Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2019-08-13
23:03:22 melwitt yeah, true
23:03:23 sean-k-mooney i proably performs better
23:03:49 openstackgerrit Dustin Cowles proposed openstack/nova master: WIP: Load custom providers to Resource Tracker https://review.opendev.org/676029
23:04:03 melwitt yeah ok, this isn't actually different, just the mechanism is different (better) and then you can get it to warn you if you didn't set the tenant keys on all your aggregates
23:05:14 sean-k-mooney yeah so its the waring/force fail behavior is slightly better but not a lot
23:06:30 mriedem the tenant to aggregate stuff was for cern i think, to model tenant-to-cell with azs / aggregates for a cell
23:06:35 mriedem since cern maps tenants to specific cells
23:08:07 mriedem and with that i'm out
23:08:55 melwitt yeah, I was thinking this was for CERN. the customer case I'm writing this up for is also dividing up the cluster per tenant
23:11:02 sean-k-mooney cern used to use the tenat isoaltion filter to do that tenat to cell mapping with cells v1 when they had two levels of scheduler
23:11:19 sean-k-mooney so ya this was a partity thing for them to be able to move to cells v2
23:15:41 melwitt ah, thanks
23:16:20 melwitt I knew the basics of that but a lot of these scheduler things confuse me, so here I am :P
23:20:13 sean-k-mooney ya cern is an edge case as they have so many host they needed to use the limit parmater on the placmenet api and and if they did that without the aggrate feature placmenet could return only host that did not map to the correct cell
23:20:38 sean-k-mooney and when it hit the tenat isolation filter it would eliminate all hosts
23:20:44 melwitt ah right
23:21:39 sean-k-mooney so now they use this pluse a limit of like 15 host or something similarly tiny
23:21:49 sean-k-mooney the default limit being 1000
23:47:02 melwitt sean-k-mooney: with forbidden aggregates, do you know off the top of your head what happens with aggregate-less hosts? will they still come back as eligible to be scheduled upon or no?
23:48:27 sean-k-mooney i belive the intent is the prefilter will append a forbidine aggreate requrest based on info in the the request spec to the placment query
23:48:56 sean-k-mooney so i guess they would be allowed
23:49:15 sean-k-mooney but no i dont know off the top of my head what the actul behavior is for certin
23:49:57 melwitt yeah, ok. that's what I was thinking
23:51:37 sean-k-mooney the spec is a little light on details
23:52:46 sean-k-mooney oh i see we went with sotring the tratit requriemetn in the hostaggreate metadata
23:52:57 sean-k-mooney * nova host agggreate metadata
23:53:53 melwitt ok, so yeah need everything to be in an aggregate in order for it to be excluded. I think.
23:54:14 sean-k-mooney yep exactlyu like my old filter worked...
23:54:27 openstackgerrit Takashi NATSUME proposed openstack/nova stable/stein: Fix misuse of nova.objects.base.obj_equal_prims https://review.opendev.org/676289
23:54:37 sean-k-mooney except that predated placmenet and match on flavor extra_specs
23:55:09 sean-k-mooney actully this is a little less flexiable but its close
23:55:19 melwitt RIP old filter
23:56:16 sean-k-mooney melwitt: how would you feel about adding cpu pinning or hugepage traits?
23:56:39 melwitt I don't know :( I know very little about those
23:57:06 sean-k-mooney they dont exisats. yet... actully i have a better way to do what i want to do if i have thos
23:57:12 sean-k-mooney *those
23:57:45 openstackgerrit Takashi NATSUME proposed openstack/nova stable/rocky: Fix misuse of nova.objects.base.obj_equal_prims https://review.opendev.org/676290
23:58:16 sean-k-mooney dpdk requried you to have hugepages to work with openstack
23:58:56 sean-k-mooney if we have a hugepage traits we could make it so your dpdk host would not allow vms without hugepages to be scheuled
23:59:22 sean-k-mooney for pinning we already have the dedicated cpus spec
#openstack-nova - 2019-08-14
00:00:19 openstackgerrit Takashi NATSUME proposed openstack/nova stable/queens: Fix misuse of nova.objects.base.obj_equal_prims https://review.opendev.org/676291
00:00:24 sean-k-mooney i would proably do this slightly differently however. e.g. not requrie the operator to confgiure aggreates and do it like my image metadtata tratis changes
00:02:09 openstackgerrit Takashi NATSUME proposed openstack/nova stable/pike: Fix misuse of nova.objects.base.obj_equal_prims https://review.opendev.org/676292
00:02:49 sean-k-mooney melwitt: i havent reviewd the code for this but i hope they are caching the aggreate metadata
00:04:21 sean-k-mooney the only why the prefilter could work would eb to retrive all of the nova host aggreates and there metadata then calulate the aggrates that wont suppor the instance and generate the &member_of=!in:<agg1>,<agg2>,<agg3> query paramater
00:05:02 melwitt sean-k-mooney: here's the code, it's close to merging so get in on it while you can https://review.opendev.org/#/q/topic:bp/placement-req-filter-forbidden-aggregates+(status:open+OR+status:merged)
00:07:33 sean-k-mooney ya so no caching...
00:07:52 sean-k-mooney the proablem with caching is keping the cache valid
00:11:05 sean-k-mooney melwitt: the schduler does not directly talk to the db right? i.e. does the schduler acess the db via the conductor like teh compute service?
00:11:35 melwitt sean-k-mooney: I think it talks directly
00:11:39 melwitt only the compute service does not
00:12:00 sean-k-mooney ok so all the "control plane" service talk directly but not comptues
00:13:25 melwitt I think so
00:13:58 sean-k-mooney ok i was not sure if it was only the api,metadata service and conductor that went direct.
00:14:29 sean-k-mooney i guess it makes sense for the schduler to be able to take a fast path to the db
00:41:46 melwitt sean-k-mooney: you can see which services go through conductor by looking for 'indirection_api' under nova/cmd
00:42:23 sean-k-mooney indirection_api is new to me
00:43:32 melwitt yeah, that's the magic for objects going through conductor
00:44:15 sean-k-mooney ah ok.
00:46:47 sean-k-mooney the way the prefilter is workign bugs me
00:47:15 sean-k-mooney i need to re read the spec but i dont think it workign the way i expect it too
00:47:21 sean-k-mooney it may be equivalent
00:48:44 sean-k-mooney if the image has no trait request i think it would allow all aggreates
00:49:10 sean-k-mooney and i think you would have to list all traits in the imgae/flavor on the aggreate for it to match.
00:49:38 sean-k-mooney i proably should be less tired when reviewing this but this seams reversed to me
00:52:44 melwitt yeah, I think it will get all the aggregates, add any without the trait to the "no" bin, and then pass the "no" bin to placement and say "these are the forbidden aggregates, don't return hosts in them"
00:53:02 sean-k-mooney i think the logic is inverted.
00:53:13 melwitt which does seem backward but I assume it's the only way to do it
00:53:26 sean-k-mooney its treating the image and flaovr as the attoritive source nto the metadata
00:53:40 sean-k-mooney this is broken in the same way the existing filter is
00:53:58 sean-k-mooney and it the different betten the out of tree one and the in tree one
00:55:01 sean-k-mooney for this prefilter to work every trait that is in the image+flavor would have to be on the host_aggreate metadata
00:55:02 melwitt it does specifically say (in the proposed doc change) that image/flavor need not be set and will still not land on hosts in aggregates with =required. I just don't know how that's done
00:55:34 sean-k-mooney its done by invertin the relationship
00:56:10 openstackgerrit Takashi NATSUME proposed openstack/nova stable/ocata: Fix misuse of nova.objects.base.obj_equal_prims https://review.opendev.org/676295
00:56:14 sean-k-mooney so what you need to do is get the metadata for all host_aggreates. then for each aggreate you check all the keys are present itn the flvore/image requst
00:56:32 sean-k-mooney if a key is set on the aggreate but not in the flaor/imgae then you add that to the forbindin list
00:56:34 melwitt oh ok
00:56:38 melwitt yeah
00:56:56 sean-k-mooney again exactly what the out of tree filter does :)
00:57:09 sean-k-mooney ok ill comment on the review
00:58:03 sean-k-mooney as implement this would work exactly the same as the aggregate_instance_extra_specs filter
00:58:24 sean-k-mooney well it sort of
00:58:41 sean-k-mooney it use traits instead of any proerty and also looks at the image
00:58:41 openstackgerrit Takashi NATSUME proposed openstack/nova stable/ocata: Fix misuse of nova.objects.base.obj_equal_prims https://review.opendev.org/676295
00:58:47 sean-k-mooney but conceptuly its the same
01:01:14 openstackgerrit Takashi NATSUME proposed openstack/nova stable/ocata: Fix misuse of nova.objects.base.obj_equal_prims https://review.opendev.org/676295
01:04:05 sean-k-mooney melwitt: actully it might be correct https://review.opendev.org/#/c/671074/6/nova/objects/aggregate.py@476
01:04:58 sean-k-mooney in need to figure out what sql that generates
01:10:15 melwitt sean-k-mooney: note the ~ (not) squiggle
01:11:47 sean-k-mooney ya looking at the doc sting i think its correct after all
01:12:16 melwitt phew good
01:12:49 sean-k-mooney so it selecting the aggrreate if the aggreate metadta key is not in the set passed in
01:16:34 openstackgerrit ZHOU YAO proposed openstack/nova master: Preserve UEFI NVRAM variable store https://review.opendev.org/621646
01:32:32 sean-k-mooney ok im goign to log off for the night o/
01:43:44 sean-k-mooney oh infra change how logs are renderd in the gate.
01:44:12 sean-k-mooney i dont think i like it just becasue its different the it has been for 6+ years
01:47:51 openstackgerrit Takashi NATSUME proposed openstack/nova master: api-ref: Fix collapse of 'host_status' description https://review.opendev.org/676301
02:23:53 openstackgerrit Sundar Nadathur proposed openstack/nova master: ksa auth conf and client for Cyborg access https://review.opendev.org/631242
02:23:53 openstackgerrit Sundar Nadathur proposed openstack/nova master: Refactor some methods for reuse by Cyborg-related code. https://review.opendev.org/673734

Earlier   Later