Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2022-08-24
09:59:38 sean-k-mooney Uggla: https://github.com/openstack/oslo.db/blob/master/oslo_db/sqlalchemy/models.py#L129-L137 is where soft delete is implemented
10:00:12 gibi the ovo base class adds deleted_at column today
10:00:16 gibi to the ovo
10:00:23 sean-k-mooney to the object sent over the wire
10:00:36 sean-k-mooney but that does not mean it needs to be in the api reponce or the db model
10:01:23 sean-k-mooney we can just ignore them. if we want to have a new base class that does not have them that is also ok
10:01:25 gibi my original comment on Uggla's patch was to not hack out the deleted_at column from the ShareMapping ovo but not add it in the first place
10:01:50 sean-k-mooney ya im fine with that too
10:02:01 gibi https://review.opendev.org/c/openstack/nova/+/839401/6/nova/objects/share_mapping.py#49
10:02:54 sean-k-mooney so they dont need to do that
10:03:05 sean-k-mooney if they just dont defien the columns in the db model
10:03:52 gibi honestly I don't want to to have columns in a new ovo that are not used
10:04:05 sean-k-mooney ack
10:04:18 sean-k-mooney we have that alredy for many object
10:04:37 gibi yes, but don't increase that debt if not really necesary
10:04:44 sean-k-mooney sure
10:05:05 sean-k-mooney so i think we are agreed Uggla should add one new base class without those fields
10:05:09 sean-k-mooney and not touch any of the rest
10:05:38 sean-k-mooney and we can disucss at the ptg if we will have time to do object cleanup in AA or BB and the faith of shadow tables
10:05:55 gibi agreed
10:05:56 gibi :)
10:06:24 Uggla sounds good to me. thanks.
10:08:28 gibi sean-k-mooney: re pci: I have to retract my original statement about nova splitting pools per PF. It splitted in my original test due to numa node differences (the fixture creates numa automatically). So on the same numa node we create pools with multiple PFs (of PCI devs). But as I already confirmed that the scheduler logic can split a request over multiple pools I can (and will) cahnge the pooling
10:08:34 gibi logic not to merge pools from different PFs or PCI devs
10:09:37 sean-k-mooney ack
10:09:39 gibi I realized that when I created a host with 3 PCI devs. so numa0 got two
10:09:47 sean-k-mooney ah
10:09:55 sean-k-mooney ya so i know it did it for numa
10:10:02 sean-k-mooney but within a numa node i was not sure
10:10:11 sean-k-mooney i think it also splits for differnnt tags
10:10:15 sean-k-mooney i.e. physnets
10:10:21 sean-k-mooney or tursted vs non trusted
10:10:24 gibi yes I assumed so, what I did not relaized that the fixture automatically split devices equally between numa0 and numa1
10:10:31 sean-k-mooney but i guess it combines if they are the same
10:10:49 sean-k-mooney ah yes it does
10:11:06 sean-k-mooney you can just create the device manually but the fixture i generally nicer to use
10:11:14 gibi yep
10:13:36 gibi basic cold migrate and resize works with miniumal changes in the conductor so I think evac and unshelve will be easy too
10:13:53 gibi and live migration is not supported for flavor based PCI so that is super easy :D
10:14:31 gibi then I will look at resize revert, reschedule, and multi create in this order
10:15:14 sean-k-mooney lol yep live migration should be trivial :P
10:15:45 sean-k-mooney fortunetly you can also verify them on real hardware for a change too once its working in the func test env
10:15:58 gibi yes I will do a final round in the lab too
10:16:17 sean-k-mooney i still need to go extend the reservation fo those nodes
10:16:20 sean-k-mooney ill do that now
10:16:22 gibi thanks
10:16:52 gibi I use those nodes to run the func and unit tests in bluk too as it is a lot faster there :D
10:17:24 gibi I rsync up my local dev repo and run tox via ssh
10:17:31 sean-k-mooney yep thats why i often do dev on my home server
10:17:49 opendevreview ribaudr proposed openstack/nova master: Default Nova persistent objects without soft delete. https://review.opendev.org/c/openstack/nova/+/854355
10:18:05 sean-k-mooney its avaiabel until 02-Nov-2022 but that seem longer then we need any prefered end date
10:18:15 sean-k-mooney we can return them early too
10:18:26 sean-k-mooney will i extend it to october 1st?
10:18:34 sean-k-mooney that gives us a month of grace period
10:19:01 gibi oct 1 is totally ok
10:20:49 sean-k-mooney ok done
10:20:58 gibi thanks
10:21:38 gibi OK, so pools are splitted on these keys https://github.com/openstack/nova/blob/master/nova/pci/stats.py#L66 and I need to actaully check parent_addr there as well
10:22:25 gibi if the parent_addr is None or not equal betweent two devs then we need to split
10:22:51 sean-k-mooney yep that makes sense
10:23:03 sean-k-mooney although hum
10:23:11 sean-k-mooney we shoudl be spliting on more then that
10:23:15 sean-k-mooney like phsynet
10:23:44 gibi yes for neutron based sriov we might need that too
10:24:02 sean-k-mooney can you add that while your adding the partent adress
10:24:32 sean-k-mooney im wondering about trusted too
10:24:49 sean-k-mooney i dont know if we need to split based on that
10:24:49 gibi wait
10:25:00 gibi I have to correct myself
10:25:18 gibi https://github.com/openstack/nova/blob/94065763d32287606895c07bd5882bab083a4e48/nova/pci/stats.py#L136-L138
10:25:29 gibi we split on both those dev fields and the devspec tags
10:25:39 gibi to physnet is handled
10:26:02 sean-k-mooney ah ok as is trusted
10:26:32 sean-k-mooney so this will auto split on traits and resouce class
10:26:50 gibi hehe :)
10:26:50 sean-k-mooney technially we allow operator provided tags too but they were never uable for anything
10:26:53 sean-k-mooney we just ignore them
10:27:46 sean-k-mooney at one point there was talk of allowign the pci alias to match on extra tags
10:27:51 gibi I had to ignore traits and resource_class https://review.opendev.org/c/openstack/nova/+/853316/4/nova/pci/stats.py to keep the pool matching work
10:28:55 sean-k-mooney ah ok hehe
10:29:10 gibi it is mostly ther to keep _filter_pools_for_spec happy as the request contains the traits tag but the pool will not mapped to traits just to RPs
10:29:35 sean-k-mooney ya thats proably fine
10:31:25 sean-k-mooney will we have the tags in the pools
10:31:28 sean-k-mooney *traits
10:31:50 sean-k-mooney i assume the intent is just to relay on placment to do the trait/rc filtering
10:32:07 sean-k-mooney and then we use the rp id to corralate the pools with the allcoation candaate
10:32:12 sean-k-mooney so we dont need to check them in the filters
10:32:27 sean-k-mooney so we can just not put them in the pools
10:33:17 gibi we use the rp_uuid to correlate the request with the pool
10:33:23 gibi the rest is doen in placemnet
10:34:16 sean-k-mooney yep that is what i was expecting
10:34:36 gibi the pooling logic uses all the dev_spec tags automatically so I needed to explicity ignore traits and resource_class there
10:34:47 gibi to not to put them into the pool
10:34:50 gibi as we don't need them
10:35:04 gibi and it also won't match with the request in generic way
10:35:09 sean-k-mooney so your going to do two change right. 1 split the pools now also by parent adress and 2 split the alias requests into multipel instance_pci_request object if the alias request more then one of something
10:35:18 gibi yes
10:35:21 gibi the later is already up
10:35:28 gibi I doing the former now
10:35:55 gibi https://review.opendev.org/c/openstack/nova/+/852771/6/nova/objects/request_spec.py#540

Earlier   Later