Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2017-08-02
16:23:59 mriedem i'll throw up the wip shortly
16:24:00 dansmith I'm really kinda confused about this doubling anyway
16:25:17 dansmith I guess it's the attempt to account for shared storage that makes this complicated,
16:25:22 dansmith and which avoids doubling for same-host
16:25:30 cdent dansmith: yes
16:26:03 mriedem yup
16:26:38 mriedem # already allocated against on the source host (like shared storage
16:26:38 mriedem "Remove any allocations against resource providers that are
16:26:39 mriedem # providers)"
16:26:49 mriedem so i guess the intention was to specifically not double up shared storage
16:26:51 mriedem on a mov
16:26:52 mriedem *move
16:26:55 dansmith which is wrong anyway
16:27:00 dansmith for certain types of shared storage
16:27:11 mriedem seemed like the right idea at the time?!
16:27:13 dansmith it's not wrong for a volume, but is wrong for a compute node using ceph
16:27:21 mriedem which was 72 hours ago?
16:27:52 dansmith I know, looking at this, I swear I've never seen it before, but I'm pretty sure I +2d it not long ago
16:28:31 mriedem ha yeah same here
16:28:54 mriedem https://review.openstack.org/#/c/487589/
16:29:01 openstackgerrit Ildiko Vancsa proposed openstack/nova master: Implement new attach Cinder flow https://review.openstack.org/330285
16:30:58 mriedem dansmith: melwitt: skip the cells v2 meeting yeah?
16:31:55 dansmith oh heh, I meant to say I have a conflict today anyway
16:31:56 dansmith so yeah
16:39:32 openstackgerrit Ildiko Vancsa proposed openstack/nova master: Implement new attach Cinder flow https://review.openstack.org/330285
16:42:31 cdent dansmith: in honor of your pain, I created a bug, which we can decided to care about or not: https://bugs.launchpad.net/nova/+bug/1708205
16:42:31 openstack Launchpad bug 1708205 in OpenStack Compute (nova) "placement allocation representation asymetric on PUT and GET" [Low,Confirmed]
16:42:40 dansmith oh I care
16:42:42 dansmith I care bigly
16:42:50 cdent tremendous
16:47:03 mriedem ok got the patch
16:47:08 mriedem pushing soon
16:47:10 mriedem prepare
16:50:49 openstackgerrit Matt Riedemann proposed openstack/nova master: WIP: Sum allocations in the scheduler when resizing to the same host https://review.openstack.org/490085
16:50:52 mriedem dansmith: cdent: ^ it's not the prettiest, and it doesn't account for a case that we'd have shared storage
16:51:09 jangutter I've got a newbie question here regarding attaching/detaching SR-IOV ports in Nova. I presume it's not going to be "just something simple" and would need work in Queens at the very least? (ref: https://review.openstack.org/#/c/139910/ )
16:51:26 openstackgerrit Chris Dent proposed openstack/nova master: [placement] Avoid error log on 405 response https://review.openstack.org/490021
16:52:17 dansmith ack I will build on it when I'm done with my thing
16:54:07 cdent mriedem: If were on a non-shared some host resize, there will only be one allocation, so perhaps we can avoid some of the looping, which makes my brain explore?
16:54:14 cdent s/some/same/
16:54:28 mriedem dansmith: ok, i have a todo question in the test - which is i'm not sure if we even need the compute to adjust things, e.g. if the new alloc for vcpu is smaller than the current alloc, wouldn't we just not sum those? in other words, shouldn't the new allocations when we're done be the max of the current + new?
16:54:41 jangutter Has someone taken a look at attaching/detaching SR-IOV ports recently? If not, I can try to take a stab at it.
16:54:49 mriedem jangutter: no
16:54:53 openstackgerrit Stephen Finucane proposed openstack/nova master: doc: Rework README to reflect new doc URLs https://review.openstack.org/480074
16:54:53 openstackgerrit Stephen Finucane proposed openstack/nova master: doc: Start using oslo_config.sphinxext https://review.openstack.org/482961
16:54:54 openstackgerrit Stephen Finucane proposed openstack/nova master: policies: Fix Sphinx issues https://review.openstack.org/480516
16:54:54 openstackgerrit Stephen Finucane proposed openstack/nova master: doc: Start using oslo_policy.sphinxext https://review.openstack.org/479358
16:54:54 cdent mriedem: for the duration of the resize we need room for both vms, don’t we?
16:54:55 openstackgerrit Stephen Finucane proposed openstack/nova master: doc: Rework index page per new sections https://review.openstack.org/478485
16:54:56 openstackgerrit Stephen Finucane proposed openstack/nova master: doc: Remove dead files https://review.openstack.org/478470
16:54:56 openstackgerrit Stephen Finucane proposed openstack/nova master: doc: Rewrite index page https://review.openstack.org/490088
16:55:05 dansmith mriedem: I don't think we should be that discriminating
16:55:06 mriedem cdent: even if it's on the same host?
16:55:18 dansmith cdent: his point is that vcpus are somewhat ephemeral, not a strict quantity of things that go away
16:55:23 dansmith except for pinning, but...
16:55:24 mriedem cdent: yes i tried a few ways to do this w/o the looping and stuff and it hurt my brain
16:55:37 dansmith I think the scheduler should stay out of deciding what it does and doesn't know about the hypervisor
16:55:44 mriedem if i'm resizing from 1 cpu to 2 cpu on the same host, my total allocation on that host should be 2, not 3
16:56:20 mriedem yeah it's fair to say we don't want to bake all that logic into the scheduler
16:56:33 mriedem just seems a bit wasteful
16:56:39 mriedem until you confirm and we fix thigns
16:56:41 cdent and for vmware and powervm and maybe others, the resize may not be on the same host, just the same n-cpu?
16:56:54 mriedem powervm should be fine
16:56:56 dansmith cdent: we're accounting based on node though
16:56:58 mriedem vmware is the freak
16:57:12 dansmith still,
16:57:31 mriedem but yeah it's per node
16:57:33 dansmith for a hypervisor with very concrete resources, we don't know that some resources add and others don't
16:57:36 dansmith and,
16:57:48 dansmith we're not using double the disk in most cases,
16:57:53 openstackgerrit Merged openstack/nova master: Accept any scheduler driver entrypoint https://review.openstack.org/484828
16:57:55 dansmith but it depends on the config of the node,
16:58:07 dansmith so we should just grab double anyway for everything, IMHO
16:58:19 mriedem bauzas: are you going to backport https://review.openstack.org/#/c/484828/ to ocata or should i?
16:58:39 bauzas mriedem: I can
16:58:42 mriedem dansmith: ok - my main concern was that ocata computes aren't going to fix the double allocation to remove the old flavor
16:58:54 dansmith mriedem: they're just going to trample all over it anyway
16:59:01 mriedem that's true
16:59:06 mriedem gd ocata computes
16:59:11 mriedem they're aholes
16:59:16 bauzas mriedem: looks like I need to resolve a conflict but I'll do
16:59:20 dansmith I know.. greedy punks
16:59:32 mriedem i will commence to lunching then
16:59:33 cdent i first read that as octopus computes
17:10:15 melwitt mriedem, dansmith: was on a call, ack to skipping cells meeting
17:16:37 openstackgerrit Steve Noyes proposed openstack/nova master: Do not attempt volume swap when guest is stopped/suspended https://review.openstack.org/389798
17:21:54 openstackgerrit Steve Noyes proposed openstack/nova master: Do not attempt volume swap when guest is stopped/suspended https://review.openstack.org/389798
17:24:46 dansmith mriedem: so that patch is not on top of jay's, nor under gibi's
17:26:36 dansmith I'm thinking I will put this on top of gibi's
17:34:56 openstackgerrit Merged openstack/nova master: placement: remove existing allocs when set allocs https://review.openstack.org/489273
17:36:47 dansmith heh, found another problem
17:37:17 dansmith at least with fake virt, we're reporting min_unit=1, max_unit=1, step_size=1, which means we can't claim 2 vcpus for the single instance
17:37:23 dansmith which could be a problem in real life,
17:37:40 dansmith if configured so that one instance can't take more than 25% of a host, but a resize would technically take 50% in a single allocation
17:37:49 dansmith which would be solved by my migration uuid thing
17:40:26 bauzas dansmith: mriedem: just saw the above discussion about whether the scheduler should know the virt logic and the move ops, tbc MHO is * NOOOOOOO *
17:40:38 bauzas because we have conductors for that
17:40:54 bauzas not for virt stuff, but at least knowing whether it's a move or a boot
17:41:13 bauzas scheduler should just give you a destination, whether it's for a move or a boot, that's it
17:42:09 bauzas if we need more reasons why we need to do that, I don't want to hit your wounts, but that's why we first tried to claim by the conductors...

Earlier   Later