Posted	Nick	Remark
#openstack-nova - 2018-10-16
19:55:20	dansmith	of course not everyone does.. OSA does I think, and hopefully all the buzz around making this a generic thing will mean in a year people are looking at it
19:55:35	dansmith	I thought you were also suggesting clarifying words in renos to help understand
19:55:56	dansmith	I was just saying flipping the default behavior now and trying to document _that_ is the confusing part
19:55:59	melwitt	yeah but IIUC that doesn't help someone upgrading to rocky if I can't backport those words
19:56:08	melwitt	oh
19:56:12	dansmith	you can backport the words,
19:56:33	dansmith	I just don't think you should change the behavior and backport more words explaining how it's changed for the third time
19:56:45	melwitt	got it, ok
19:57:30	mriedem	what would the nova-status upgrade check look for? that [workarounds]/enable_consoleauth is False and return a warning?
19:58:02	dansmith	yeah, and maybe check the services table or current tokens to see if you even use that stuff
19:58:09	dansmith	if you don't use console, then you don't need to warn,
19:58:26	dansmith	but if you do and you're rolling, pretty much should have that set right?
19:58:33	melwitt	yeah
19:58:45	mriedem	you can also tell if there are no console auth entries in the db right?
19:58:50	dansmith	I said that
19:59:08	mriedem	i said it with an accent
19:59:19	dansmith	fancy
19:59:21	mriedem	dansmith: btw, https://review.openstack.org/#/c/611094/ needs an assertion on it
19:59:23	melwitt	there wouldn't be, before rocky though. they'd be in the nova-consoleauth service
19:59:31	mriedem	melwitt: well, that's the point right?
19:59:45	mriedem	if you're using nova-consoleauth in queens, and upgrading to rocky, you want the workaround enabled
20:00:05	dansmith	mriedem: ack, I have to run off for a bit but will hit that when I get back
20:00:05	mriedem	and nova-consoleauth would show up in the services table in....one of the dbs
20:00:08	melwitt	yeah, I mean, if you are checking a queens deployment for whether they use consoles at all, you'd have to check the nova-consoleauth service
20:00:26	mriedem	we don't want to make an rpc call from the status check
20:00:32	mriedem	but we could check the services table to see if it's been started
20:00:35	melwitt	oh, you're thinking if they don't use consoles they won't run the service at all. that makes sense too
20:00:37	melwitt	yeah
20:00:51	mriedem	i just don't know which db that'd be in
20:00:58	mriedem	api?
20:01:00	mriedem	no,
20:01:02	mriedem	wrong schema
20:01:09	mriedem	i guess just iterate the cell dbs
20:01:40	mriedem	if you find a non-deleted nova-consoleauth service record in that db, but no console auth tokens in the db, and workarounds is false, then fail
20:02:56	melwitt	yeah, or warn like dansmith said. only matters if you're rolling
20:03:33	melwitt	i.e. it will only mess you up if you're rolling
20:04:43	mriedem	i left a comment on the bug with the status ugprade check idea
20:05:01	melwitt	thanks
20:05:29	openstackgerrit	Matthew Edmonds proposed openstack/nova master: Use tempfile for powervm config drive https://review.openstack.org/610174
20:11:27	mriedem	dansmith: np i got it, it was 1 line
20:11:39	openstackgerrit	Matt Riedemann proposed openstack/nova master: Fix formatting non-templated cell URLs with no config https://review.openstack.org/611094
20:12:27	mriedem	easy fix for another core ^
20:17:22	melwitt	+W
20:27:14	openstackgerrit	Matt Riedemann proposed openstack/nova master: Ignore uuid if already set in ComputeNode.update_from_virt_driver https://review.openstack.org/611162
20:27:15	mriedem	efried: jroll: ^
20:30:07	openstackgerrit	melanie witt proposed openstack/nova master: Bump os-brick version to 2.6.1 https://review.openstack.org/611109
20:30:52	efried	I don't get it.
20:32:25	openstackgerrit	Sundar Nadathur proposed openstack/nova-specs master: Nova Cyborg interaction specification. https://review.openstack.org/603955
20:33:58	spatel	Folks! i have 64G compute node
20:34:09	dansmith	spatel: it's not nice to brag
20:34:14	dansmith	mriedem: thanks home skillet
20:34:16	spatel	Should i go with 1G hugepage or 2M
20:34:46	spatel	dansmith: i was going to write question but hit enter middle of i
20:34:49	spatel	dansmith: i was going to write question but hit enter middle of it
20:34:55	dansmith	spatel: I know, I'm just joking :P
20:35:00	spatel	:)
20:35:49	spatel	what do you recommend if that is the case
20:36:08	spatel	Problem is if i launch application then it will be hard to adjust those value
20:36:36	spatel	currently i have "hugepagesz=2M hugepages=27000 transparent_hugepage=never"
20:36:50	dansmith	spatel: you probably want cfriesen
20:37:09	spatel	cfriesen: ^^
20:38:36	spatel	He may be not around
20:38:42	mriedem	efried: you were on the original regression patch of mine so figured you'd have context
20:38:51	mriedem	this https://review.openstack.org/#/c/571535/
20:38:56	efried	mriedem: Yeah, I think I get it now.
20:39:07	efried	See if my review comment makes sense.
20:40:12	mriedem	efried: yup
20:40:32	efried	mriedem: ...and another update
20:40:33	cfriesen	spatel: in our testing 2M gave a noticeable benefit. 1G gave some additional benefit but only for specific testcases
20:40:50	spatel	There you go!! thanks you
20:41:15	mriedem	efried: yup
20:41:21	efried	cool
20:41:21	cfriesen	spatel: are you using dedicated CPUs?
20:41:22	spatel	Do you think 27000 is good number on 64G compute node?
20:41:30	spatel	yes I am pinning CPU
20:41:56	spatel	i am going to run riak cluster application on this compute node
20:42:11	spatel	riak love memory
20:42:16	cfriesen	you don't need to allocate hugepages at boot. you can allocate them at runtime
20:42:33	spatel	i heard sometime it cause issue during runtime
20:42:58	cfriesen	spatel: If you allocate them early during startup the memory hasn't gotten fragmented yet
20:43:14	spatel	hmm!
20:43:39	spatel	i will keep that in mind then..
20:43:48	cfriesen	it's a bit of a tradeoff, since any memory you reserve for hugepages can't be allocated to small-page instances. Also, you need to keep some 4K memory around for the host itself.
20:43:59	spatel	but you have to reboot your flavor also right after adjust hugepage
20:44:35	cfriesen	not sure what you mean by "reboot your flavor" :)
20:44:39	spatel	I kept 8G memory for host that is why i pick 27000 pages
20:44:50	spatel	i meant reboot your VM
20:44:59	cfriesen	spatel: that should be fine as long as most of your guests are using hugepages
20:45:27	spatel	cfriesen: thanks! in that case i will go with 27000
20:46:31	cfriesen	spatel: changing a flavor and then rebooting your vm won't do anything. the flavor information was cached in the VM at creation time.
20:47:33	cfriesen	in the instance object, rather
20:47:38	spatel	oh!!
20:47:41	spatel	cool
20:56:01	mriedem	i still don't understand how this forbidden aggregate placement API thing is going to be used via nova for blazar which is use case in this spec https://review.openstack.org/#/c/603352/
20:56:19	efried	mriedem: You kind of had to be in the room, unfortunately.
20:56:21	mriedem	it seems we're gung ho about adding a placement api without any details on how to ues it
20:56:22	mriedem	*use
20:56:43	mriedem	well, the place to document that is in the spec for those not in the room right?
20:56:58	mriedem	if the answer is, "we're going to fork nova and make pre-request filters an extension point" then say that
20:57:00	efried	Yes, I agree. If it's not clear to someone who wasn't in the room, it needs a rewrite.

Earlier   Later