Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2017-08-09
15:23:23 mnaser oh boy
15:23:37 mnaser did i just forget how to math
15:23:50 mnaser and the fact it had 29 cores probably messed up the math
15:24:06 danpb sibling_sets gets populated from available_siblings iiuc
15:24:15 danpb so presumably available_siblings is empty too ?
15:24:42 mnaser danpb: let me put some log.debug's, but i confirmed that siblings_set was empty
15:24:53 mnaser fyi this issue only occurs on the *last* VM to fit on the compute node
15:26:12 danpb well i think you'd need to work backwards through the call stack to figure out why it becomes empty
15:26:17 mnaser danpb: AVAILABLE_SIBLINGS: [CoercedSet([]), CoercedSet([]), CoercedSet([]), CoercedSet([]), CoercedSet([]), CoercedSet([]), CoercedSet([])]
15:27:19 mnaser ok i guess ill have to see why host_cell.free_siblings is being set to that
15:27:30 dansmith danpb: doesn't this come from libvirt?
15:28:12 mnaser the only thing i can imagine which can cause a corner case is the fact that we reserve 2 cores for the OS, so vcpu_pin_set=2-31 .. maybe that's not taken in consideration (guessing)
15:28:17 openstackgerrit Matt Riedemann proposed openstack/nova master: [placement] Add api-ref for usages https://review.openstack.org/480563
15:29:51 danpb dansmith: libvirt will provide info on the siblings present on the host, but iiuc free_siblings is populated by nova
15:30:17 danpb so if mnaser is saying it works correctly for all VMs until the last one, it sounds like nova is filtering the info from libvirt and ending up with the empty set
15:30:24 dansmith danpb: oh okay, I hadn't traced it very far up the stack because I assumed we must just be getting an empty set of things from libvirt because we had no numa info or something
15:30:32 dansmith danpb: yeah, makes sense
15:30:35 sahid danpb: mnaser danpb i just for information i was working on this but did not find the root cause
15:30:37 sahid https://review.openstack.org/#/c/458848/
15:30:42 mnaser im checking the values of host_cell and instance_cell
15:31:15 mnaser HOST_CELL: NUMACell(cpu_usage=14,cpuset=set([2,4,6,8,10,12,14,16,18,20,22,24,26,28,30]),id=0,memory=196562,memory_usage=57344,mempages=[NUMAPagesTopology,NUMAPagesTopology],pinned_cpus=set([2,4,6,8,10,12,14,18,20,22,24,26,28,30]),siblings=[set([8,24]),set([2,18]),set([10,26]),set([12,28]),set([6,22]),set([14,30]),set([4,20])])
15:31:36 danpb how many VMs are you running ?
15:32:09 danpb you've got 7 pairs of siblings there, so if each VM wanted one pair, you'd be able to run 7 VMs
15:32:42 mnaser booted with 120x 1GB hugepages, 30 cores available, trying to boot 15 VMs with 2 cores each + 8gb of memory each
15:32:43 mnaser but im not using isolate, im using prefer which i believe should try to schedule them on the same hyperthread
15:33:17 mnaser my flavor has properties: hw:cpu_policy=dedicated, hw:cpu_thread_policy=prefer, hw:mem_page_size=1048576, hw:numa_nodes=2
15:33:58 mnaser (also hw:cpu_thread_policy=prefer forgot to put that in)
15:34:23 mnaser i was thinking hw:numa_nodes=2 and hw:cpu_thread_policy=prefer might be the source of the issue because they are a bit the opposite of each other
15:35:48 mnaser getting memory split from 2 numa nodes but also wanting to be on the same hyperthread isnt really possible considering you'll want to be on two physical cpus to access the numa node
15:36:07 openstackgerrit Matt Riedemann proposed openstack/nova master: [placement] Add api-ref for RP usages https://review.openstack.org/450105
15:37:50 danpb mnaser: how many vcpus per guest ?
15:38:00 danpb (i mean what is vcpu count in the flavour ?)
15:38:04 mnaser danpb: 2
15:38:18 danpb that flavour config looks flawed then
15:38:33 danpb you've got 2 vcpus, and you've said to give the VM 2 virtual numa nodes
15:38:55 danpb nova will want to place each virtual numa node on a separate host numa node
15:39:03 danpb so each vcpu will need to be on a separate socket
15:39:09 openstackgerrit Matthew Edmonds proposed openstack/nova master: use conf for keystone session creation https://review.openstack.org/485121
15:39:16 danpb so asking nova to put the vcpus in the same hyperthread sibling is nonsensical
15:39:29 danpb as you can't have siblings if you have separate sockets
15:39:43 mnaser that is what i kinda theorized a bit
15:39:54 mnaser so will have to fallback to 1 numanode default
15:40:08 danpb nova shouldn't be trying to use hypthread siblings at all in ths case
15:40:43 mnaser that was my assumption that what it would do, thats why i put 'prefer' in there because i figured that was the more reasonable choice
15:40:46 cfriesen with hw:cpu_thread_policy=prefer it should just give separate pCPUs. if you had hw:cpu_thread_policy=require I think it'd fail
15:40:59 danpb putting in prefer should be an no-op in this case
15:41:14 danpb as the need to pick separate sockets should have taken priority in nova to satisfy the numa constraint
15:41:28 danpb and if you have thread_policy=require, then nova ought to raise a fatal error
15:41:59 cfriesen hw:cpu_thread_policy=prefer is always a no-op, technically. it's the default behaviour
15:42:43 mnaser right: so in that case, the config seems to be ok (yes, the prefer is noop/will never work) but it should still work
15:42:52 mnaser numastat -m reports 4096 free hugepages in each node (for a total of 8192)
15:45:23 cfriesen too bad sfinucan is on holidays. :)
15:45:59 mnaser the used cores are: 2-15,18-31
15:46:12 mnaser which leaves 16/17 free which are on two different sockets
15:46:43 mnaser so technically it should schedule on 16,17 which will be able to get 4096MB from each numa socket
15:47:53 cfriesen mnaser: which version of the code is this?
15:47:58 mnaser newton
15:48:09 mnaser python-nova-14.0.7-1.el7.noarch
15:48:10 mnaser more specifically
15:49:47 mnaser sahid i looked at your change however it seems siblings_set being empty is a possiblity :(
15:50:03 cfriesen dansmith: for some reason I thought that pCPUs without a sibling would still be in sibling_set, just as a single-item set
15:50:46 efried sdague https://github.com/openstack/nova/blob/master/nova/cmd/status.py#L198 Am I missing something, or is this untrue?
15:51:11 hongbin oomichi: hi, my team is working on deciding the default api version, a team member asked me to check with you since you did the api version work in nova, i mainly wanted to know how nova maintain the default api version (pick the latest version or a stable version? bump the default version everytime a new version is introduced? etc.) Are you the right person to ask?
15:51:28 mnaser does host_cell get assigned by the scheduler?
15:52:44 cfriesen mnaser: danpb: dansmith: this looks sort of related, but the fix should be in newton: https://bugs.launchpad.net/nova/+bug/1578155
15:52:45 openstack Launchpad bug 1578155 in OpenStack Compute (nova) newton "'hw:cpu_thread_policy=prefer' misbehaviour" [Medium,Fix committed] - Assigned to Stephen Finucane (stephenfinucane)
15:52:57 dansmith I really don't know anything about this stuff
15:54:53 jaypipes mriedem, dansmith: I will be afk for 2 hours, just FYI.
15:54:55 cfriesen mnaser: so are 16/17 the siblings of 0/1 which are not included in vcpu_pin_set?
15:55:08 openstackgerrit Matt Riedemann proposed openstack/nova master: [placement] Add api-ref for usages https://review.openstack.org/480563
15:55:11 artom About the fix, at any rate.
15:55:13 mnaser cfriesen nope, 0/1 are the ones which are not included
15:55:26 mnaser so our vcpu_pin_set is 2-31
15:55:30 mnaser (32 core machine)
15:55:33 cfriesen mnaser: yes, but which are the HT siblings of 0/1?
15:56:15 danpb hmm, yes, it seems numa node 0 has even numbered cpus, and node 1 has odd numbered cpus
15:56:19 mnaser cfriesen sorry, not sure i follow, but here's lscpu output if that helps? http://paste.openstack.org/show/617955/
15:56:34 danpb so excluding 0-1, kills 1 cpu each host numa node, instead of 2 cpus
15:56:35 cfriesen mnaser: based on the pattern above, the HT siblings of 0/1 would be 16/17....I'm wondering if that's confusing the logic in hardware.py
15:56:42 mnaser but i think that they are
15:57:19 cfriesen mnaswer: running "virsh capabilities" will show sibling info
15:57:26 danpb so that in turn kills a entire sibling pair from each numa node
15:57:37 danpb effectively making 4 cpus unavailable
15:57:53 cfriesen danpb: theoretically it shouldn't kill the sibling pair though...we should have two pCPUs with no siblings
15:58:03 mnaser here is output of virsh caps. http://paste.openstack.org/show/617956/
15:58:06 cfriesen but maybe the logic can't handle a mix of sibs and no sibs
15:58:15 mriedem someone want to approve this? https://review.openstack.org/#/c/491855/
15:58:18 danpb yeah that wouldn't be surprising
15:58:46 mnaser so am i better off reserving 2 siblings and maybe that might get rid of the 'confusion' ?
15:58:53 cfriesen mnaswer: seems likely
15:58:55 mnaser so instead of 0,1 .. do 0,16 instead?
15:59:02 danpb yeah
15:59:53 mnaser let me test that theory out
16:00:03 openstackgerrit Ed Leafe proposed openstack/nova master: Handle addition of new nodes/instances in ironic flavor migration https://review.openstack.org/487954
16:00:12 edleafe dansmith: dtantsur: ^^ updated to remove the node RC change handling
16:00:37 dtantsur awesome! I'm just out of a meeting, will do the ironic part now
16:00:50 mnaser so my vcpu_pin_set should technically be 1-15,17-31 in that case, correct?
16:00:53 mnaser leaving 0,16 for the OS
16:00:53 mriedem gibi: wasn't that also a problem in ocata then? CoreFilter isn't enabled by default
16:00:59 cfriesen mnaswer: either way, please raise a bug for this issue, it'd be nice to handle it properly

Earlier   Later