Earlier  
Posted Nick Remark
#openstack-nova - 2018-02-06
22:13:27 edleafe efried: according to the people who seemed to know more about it than I do, a machine with NUMA can be configured in a variety of ways.
22:14:05 edleafe efried: so if all the resources were provided by a plain compute node, wouldn't that also get returned?
22:14:21 sean-k-mooney edleafe: yes it they can. i know more about this topic then is good for my health
22:15:04 sean-k-mooney edleafe: if that compute node has only one memoy controler then by definiton all resocues on that node belong to the same numa node
22:15:07 efried edleafe: Okay, that's a fair point. Which we could resolve, as you say, by tagging the NUMA RPs with a NUMA trait.
22:15:20 edleafe ah good. So efried, just direct your questions to sean-k-mooney
22:15:32 sean-k-mooney so all compute nodes always have at least 1 numa node
22:16:29 efried Okay, so the only way to get non
22:16:47 efried non-NUMA-ness is to go to a system that has more than one NUMA node and get resources spread out across them.
22:17:05 efried In which case we *don't* need the trait.
22:17:11 sean-k-mooney efried: and if you incldue a distance to restict it
22:18:07 sean-k-mooney if you have multiple numa node and put no constraitn on the distance between resocues in the tree then you get resouce from different numa nodes
22:18:17 sean-k-mooney your performace will be less but it will still work
22:18:37 sean-k-mooney and you never asked for any afintiy so it fine
22:21:11 efried Okay, so this is a workable thing for NUMA affinity. I don't think the same mechanism can be made to work for affinity in general, or for anti-affinity.
22:21:55 openstackgerrit Matt Riedemann proposed openstack/nova master: Move the nova-next job in-tree and update it https://review.openstack.org/541474
22:22:03 sean-k-mooney efried: if you dont have to deal with aggreates then i think it will work not sure how distance is compute for aggregates however.
22:22:40 efried sean-k-mooney: I'm talking about using one aggregate per NUMA node (so nothing to do with distance).
22:23:03 sean-k-mooney when modeling afinity via distance affinty becoures deistnace < X and anti afinity is distance > X
22:23:48 sean-k-mooney efried: oh then ya if numa nodes are aggrages then that is different
22:23:50 efried Yeah, I get that, but *modeling* and *expressing* distance seems really complicated for people to understand.
22:24:10 cfriesen efried: one aggregate per numa node? would you then associate a pool of RAM, CPUs, PCI devices, etc, with that aggregate?
22:24:48 cfriesen at probably 2-4 numa nodes per compute node that seems like a lot of aggregates
22:24:55 sean-k-mooney you would make the cpus/ram/pci devices resouce providers members of the aggregate
22:25:03 efried cfriesen: You associate the resource *providers* with the aggregate. So whatever providers provide the resources associated with your NUMA node.
22:25:06 efried yeah, what sean said.
22:25:18 efried No, it would be one aggregate per NUMA node, period.
22:25:39 sean-k-mooney efried: not one aggregate per numa node per host?
22:25:51 efried sean-k-mooney: Yes.
22:26:02 efried One aggregate per NUMA node, in the global sense :)
22:26:24 sean-k-mooney oh ame that aggreate will be am large
22:26:35 efried Is that "a lot" of aggregates? Meh. Aggregates are cheap.
22:27:16 efried I feel like we're not talking in the same terms. Here: If you have 10 hosts, and each host has 2 NUMA nodes, you have 20 aggregates total.
22:27:18 sean-k-mooney i think i would prefer lots of small aggregate then a few giant cluster wide ones
22:27:58 sean-k-mooney efried: ah ok that what i was thinking we should have if we used aggregates so im cool with that
22:29:04 sean-k-mooney we dont have traits on aggregates but at least for numa i dont think we need them on teh aggreages so it could work
22:29:21 cfriesen so you'd have these per-numa-node aggregates for system use, then per-host aggregates to match against flavor/image properties and act as availability zones?
22:29:44 efried sean-k-mooney: We can always stamp the involved providers with traits. But not sure we need 'em. At least to mark NUMA nodes.
22:30:02 efried cfriesen: I know nothing about host aggregates or availability zones.
22:30:16 sean-k-mooney cfriesen: well an aggreaget is just a bag of resouce providers nothing more or less. you can use them to group anything you want
22:30:18 efried cfriesen: Except that placement aggregates != host aggregates
22:30:39 efried cfriesen: The other main use case for aggregates is to associate sharing resource providers, like a shared storage pool.
22:31:25 efried ^ placement aggregates, that is.
22:32:07 cfriesen efried: okay...I'm not up on the placement terminology. :)
22:32:35 efried cfriesen: I'm not up on... pretty much everything else. Between the two of us, we're gonna need a pidgin.
22:33:22 sean-k-mooney cfriesen: a sharing resouce provider is jsut a normal resouce provider with a trait and then its inventores are consumable by anything in the same aggreates as it
22:34:36 cfriesen so how do we handle the idea of an instance consuming cpu/ram from one placement aggregate but a PCI device from another placement aggregate? (for the non-strict PCI affinity case)
22:34:44 openstackgerrit Matt Riedemann proposed openstack/nova master: Test websocketproxy with TLS in the nova-next job https://review.openstack.org/513160
22:34:49 mriedem melwitt: cleaned that up ^
22:35:28 sean-k-mooney cfriesen: instance consume resocues form invetories provided by resouce providers
22:35:40 melwitt mriedem: thanks. sorry I lost track of that one
22:35:47 sean-k-mooney cfriesen: and resouce provires can be member of 0-n aggregates
22:36:30 sean-k-mooney so you are not consuming form the aggregates
22:37:22 efried cfriesen: In the design I'm strawpersoning, if you care about CPU/mem affinity and you don't care about VF affinity, you would say affine_resources=VCPU,MEMORY_MB but not SRIOV_NET_VF
22:37:46 efried cfriesen: s/VF/PCI device/ if you wish.
22:39:08 sean-k-mooney efried: that may break down a bit with anti afinity
22:39:33 efried sean-k-mooney: I'm saying we're not handling anti-affinity at all with this model.
22:39:34 sean-k-mooney e.g. i want 2 VF whoes parent PF are anti affined
22:39:48 sean-k-mooney efried: ah ok
22:39:52 efried sean-k-mooney: Yeah, I get that. But that's not a NUMA use case, right?
22:40:01 efried It's more about HA
22:40:05 efried redundancy
22:40:15 efried powerfail grouping
22:40:18 efried that kind of thing.
22:40:21 sean-k-mooney ya its for ha bonding so my bond dont come from the same card
22:40:51 efried So yeah, this doesn't cover that. And if there's a generic solution that encompasses both, great. But I'm gonna have to pick Jay's brain for that.
22:40:51 sean-k-mooney e.g. pulling out one cable wont bring donw the bond
22:42:24 sean-k-mooney efried: if you can show your propasl support affinity in a clean way and does not block future addtion of anti afinity by design then its better then nothing
22:43:02 sean-k-mooney e.g. we dont have to support everything in one go in rocky
22:43:33 sean-k-mooney it would be nice however
22:44:25 efried That's kinda what I was thinking. Propose it this way, show that it wouldn't prevent a more generic solution in the future, and then Jay can propose (or not) that more generic solution.
22:46:14 cfriesen efried: what about multi-numa-node guests? they would need to essentially do numa anti-affinity (but host affinity) between the two pools of memory
22:47:02 efried cfriesen: What does that mean? Multi-numa-node guests? You mean guests that don't care about NUMA affinity? Or do they explicitly care about *anti* affinity?
22:47:05 sean-k-mooney cfriesen: multi numa guest do not need to be on multiple host numa nodes
22:47:14 cfriesen efried: no, I mean guests with multi numa nodes
22:47:22 sean-k-mooney cfriesen: i consider that limitation in libvirt to be a bug not a feature
22:47:30 sean-k-mooney libvirt virt driver that is
22:47:45 cfriesen sean-k-mooney: one reason why you might "require" multi-numa-node guests is to get access to more memory bandwidth
22:48:48 sean-k-mooney cfriesen: yes but there is a flavour extra spec defined for that that only the hyperv dirver support
22:49:00 sean-k-mooney cfriesen: i think the libvirt one should too
22:49:17 cfriesen sean-k-mooney: sorry, don't follow.
22:50:09 sean-k-mooney cfriesen: hw:numa_mempolicy
22:50:13 cfriesen sean-k-mooney: I can currently specify a multi-numa-node guest with libvirt, and I get access to that many host numa nodes, with accordingly more host memory bandwidth. (presumably combined with dedicated cpus)
22:50:59 cfriesen numa_mempolicy isn't mentioned anywhere in nova
22:51:41 sean-k-mooney cfriesen: i used to be in the hyperv driver
22:52:06 sean-k-mooney anyway that is off topic
22:52:25 sean-k-mooney efried: a multi numa gues is a guest with multiple virtual numa nodes
22:52:55 cfriesen efried: and currently they map to different physical numa nodes of the same host.
22:53:03 sean-k-mooney efried: you can choose what virtual cores are associated to each virtual numa node and you can device the memoy between each virtual numa node
22:54:09 efried Well, if you wanted to get that specific, you could probably swang something with granular request groups.
22:54:29 efried But we still wouldn't be able to guarantee you anti-affinity.
22:54:49 cfriesen if you can't guarantee numa-antiaffinity it'll change the current behaviour
22:55:02 cfriesen (ie if both virtual numa nodes could end up on the same host numa node)
22:55:06 sean-k-mooney placement does not need to give that guarentee
22:55:14 sean-k-mooney the virt driver can via the resouce tracker
22:55:40 efried sean-k-mooney: Yeahhhh, but we eventually *should* be able to implement that filter in placement.
22:55:44 sean-k-mooney or the nova numa topology filter
22:56:09 sean-k-mooney efried: talk to jay about that. in principal i agree
22:57:23 sean-k-mooney i would like placement to fully handel numa. but if it can handel 80% of numa in rocky with the rest in the filetr i would be happy with that too
22:57:46 efried Okay. I'll try to put up that spec later this week.

Earlier   Later