: the live migration *might* succeed, but it might end up putting multiple guest CPUs on the same host CPU, or it might fail due to different host topology. | |
| 15:14:15 | tobias-urdin | ok, that explains why it worked atleast thanks | |
| 15:14:46 | zzzeek | jaypipes-ooo: think fast! what's your favorite innodb_flush_method setting ! don't think! | |
| 15:16:43 | jaypipes-ooo | zzzeek: direct? | |
| 15:16:50 | zzzeek | jaypipes-ooo: OK | |
| 15:17:22 | zzzeek | jaypipes-ooo: we're adding the option to tripleo | |
| 15:17:46 | jaypipes-ooo | zzzeek: but only if there is a battery-backed write cache on the disks. | |
| 15:17:55 | jaypipes-ooo | zzzeek: otherwise, keep it normal. | |
| 15:18:39 | openstackgerrit | Merged openstack/nova stable/rocky: Fix nova-status "_check_resource_providers" check https://review.openstack.org/600098 | |
| 15:19:11 | openstackgerrit | Brin Zhang proposed openstack/python-novaclient master: Add support changes-before for microversion 2.66 https://review.openstack.org/603549 | |
| 15:20:07 | zzzeek | jaypipes-ooo: OK is that impacting the durability of the logs only? usually b.c. we use galera, we assume write sets have also gone to other nodes | |
| 15:20:33 | zzzeek | jaypipes-ooo: so w/ these kinds of options, we have to worry about all galera nodes power off simultaneously, but if only one machine, then not | |
| 15:20:42 | openstackgerrit | Balazs Gibizer proposed openstack/nova master: Run negative server moving tests with nested RPs https://review.openstack.org/604125 | |
| 15:20:53 | zzzeek | jaypipes-ooo: we certainly arent changing this setting by default | |
| 15:21:08 | zzzeek | jaypipes-ooo: Im just trying to create my own importance by curating what option we expose :) | |
| 15:21:44 | jaypipes-ooo | zzzeek: :) just leave it defaulted to normal then. It's mostly micro-tuning for write-heavy environments anyway. | |
| 15:21:54 | jaypipes-ooo | zzzeek: which openstack's DBs definitely are *not*. | |
| 15:22:04 | zzzeek | jaypipes-ooo: yes. I just need to have an intelligent answer for why someone would want to set this | |
| 15:22:19 | zzzeek | jaypipes-ooo: it's usually folks with old OSP and struggling w/ huge keystone token table | |
| 15:22:50 | jaypipes-ooo | zzzeek: ack. there's a solution to that, though... TRUNCATE TABLE tokens; :P | |