| Posted | Nick | Remark | |
|---|---|---|---|
| #openstack-nova - 2019-09-06 | |||
| 10:38:32 | sean-k-mooney | oh that should be the union of hte cores used for the vcpus | |
| 10:38:39 | sean-k-mooney | and should be updated | |
| 10:38:42 | artom | It only matters when an instance without a NUMA topology lands on a host with vcpu_pin | |
| 10:38:46 | stephenfin | alex_xu: You around today? I'm planning to work through the VPMEM series again, but is there any chance you could look at https://review.opendev.org/#/c/671800 ? | |
| 10:38:53 | sean-k-mooney | artom: yes | |
| 10:39:03 | artom | sean-k-mooney, right, so is that... a real thing? | |
| 10:39:10 | sean-k-mooney | yes | |
| 10:39:10 | artom | We tell people to split them by aggregates | |
| 10:39:17 | stephenfin | artom: wait for it | |
| 10:39:23 | artom | They don't? | |
| 10:39:29 | sean-k-mooney | you shoudl set vcpu_pin_set on all host ideally | |
| 10:39:37 | stephenfin | (sean-k-mooney is about to tell you that vcpu_pin_set has nothing to do with pinning and should always be set) | |
| 10:39:39 | sean-k-mooney | and it will be need for the new stuff stephenfin is working on | |
| 10:39:53 | sean-k-mooney | stephenfin: :) | |
| 10:39:53 | artom | Haha | |
| 10:40:18 | openstackgerrit | Gorka Eguileor proposed openstack/nova master: Use os-brick locking for volume attach and detach https://review.opendev.org/614190 | |
| 10:40:19 | stephenfin | but yeah, vcpu_pin_set != pinning (or NUMA, for that matter) | |
| 10:40:24 | stephenfin | we use it for this stuff | |
| 10:40:36 | sean-k-mooney | its the set of core that the hypvervior may use for vms | |
| 10:40:41 | sean-k-mooney | nothing more or less | |
| 10:40:41 | stephenfin | and also to decide how much VCPU inventory to report to placement | |
| 10:41:05 | sean-k-mooney | yep its used for the inventory in the RT which is used for plamcent | |
| 10:41:22 | stephenfin | I don't agree with it being set by default, but yeah, it being set even when you don't have NUMA or pinning is a thing | |
| 10:41:25 | artom | OK, so it needs to be handled | |
| 10:41:30 | stephenfin | afraid so | |
| 10:41:38 | sean-k-mooney | stephenfin: you realise that for you vcpu work | |
| 10:41:51 | sean-k-mooney | we shoudl also be handelign shared and dedicated sets | |
| 10:42:03 | artom | sean-k-mooney, I'll most likely as you to test that bit when it's implemented | |
| 10:42:17 | stephenfin | sean-k-mooney: what do you mean? | |
| 10:42:18 | artom | Or if we can set it in the gate and show that working... | |
| 10:42:57 | sean-k-mooney | stephenfin: well for non pinned instance the cpuset will need to be updated based on the dest cpu_shared_set | |
| 10:43:24 | sean-k-mooney | and for pinned instnce it need to be updated based on the dedicated set | |
| 10:43:40 | sean-k-mooney | although the second half will be done by artoms exisitng code | |
| 10:44:07 | stephenfin | sean-k-mooney: Yeah, correct. That code has been updated here https://review.opendev.org/#/c/671800/28/nova/virt/libvirt/driver.py | |
| 10:44:42 | openstackgerrit | Luyao Zhong proposed openstack/nova master: object: Introduce Resource and ResourceList objs https://review.opendev.org/678448 | |
| 10:44:43 | openstackgerrit | Luyao Zhong proposed openstack/nova master: Retrieve the allocations early https://review.opendev.org/678450 | |
| 10:44:43 | openstackgerrit | Luyao Zhong proposed openstack/nova master: Add resources dict into _Provider https://review.opendev.org/678449 | |
| 10:44:44 | openstackgerrit | Luyao Zhong proposed openstack/nova master: Claim resources in resource tracker https://review.opendev.org/678452 | |
| 10:44:45 | openstackgerrit | Luyao Zhong proposed openstack/nova master: libvirt: report VPMEM resources by provider tree https://review.opendev.org/678454 | |
| 10:44:45 | openstackgerrit | Luyao Zhong proposed openstack/nova master: libvirt: Enable driver discovering PMEM namespaces https://review.opendev.org/678453 | |
| 10:44:46 | openstackgerrit | Luyao Zhong proposed openstack/nova master: Parse vpmem related flavor extra spec https://review.opendev.org/678456 | |
| 10:44:46 | openstackgerrit | Luyao Zhong proposed openstack/nova master: libvirt: Support VM creation with vpmems and vpmems cleanup https://review.opendev.org/678455 | |
| 10:44:47 | openstackgerrit | Luyao Zhong proposed openstack/nova master: Add functional tests for virtual persistent memory https://review.opendev.org/678470 | |
| 10:44:47 | openstackgerrit | Luyao Zhong proposed openstack/nova master: libvirt: Enable driver configuring PMEM namespaces https://review.opendev.org/679640 | |
| 10:44:48 | openstackgerrit | Luyao Zhong proposed openstack/nova master: doc: attaching virtual persistent memory to guests https://review.opendev.org/680300 | |
| 10:46:22 | sean-k-mooney | by the way people know that cpuset in the vcpu element is funtionally the same as generatin vcpuin elemnt for each core pinning the core to the set of host core set in vcpu pin set | |
| 10:46:31 | artom | stephenfin, sean-k-mooney, actually, while we're on the topic, I thought you could get CPU pinning with https://libvirt.org/formatdomain.html#elementsCPUTuning without NUMA pinning with https://libvirt.org/formatdomain.html#elementsNUMATuning | |
| 10:46:47 | artom | For example by using only the cpu_policy=dedicated extra spec | |
| 10:46:49 | sean-k-mooney | artom: yes you can | |
| 10:46:57 | sean-k-mooney | no | |
| 10:47:05 | sean-k-mooney | if you use cpu_policy=dedicated | |
| 10:47:11 | sean-k-mooney | you will have a numa toplogy | |
| 10:47:20 | sean-k-mooney | if you set vcpu_pin_set on the host | |
| 10:47:22 | artom | But looking at the code of _get_guets_numa_config, IIUC if it sets vcpupin, it'll also set the memnode stuff | |
| 10:47:26 | sean-k-mooney | and set nothing on the flavor | |
| 10:47:42 | sean-k-mooney | you get <vcpu placement='static' cpuset="1-4,^3,6" current="1">2</vcpu> | |
| 10:47:45 | artom | sean-k-mooney, right, that's the case with <vcpu cpuset=blah> that I'm not handling | |
| 10:47:55 | sean-k-mooney | or whatever the vcpu_pin_set is but no numa affinity | |
| 10:48:05 | artom | I was talking about <cputune> without <numatune> | |
| 10:48:17 | artom | Is that possible? Don't think so | |
| 10:48:19 | sean-k-mooney | oh libvirt allows that but we never generate it | |
| 10:48:23 | artom | Aha, ok | |
| 10:48:53 | artom | Confirms what I was seeing in the code, thanks | |
| 10:49:03 | sean-k-mooney | we only generate the cputune element if you have pinning nebaled which give you a numa toplogy | |
| 10:49:16 | sean-k-mooney | we do generate numatune without cpu tune i think | |
| 10:49:26 | sean-k-mooney | e.g. if you just enable hugepages but no pinning | |
| 10:49:32 | sean-k-mooney | or hw:numa_nodes=X | |
| 10:50:23 | artom | sean-k-mooney, you sure? It would happen only if https://github.com/openstack/nova/blob/master/nova/virt/libvirt/driver.py#L4609 doesn't iterate at all | |
| 10:50:52 | openstackgerrit | Luyao Zhong proposed openstack/nova master: libvirt: report VPMEM resources by provider tree https://review.opendev.org/678454 | |
| 10:50:52 | sean-k-mooney | we definitly numa affine hugepages | |
| 10:50:53 | openstackgerrit | Luyao Zhong proposed openstack/nova master: Parse vpmem related flavor extra spec https://review.opendev.org/678456 | |
| 10:50:53 | openstackgerrit | Luyao Zhong proposed openstack/nova master: libvirt: Support VM creation with vpmems and vpmems cleanup https://review.opendev.org/678455 | |
| 10:50:54 | openstackgerrit | Luyao Zhong proposed openstack/nova master: Add functional tests for virtual persistent memory https://review.opendev.org/678470 | |
| 10:50:54 | openstackgerrit | Luyao Zhong proposed openstack/nova master: libvirt: Enable driver configuring PMEM namespaces https://review.opendev.org/679640 | |
| 10:50:55 | openstackgerrit | Luyao Zhong proposed openstack/nova master: doc: attaching virtual persistent memory to guests https://review.opendev.org/680300 | |
| 10:51:10 | sean-k-mooney | im just not sure if we use vcpupin to pin the cores to the numa node or cpuset | |
| 10:51:12 | artom | sean-k-mooney, right, I'm saying in those cases we also affine guest CPUs with cputune | |
| 10:51:19 | sean-k-mooney | ah ok | |
| 10:51:34 | sean-k-mooney | well as i said cpuset is just a shortcurt for vcpupin | |
| 10:51:43 | sean-k-mooney | let me give an example | |
| 10:52:28 | sean-k-mooney | those are the same | |
| 10:52:31 | sean-k-mooney | <vcpu placement='static' cpuset="1-4" current="1">2</vcpu> | |
| 10:52:33 | sean-k-mooney | <cputune> | |
| 10:52:35 | sean-k-mooney | <vcpupin vcpu="0" cpuset="1-4"/> | |
| 10:52:37 | sean-k-mooney | <vcpupin vcpu="1" cpuset="1,2,3,4"/> | |
| 10:52:39 | sean-k-mooney | <cputune> | |
| 10:52:54 | sean-k-mooney | libvirt internally generate the vcpupin element form the cpu set | |
| 10:53:04 | sean-k-mooney | and generge the same qemu commandline | |
| 10:53:12 | artom | Wait a second. | |
| 10:53:24 | artom | My *entire* code triggers if the instance has a NUMA topology | |
| 10:53:32 | sean-k-mooney | yes | |
| 10:53:53 | artom | So the insance with no NUMA topology on a host with vcpu_pin_set case will have to...? | |
| 10:54:12 | sean-k-mooney | you will need to add support for it | |
| 10:54:22 | sean-k-mooney | well you or stepen | |
| 10:54:30 | sean-k-mooney | technically this is a long standing issue | |
| 10:54:43 | sean-k-mooney | we always should have been updating the xml | |
| 10:54:54 | sean-k-mooney | it will be fixed via a hard reboot however | |
| 10:55:15 | sean-k-mooney | libvirt will reject the migration if the cores do not exist on the dest | |
| 10:55:25 | artom | Right, true | |
| 10:55:33 | artom | I'm surprised more people haven't been hitting that | |