Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2018-10-10
16:28:35 stephenfin sean-k-mooney: I'm going to see if we can do the same thing claudiub did here to avoid that happening again https://review.openstack.org/#/c/470775/
16:29:30 sean-k-mooney stephenfin: im not sure that would have help in this case but i might not under stand awhat that dose enough either
16:30:28 stephenfin sean-k-mooney: Yeah, it wouldn't actually. I thought https://review.openstack.org/#/c/609414/1/os_vif/tests/unit/internal/command/ip/windows/test_impl_netifaces.py was mocking the wrong stuff but it's not
16:31:02 sean-k-mooney this is the improtant test https://review.openstack.org/#/c/609414/1/os_vif/tests/unit/internal/command/ip/test_api.py
16:31:26 sean-k-mooney before we were returing an instance of the ip lib for windows and the module for linux
16:31:43 stephenfin sean-k-mooney: Just left a nit comment on that, actually
16:32:33 mriedem dansmith: on that initial allocation ratios spec, there is one place we still use the compute node fields, and that's in the scheduler https://github.com/openstack/nova/blob/6bf11e1dc14afad78b11d980c2544a3dc41579ff/nova/scheduler/host_manager.py#L256
16:32:46 sean-k-mooney ya i realsied i coudl have used that too. im going to be adding more unit tests wehn ir change the module layout so ill change that then if your ok with that
16:32:51 mriedem so we set the values on the compute (in the RT) and then read in the scheduler
16:33:00 dansmith mriedem: but aren't those just for the filters?
16:33:04 dansmith what else cares about those?
16:33:28 mriedem weighers use it too i guess
16:33:46 dansmith to pick things that aren't set for oversubscription?
16:33:49 dansmith that seems weird
16:34:11 mriedem the numa topology filter is also using it
16:34:12 dansmith oh, numa filter uses it
16:34:18 dansmith hah
16:34:24 dansmith christ
16:34:37 dansmith so
16:34:43 dansmith (a) we should try to get someone to fix that
16:35:02 stephenfin sean-k-mooney: Yup, it's just a nit
16:35:02 dansmith (b) we can just update the compute node record with the same policy as we update placement
16:35:28 mriedem which is use config if set
16:35:35 mriedem that's what the spec proposed yeah
16:35:39 dansmith yep
16:35:44 dansmith okay I see the use in the weigher
16:35:48 mriedem the question was about upgrading from compute nodes that have 0.0 values in the db
16:36:01 mriedem and i said we could online data migrate those by reading config if the values are 0.0
16:36:13 dansmith the 0.0 means that scheduler uses its own config, right?
16:36:26 sean-k-mooney regarding the numa topology filter im not sure it actully need the allocation ratiors
16:36:27 mriedem no
16:36:43 dansmith allocation_ratio of zero makes no sense otherwise right?
16:36:47 mriedem nothing in the scheduler reads config for allocation ratios
16:36:58 dansmith mriedem: it used to though yeah?
16:37:04 mriedem in the long long ago i guess,
16:37:07 mriedem oh well,
16:37:12 mriedem the facade would i guess yeah...
16:37:30 dansmith point being, 0.0 is nonsense
16:37:30 mriedem the ComputeNode._from_db_object would read from config if 0.0 in the db
16:37:38 jaypipes dansmith: ++
16:37:55 dansmith so we could leave that, and only set it if they override per compute.. same logic as placement
16:38:05 dansmith and anything existing with 0.0 gets the same treatment as usual I guess
16:38:26 mriedem leave the facade in ComputeNode._from_db_object?
16:38:49 dansmith i imagine it won't affect compute's use if it is ignoring what is set on the object, so .. sure?
16:38:58 sean-k-mooney stephenfin: for https://github.com/openstack/nova/blob/0163b9bfb54aaa89b0574c86e7fd36321eebccfe/nova/scheduler/filters/numa_topology_filter.py#L91-L102 can you think of a use case wehre we actully need the allcoation ratiios here
16:39:33 sean-k-mooney stephenfin: we are not allowed to over subsibe against our selves wehn fitting to a host so setting them to 1.0 i think would be valid
16:42:05 sean-k-mooney stephenfin: i would have to go through the code to check but i think we could make the numa topolgy filter work without them.
16:42:50 sean-k-mooney dansmith: although it sound like we dont need to remove them if we go wtih the facade right
16:43:14 dansmith sean-k-mooney: it won't be right if you do
16:43:32 dansmith sean-k-mooney: because if people set the ratio in placement per-compute, which is what we're trying to enable with all of this work,
16:43:38 dansmith the filter will consider a value other than what is in placement
16:44:04 dansmith I'm guessing we don't get back ratios in /a_c, but if we did, we could update our host states before we call the filters
16:44:12 sean-k-mooney dansmith: yes but this is not using the placement value anyway
16:44:18 jaypipes mriedem: why would we keep the facade stuff in ComputeNode._from_db_object()?
16:44:38 dansmith sean-k-mooney: right but right now they have to be the same
16:44:39 dansmith sean-k-mooney: in the future they will not be
16:44:52 dansmith jaypipes: see the discussion just now on the filters that use it, and existing computes in the db with 0.0 set
16:45:21 mriedem jaypipes: we either need to leave that or online data migrate the 0.0 entries from existing records in the db on read
16:45:21 sean-k-mooney dansmith: placement woudl have already filtered out any host that did not pass its allocation right. so we should not have to check twice and sicne we can oversubsibe againat ourselve an allocation ration of 1 i think would still be correct
16:45:46 sean-k-mooney dansmith: anywway its not important right now i guess
16:45:49 mriedem jaypipes: because the compute won't deal with those fields on the object, and if not set, the scheduler reads them from config via the facade
16:45:55 dansmith sean-k-mooney: I didn't look to see what that code in the filter was doing, so maybe?
16:46:00 dansmith sean-k-mooney: if so, we just remove it right?
16:46:43 sean-k-mooney dansmith: ya i think soo but i would have to double check the hardware.numa_fit_instance_to_host fucntion first
16:47:28 dansmith we get back the resource summaries from the providers in /a_c, so including the allocation_ratios in there might be useful for things like this and for the weigher case
16:48:30 sean-k-mooney dansmith: ya. the filters dont currently have access to the allocation candiates today is that correct
16:48:39 sean-k-mooney unless there in the spec_objec?
16:48:42 dansmith sean-k-mooney: that isn't what I'm saying
16:49:01 dansmith I'm saying the scheduler, when it gets back candidates, gets a summary of all covered providers, with inventory information
16:49:11 dansmith if that included the ratios, it could update host_states before calling the filter loop
16:49:46 jaypipes sean-k-mooney: pls see my comment on https://review.openstack.org/#/c/609414/
16:49:59 sean-k-mooney oh ok that would work too ya i was assuming you were suggsting passing in the candiates but your way we do that update once and dont have to update any code in the filters
16:50:52 sean-k-mooney jaypipes: there isnt a bug number because i just got pingged on irc this morning but i can open one
16:51:11 dansmith mriedem: melwitt tssurya: cells meeting today?
16:51:17 mriedem nack
16:51:24 mriedem we might want to just cancel that meeting
16:51:39 dansmith I'd also be fine with that
16:51:50 dansmith I think I suggested that last year even
16:53:39 jaypipes sean-k-mooney: yes pls. if this is truly a "currently broken for hardware offloaded ovs" scenario, it definitely should be a bug.
16:54:50 tssurya dansmith: no problems in cancelling
16:55:03 sean-k-mooney its broke for all ovs backends that use ip command because api _get_impl on linux retrun a module instead of the insatnce of the pyroute2 class
16:56:12 mriedem dansmith: you want to propose the change to cancel the meeting or want me to?
16:56:13 sean-k-mooney jaypipes: os that is ovs + iptables or ovs + hardware offloads. i have added some extra test to catch this case
16:56:22 dansmith melwitt: what do you think about canceling the cells meeting altogether and making it ad-hoc as needed?
16:56:35 dansmith mriedem: I figure we don't need to make a federal case out of it
16:56:42 dansmith if everyone agrees, we just take it off the schedule
16:56:45 mriedem dansmith: it's literally in a schedule though
16:56:56 dansmith I know
16:56:56 mriedem http://git.openstack.org/cgit/openstack-infra/irc-meetings/tree/meetings/nova-cells-v2-meeting.yaml
16:57:02 dansmith oh,
16:57:04 dansmith that schedule
16:57:16 mriedem yeah, free up the time slot in that channel if we're not going to use it
16:57:18 dansmith I thought it was just on the old wiki list
16:57:28 mriedem no this is very official and federale
16:57:34 dansmith I'll propose.. I need to do something useful today
17:01:44 melwitt dansmith: yup sounds ok to me
17:01:57 dansmith https://review.openstack.org/#/c/609496/
17:09:35 mriedem hmm live migration failure in the gate, not something i've seen before i don't think, looks like it was aborted but i'm not sure why
17:09:35 mriedem http://logs.openstack.org/31/606031/4/check/nova-live-migration/9d106bb/logs/subnode-2/libvirt/libvirtd.txt.gz#_2018-10-10_15_27_01_313
17:09:40 mriedem 2018-10-10 15:27:01.313+0000: 18210: error : qemuMigrationFinish:5533 : migration successfully aborted

Earlier   Later