| Posted | Nick | Remark | |
|---|---|---|---|
| #openstack-nova - 2018-05-17 | |||
| 13:43:50 | efried | ? | |
| 13:43:50 | efried | compute_provider_additional_traits = CUSTOM_FOO, CUSTOM_BAR | |
| 13:43:50 | efried | belmoreira: You're suggesting a config option like | |
| 13:44:06 | mriedem | yes, and i don't think we should do that | |
| 13:44:07 | belmoreira | efried yes | |
| 13:44:10 | dansmith | you'd have to have a negative as well so you could ever remove traits you had previously set | |
| 13:44:13 | mriedem | it makes nova a proxy, | |
| 13:44:15 | dansmith | mriedem: agree | |
| 13:44:23 | mriedem | and blocks everyone on nova supporting this thing | |
| 13:44:28 | mriedem | when you can already do it external to nova | |
| 13:44:34 | efried | belmoreira: yes, I agree with mriedem and dansmith. | |
| 13:44:47 | mriedem | i already don't really care for the ironic trait proxy stuff we do | |
| 13:45:02 | belmoreira | my issue is how to manage all this things at scale | |
| 13:45:28 | efried | belmoreira: You would be editing config files, one per host, in any case. | |
| 13:45:40 | mriedem | your config mgmt tool is already doing the logic to put the custom traits in the specific compute host config file yes? | |
| 13:45:41 | efried | belmoreira: Rather than doing that, run a placement CLI once per host. | |
| 13:45:47 | belmoreira | a configuration management tool doing API calls doesn't looks good to me | |
| 13:45:56 | mriedem | why can't that same code just put the traits in placement once nova-compute is running and the rp is registered? | |
| 13:46:15 | mriedem | belmoreira: how do you create keystone endpoints? | |
| 13:46:25 | mriedem | and users/projects | |
| 13:46:38 | efried | Are you equating "API calls" with invoking CLIs? | |
| 13:47:07 | mriedem | openstack-ansible does all of this quite a bit i think | |
| 13:47:13 | mriedem | that's how they do the cells v2 setup | |
| 13:47:24 | mriedem | it's really the only way since we don't have rest apis for cell and host mappings | |
| 13:47:55 | belmoreira | mriedem is true for keystone. We built our own machinery to create/delete projects. And the experience is not great | |
| 13:49:36 | belmoreira | efried: no. The problem is given credentials to a configuration management tool to populate these properties | |
| 13:50:19 | belmoreira | changing configuration files is much easier and less worrying | |
| 13:54:20 | openstackgerrit | Matt Riedemann proposed openstack/osc-placement master: Fix error message asserts in functional test https://review.openstack.org/569048 | |
| 13:54:21 | openstackgerrit | Matt Riedemann proposed openstack/osc-placement master: Resource class set (v1.7) https://review.openstack.org/514644 | |
| 13:54:22 | openstackgerrit | Matt Riedemann proposed openstack/osc-placement master: Usages per project and user (v1.8, v1.9) https://review.openstack.org/514646 | |
| 13:54:23 | openstackgerrit | Matt Riedemann proposed openstack/osc-placement master: CLI allocation candidates (v1.10) https://review.openstack.org/514647 | |
| 13:54:24 | openstackgerrit | Matt Riedemann proposed openstack/osc-placement master: New dict format of allocations (v1.11, v1.12) https://review.openstack.org/542819 | |
| 13:54:25 | openstackgerrit | Matt Riedemann proposed openstack/osc-placement master: Transactionally update allocations (v1.13) https://review.openstack.org/546674 | |
| 13:54:26 | openstackgerrit | Matt Riedemann proposed openstack/osc-placement master: Add nested resource providers (v1.14) https://review.openstack.org/546675 | |
| 13:54:27 | openstackgerrit | Matt Riedemann proposed openstack/osc-placement master: Limit allocation candidates (v1.15, v1.16) https://review.openstack.org/548043 | |
| 13:54:28 | openstackgerrit | Matt Riedemann proposed openstack/osc-placement master: Allocation candidates parameter: required (v1.17) https://review.openstack.org/548326 | |
| 13:54:53 | mriedem | gibi: https://review.openstack.org/#/c/514644/ should be ready to go now | |
| 13:55:19 | openstackgerrit | Yikun Jiang (Kero) proposed openstack/nova master: Add policy to InstanceGroup object and api models. https://review.openstack.org/563375 | |
| 13:55:20 | openstackgerrit | Yikun Jiang (Kero) proposed openstack/nova master: Add policy field to ServerGroup notification object https://review.openstack.org/563401 | |
| 13:57:08 | gibi | mriedem: done | |
| 13:57:18 | mriedem | yay | |
| 13:57:58 | melwitt | nova meeting in 3 min | |
| 14:09:17 | openstackgerrit | Sylvain Bauza proposed openstack/nova-specs master: Proposes Multiple GPU types https://review.openstack.org/557065 | |
| 14:17:13 | mriedem | tssurya: check the max_attempts config option, if it's 1 then we shouldn't be calculating alternate hosts in the scheduler | |
| 14:17:23 | mriedem | and belmorei_ said in boston that cern doesn't do retries | |
| 14:17:41 | dansmith | mriedem: well, let me just hand you my thunder then | |
| 14:18:28 | mriedem | tssurya: dansmith reminded me that max_attempts=1 means no retries and dansmith reminded me that belmorei_ said in boston that cern doesn't do retries | |
| 14:18:40 | dansmith | haha | |
| 14:20:44 | belmorei_ | mriedem: correct :) we have max_attempts=1 | |
| 14:21:31 | dansmith | so if it's trying to get alternates in that case, it's probably a bug | |
| 14:22:28 | mriedem | yeah need to figure that part out, although i thought edleafe added unit tests for that scenario to assert that we didn't try calculating alternates if max_attempest=1 | |
| 14:22:36 | mriedem | tempest on the brain | |
| 14:23:01 | dansmith | ah, | |
| 14:23:08 | dansmith | max_attempts grabs one alternate it looks like | |
| 14:23:15 | dansmith | instead of max_attempts-1 ? | |
| 14:24:06 | dansmith | https://github.com/openstack/nova/blob/master/nova/scheduler/filter_scheduler.py#L161-L161 | |
| 14:24:16 | edleafe | num_alts = CONF.scheduler.max_attempts if return_alternates else 0 | |
| 14:24:36 | edleafe | so yeah, it will return 1 alternate with max_attempts=1 | |
| 14:24:52 | dansmith | which is wrong | |
| 14:25:20 | mriedem | arvindn05: how are you coming on that traits-in-images + rebuild patch? | |
| 14:25:38 | dansmith | we also would do too much extra work if num_alts is zero | |
| 14:26:06 | dansmith | belmorei_: tssurya: if you can confirm that with max_attempts the scheduler is asking for one extra alternate that would be good | |
| 14:26:29 | dansmith | would mean about an exact doubling of placement and filter load I guess | |
| 14:26:42 | mriedem | shouldn't be any more placement load, | |
| 14:26:44 | mriedem | just filtering | |
| 14:26:58 | mriedem | the alternates are pulled from the compute node list we get after placement allocation candidates are retrieved | |
| 14:26:59 | dansmith | it asks placement again doesn't it? | |
| 14:27:07 | mriedem | not for new allocation candidates | |
| 14:27:09 | edleafe | alternates only get re-filtered when building > 1 instance per request | |
| 14:27:10 | mriedem | or it shouldn't... | |
| 14:27:19 | belmorei_ | dansmith: will do | |
| 14:27:23 | dansmith | mriedem: okay | |
| 14:29:09 | mriedem | there are some unit tests setting max_attempts to 1 but i'm having a hard time seeing what they assert | |
| 14:29:28 | mriedem | probably just need a simple tests where max_attempts=1, single instance in the request, and make sure we don't mess with alternate calculation | |
| 14:29:34 | mriedem | *test | |
| 14:29:42 | openstackgerrit | Stephen Finucane proposed openstack/nova-specs master: Rename spec to match blueprint https://review.openstack.org/568870 | |
| 14:29:43 | openstackgerrit | Dan Smith proposed openstack/nova master: WIP: Fix interpretation of max_attempts for scheduling alternates https://review.openstack.org/569127 | |
| 14:29:48 | mriedem | dansmith: want to keep this backport train a comin https://review.openstack.org/#/c/568369/ | |
| 14:29:51 | dansmith | belmorei_: tssurya ^ is what I expect we'll do | |
| 14:30:00 | dansmith | *we'll need | |
| 14:30:05 | belmorei_ | this is an interesting issue. We get in average 20s for scheduling... We are debugging and let you know when we have more info | |
| 14:30:27 | dansmith | 20s to schedule one? that's.. uh, terrible | |
| 14:33:49 | melwitt | yeah ... | |
| 14:37:01 | openstackgerrit | Eric Fried proposed openstack/nova master: libvirt: get_inventory => update_provider_tree https://review.openstack.org/560444 | |
| 14:37:02 | openstackgerrit | Eric Fried proposed openstack/nova master: libvirt: Don't report DISK_GB if sharing https://review.openstack.org/560459 | |
| 14:37:03 | openstackgerrit | Eric Fried proposed openstack/nova master: WIP: Normalize inventory from update_provider_tree https://review.openstack.org/569132 | |
| 14:37:18 | efried | bhagyashris_: ^ | |
| 14:42:49 | dansmith | belmorei_: tssurya: something else to maybe look at.. by default we ask for 1000 placement results on each schedule, which is probably insane | |
| 14:43:04 | dansmith | you might try tweaking down max_placement_results to see if that makes a difference | |
| 14:43:12 | dansmith | depending on where you see the hotspot is | |
| 14:43:48 | belmorei_ | in our case that shouldn't happen because we use the request_filter | |
| 14:44:10 | dansmith | what do you mean? | |
| 14:44:20 | dansmith | you mean because you're limiting to a cell with only 200 hosts? | |
| 14:44:32 | belmorei_ | the request_filter to use placement_aggregates | |
| 14:44:42 | dansmith | yeah, okay | |
| 14:45:04 | dansmith | well, still, on an empty cell you'd get back the whole cell, but fair enough | |
| 14:45:10 | mriedem | are the compute node resource providers in placement within aggregates? | |
| 14:45:24 | mriedem | in other words, are you getting back 200 allocation candidates or 10K? | |
| 14:45:24 | dansmith | mriedem: I think that's the implication :) | |
| 14:45:37 | mriedem | sure but have they asserted that's all linked up properly | |