Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2017-08-08
11:17:12 gibi cdent: I have no good answer for that
11:17:15 cdent :)
11:17:58 cdent I’m currently looking at coverage results when running just functional/test_servers.py to see if that raises any alarms. But because I’m looking at master I now it is missing several of the things that are in progress.
11:20:50 cdent gibi: a lot of what is missing is related to custom resource classes, so your plans for that will be useful
11:23:31 openstackgerrit Alex Xu proposed openstack/nova master: placement: ensure sharing RPs maps combinates with correct shared RP https://review.openstack.org/480379
11:23:32 openstackgerrit Balazs Gibizer proposed openstack/nova master: replace chance with filter scheduler in func tests https://review.openstack.org/491529
11:23:52 alex_xu cdent: ^ remove the 'root', instead to use 'sharing' and 'shared'
11:24:02 gibi cdent: now I just have to find the time to write them :)
11:24:08 cdent thank you alex_xu
11:24:27 gibi bauzas: I removed the vcpu=2 from the SmallFakeDriver to see what fails
11:24:46 gibi bauzas: I think we have a problem with the default 16.0 allocation ration. I don't see that it is applied at all
11:24:58 alex_xu cdent: hope that works :)
11:26:26 cdent gibi: do you get reasonable results from placement, but then the fake driver refuses then? If so, it’s probably a bug in the driver itself. If you’re not getting results from placement then is the inventory being set properly?
11:26:26 gibi bauzas: here is an example test failure http://paste.openstack.org/show/617764/
11:27:39 gibi cdent: L66 worries me http://paste.openstack.org/show/617764/
11:28:15 cdent gibi: it’s max_unit,
11:28:21 cdent that’s the problem
11:28:31 cdent and is an actual problem:
11:29:10 cdent we set max_unit to be the number of real cpus because we don’t think any single consumer should occupy more than the number of cpus, whatever allocation ratio says
11:29:13 cdent which makes sense
11:29:31 cdent but when we create a doubling allocation for the resize to same host, we’re breaking that
11:30:13 gibi bauzas: ^^
11:30:17 cdent so as the code is currently designed, if using doubling, you can never resize to same host a guest with #vcpus == #pcpus
11:30:29 cdent the quick fix for the tests is to use a bigger driverr
11:30:44 gibi yes, that was what I proposed the bauzas suggested allocation ratio
11:30:54 gibi but then I guess we cannot use the allocation ratio trick here
11:31:02 cdent but we may need to consider that the doubling concept is problematic on small hosts…
11:31:14 cdent yeah, allocation ratio doesn’t do anything for max_unit
11:32:34 gibi OK I wait for bauzas to agree then I will push back the previos patch set of https://review.openstack.org/#/c/491529/ that has the SmallFakeDriver modification to 2 vcpus
11:33:07 cdent alex_xu: it isn’t quite right. I’m trying to come up with a suitable alternative. I’ll push something up if that’s okay with you?
11:33:53 alex_xu cdent: yea, appreciate the help, looks like I understand 'shared' and 'sharing' incorrectly
11:33:58 openstackgerrit Sean Dague proposed openstack/nova master: Fix all >= 2 hit 404s https://review.openstack.org/491761
11:34:27 cdent gibi: seems reasonable. I wonder who has an answer to the question about the legitimacy of resizing a large guest on a small host. I don’t really know. I wouldn’t want to do it, but I’m not paying for hardware...
11:35:44 gibi resize on same host feels like a feature to support testing resize on single node devstack
11:39:42 openstackgerrit Sean Dague proposed openstack/nova master: Create reference subpage https://review.openstack.org/490994
11:51:45 openstackgerrit Chris Dent proposed openstack/nova master: placement: ensure RP maps to those RPs that share with it https://review.openstack.org/480379
11:54:12 cdent alex_xu: ^ that may be a bit better. I ended up finding it difficult to make clear.
11:55:33 alex_xu cdent: thanks
11:58:22 openstackgerrit OpenStack Proposal Bot proposed openstack/nova master: Imported Translations from Zanata https://review.openstack.org/477091
12:01:24 openstackgerrit Balazs Gibizer proposed openstack/nova master: Raise NoValidHost if no allocation candidates https://review.openstack.org/491491
12:02:31 gibi bauzas: fixed your comment about the test ^^
12:12:33 bauzas cdent: looking at alex_xu change, thanks for it
12:13:32 bauzas gibi: cdent: looking at the above discussion, could you please summarize the problem with ratios ?
12:14:23 gibi bauzas: we set max_unit of the vcpu resource based on the number of pcpus therefore we cannot ask for more than pcpu amount of vcpu in a single allocation
12:14:39 alex_xu bauzas: thanks for the review
12:14:48 bauzas gibi: mmm, sadly then
12:21:17 gibi bauzas: so are you OK with the SmallFakeDriver vcpu=2 change?
12:21:24 bauzas gibi: looking
12:23:39 bauzas gibi: just reviewing alex_xu and then it's your turn :p
12:25:57 gibi bauzas: OK, thanks
12:28:53 openstackgerrit Alexandra Settle proposed openstack/nova master: Create reference subpage https://review.openstack.org/490994
12:29:21 bauzas cdent: gibi: so, question about max_unit
12:29:47 bauzas cdent: gibi: when we have a flavor asking for 10 CPUs, do you think it should be 10 pCPUs or 10 vCPUs ?
12:30:01 bauzas cdent: gibi: MHO is that's for virtual ones
12:30:02 asettle Ah shite, sdague and stephenfin - I got ahead of myself and updated the wrong nova patch (didn't tab to the right PR). I just updated this patch: https://review.openstack.org/#/c/490994/ with the change from Administrators to Adminstration Guide
12:30:04 asettle I can revert if you'd like?
12:30:14 cdent bauzas: it _is_ for vcpus
12:30:18 asettle Otherwise, I was just going to make the same change on stephenfin 's patch: https://review.openstack.org/#/c/490952/
12:30:21 bauzas cdent: I know
12:30:30 cdent and it should be
12:30:41 bauzas cdent: gibi: but the point is then, why max_unit should be about *physical* resources ?
12:31:56 cdent max_unit, in general, is abstract. where we make concrete assertions about its meaning is when we set inventory from the compute manager. In there we make the assertion that it is bad for a service to host a _single_ vm that has more vcpus than there are pcpus
12:32:34 cdent when we discussed this before, several people mentioned that doing so would cause pathalogical context switching
12:32:44 bauzas cdent: tbc, flavor * ratio should be <= max_unit
12:33:16 cdent ratio has _nothing_ to do with max unit. ratio is a measurement for all all guests, max_unit is for one
12:33:16 bauzas not sure I understand the problem about max_unit being related to physical resources :(
12:33:37 bauzas cdent: sure, I know
12:34:09 cdent bauzas: okay, so I dont understand what you’re asking or suggesting then?
12:34:16 sdague asettle: don't work, I've got a patch sitting on top of it that I'm about to push and it should zero it back
12:34:31 asettle sdague: dont work?
12:34:37 cdent asettle: woot. you never need to work again!
12:34:44 asettle cdent: I have seriously been waiting for this day for ages.
12:34:49 bauzas cdent: the problem is that max_unit is blocking us for https://review.openstack.org/#/c/491529/4/nova/virt/fake.py@599
12:34:57 sdague asettle: don't worry
12:35:14 cdent bauzas: yes, in a way we want and expect it to. so either we change the fake driver or use a different one.
12:35:14 asettle sdague: hahaha okay :)
12:35:30 asettle I will go make my two second patch on stephenfin 's then
12:35:52 bauzas cdent: I'm unclear, why should we need to modify the resources if we have a ratio around 16.0 ?
12:36:08 bauzas cdent: because of max_unit, because max_unit == pCPUs, right?
12:36:30 bauzas that ^ looks weird to me, and changing the behaviour we had for 5 years
12:36:37 cdent because we are creating an allocation, when we do the doubling, for a single consumer that is greater than max_unit
12:36:49 cdent the change in behavior is the doubling
12:36:53 cdent that’s what’s new
12:37:01 cdent the max_unit behavior has been around for a while
12:37:13 bauzas imagine a world where we would double the allocation, but have the previous behaviour
12:37:23 cdent if we were able to do the doubling with two different consumer uuids, we wouldnt have this problem
12:37:45 bauzas say I have a single pCPU with 16.0 ratio, I can still ask for 2 flavors of 1vCPU, right?
12:37:53 cdent bauzas: yes
12:38:05 bauzas so, why can't we do that now ?
12:38:06 cdent but that’s not what is being asked for here
12:38:21 cdent we are literally asking for a single consumption with 2vcpu
12:38:22 jaypipes morning folks
12:38:50 bauzas cdent: if we have a ratio of 16.0, I don't think it's a problem
12:39:04 cdent morning jaypipes: i hope your potty emergency is cleared up
12:39:15 openstackgerrit Sean Dague proposed openstack/nova master: Create For End Users index section https://review.openstack.org/491785
12:39:17 bauzas cdent: in a previous world, I could ask for one flavor of 2vCPUs with a single compute having 1pCPU
12:39:18 jaypipes dansmith, cdent, bauzas: just FYI, I'm down to two test failures left on the confirm/resize patch series...
12:39:27 jaypipes cdent: seems to be, yes.
12:39:33 sdague asettle: ah, I guess it doesn't overwrite it, however it should be fine
12:39:46 asettle sdague: okay :) that's not so bad. Just a wording nit anyway.

Earlier   Later