Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2019-02-05
16:14:20 sean-k-mooney mriedem: by the way i agree with your ml post regardign changing the jobs. when i get my third party ci triggering of upstream it would be good to get your input on what you would like it to test
16:15:07 mriedem ok
16:17:33 openstackgerrit sean mooney proposed openstack/nova master: Libvirt: do not set mac when unplugging macvtap vf https://review.openstack.org/624842
16:17:34 openstackgerrit sean mooney proposed openstack/nova master: Allow per-port modification of vnic_type and profile https://review.openstack.org/607365
16:17:34 openstackgerrit sean mooney proposed openstack/nova master: Add free for claimed, allocated devices https://review.openstack.org/616120
16:17:35 openstackgerrit sean mooney proposed openstack/nova master: SR-IOV Live migration indirect port support https://review.openstack.org/620115
16:17:35 openstackgerrit sean mooney proposed openstack/nova master: Add get_instance_pci_request_from_vif https://review.openstack.org/619929
16:17:36 openstackgerrit sean mooney proposed openstack/nova master: libvirt: auto detach/attach sriov ports on migration https://review.openstack.org/629589
16:18:02 sean-k-mooney i fixed the merge conflict ^ otherwise its the same
16:23:59 openstackgerrit Rodolfo Alonso Hernandez proposed openstack/os-vif master: Retrieve the table schema using ovsdb-client command https://review.openstack.org/634967
16:36:43 artom sean-k-mooney, so, explain again why the nodeset in memoryBacking isn't necessary? If the dest compute has differently-numbered NUMA nodes, doesn't that need updating?
16:37:48 sean-k-mooney it is but its is carried in the numa_mappings info already
16:37:54 sean-k-mooney https://etherpad.openstack.org/p/numa_migration
16:40:34 sean-k-mooney i think the object i have listed in https://etherpad.openstack.org/p/numa_migration should have all the info you need?
16:42:10 artom sean-k-mooney, sorry, still don't get how if the source has <page size="1" unit="G" nodeset="4"/> and on the dest there is no nodeset 5 because there's only 4 NUMA nodes, how the source is supposed to update that bit of the XML
16:43:30 openstackgerrit Rodolfo Alonso Hernandez proposed openstack/os-vif master: Retrieve the table schema using ovsdb-client command https://review.openstack.org/634967
16:43:48 sean-k-mooney simple if it recives numa_maping:{0:2, 1:3} from the dest it sets nodeset=2,3
16:44:30 sean-k-mooney nodeset will be the union of of all the host numa nodes listed in numa_maping
16:44:50 artom Ah, yep, that's annoyingly obvious
16:45:11 artom The kernel will spread the pages across all nodes as equally as it can
16:45:13 sean-k-mooney libvirt allows mixed page size for the xml but nova does not
16:45:21 sean-k-mooney yes
16:45:25 artom Right, I knew the size won't change
16:45:26 sean-k-mooney by default
16:45:35 sean-k-mooney you can allocate pages differently if you choose too
16:46:24 sean-k-mooney artom: the mempage size could cange form 2M to 1G on migration if the flavor had hw:mem_page_size=large
16:46:40 sean-k-mooney or at leaste we woudl be allowed to do that
16:46:47 artom I feel like that should not be allowed?
16:46:48 sean-k-mooney if we choose to support that is another matter
16:47:06 artom I don't think the guest would like that...
16:47:06 sean-k-mooney well large state you will recive any pagezie other then the smallest available
16:47:15 sean-k-mooney the guest wont be able to tell
16:47:38 artom So wait, the guest will continue to see 1G pages, but they'll be backed by 2M?
16:47:51 sean-k-mooney the guest does nto see 1G hugepages
16:48:06 sean-k-mooney this all happens in qemu
16:48:21 sean-k-mooney the guest sees virtual ram dims
16:48:41 dims o/ :)
16:48:50 sean-k-mooney hehe o/
16:49:09 dims LOL
16:49:52 sean-k-mooney artom: if the guest expcitly requested 1G pages in the flavor and the destiation cant provide that the resouce tracker will fail to claim it
16:50:12 sean-k-mooney artom: but this is why i added numa_pagesize
16:50:26 artom sean-k-mooney, I'm confused because when I first set up my env for this, I gave the devstack machines 2M hugepages, but didn't configure anything on the host. With that I could boot only like, 1 VM out of 10
16:50:29 sean-k-mooney the guess will have only 1 page size for all its memory
16:50:32 artom The rest hung in 'spawning'
16:50:47 artom So I assumed guest and host pagesize need to match up, otherwise dragons!
16:51:22 sean-k-mooney no hugepages in the guest is totally independat
16:51:30 artom Which is why the idea of live-migrating a guest with 1G pages to a host that only has 2M pages (and this would get accepted because hw:numa_pagesize=large) confuses me
16:51:39 sean-k-mooney you can create 1G pages in the gues and have 4k pages on the host
16:52:12 artom And there's no performance impact? That's not been my experience... Or maybe I was doing something else wrong
16:52:25 openstackgerrit Rodolfo Alonso Hernandez proposed openstack/os-vif master: Add native implementation OVSDB API https://review.openstack.org/482226
16:52:42 artom I kinda feel we need danpb
16:52:57 sean-k-mooney of corse there is a performacne impact but if you did not ask for hugepages for the vm you cannot expect to have the performace benifit
16:53:44 artom In my case it appeared to be a debilitating performance impact
16:53:45 sean-k-mooney the point is the guest internal memory allocation stragy is not something nova shoudl care about
16:54:30 sean-k-mooney artom: it will perfom better then 4k pages in the geuset but it wont significatly improve performance
17:13:29 openstackgerrit Matt Riedemann proposed openstack/nova master: WIP: Fix legacy-grenade-dsvm-neutron-multinode-live-migration https://review.openstack.org/634962
17:15:40 sean-k-mooney artom: i updated https://etherpad.openstack.org/p/numa_migration
17:15:50 artom sean-k-mooney, yep, cheers for that
17:16:03 artom I've kinda deviated on trying to better understand hugepages
17:27:31 openstackgerrit Jack Ding proposed openstack/nova master: Correct instance port binding for rebuilds https://review.openstack.org/603844
17:28:12 artom sean-k-mooney, so next problem is, with hw:mem_page_size=large, will the scheduler and/or resource tracker claim allow migrating from a 1G host to a 2M host?
17:28:56 sean-k-mooney it would if we used the value form the flavor directly
17:29:44 artom So we need a way to prevent that...
17:30:23 sean-k-mooney artom: when you do the claim here https://review.openstack.org/#/c/634606/6/nova/compute/resource_tracker.py@255
17:30:28 openstackgerrit Jack Ding proposed openstack/nova master: Flavor extra spec and image properties validation https://review.openstack.org/620706
17:30:40 sean-k-mooney we will need to pas the pageszie used on the source node
17:31:04 artom sean-k-mooney, yesh, the claims aren't really set up to receive a pagesize separately
17:31:11 openstackgerrit Jack Ding proposed openstack/nova master: Preserve UEFI NVRAM variable store https://review.openstack.org/621646
17:31:13 artom What if we just hax the flavor we pass it? ;)
17:31:56 sean-k-mooney artom: well you are passing the migration so why not prepoulate the instance_host_mapping with the pagesize info in the migration
17:32:08 openstackgerrit Jack Ding proposed openstack/nova master: Improve libvirt image and snapshot handling https://review.openstack.org/616692
17:33:08 artom sean-k-mooney, instance_host_mapping?
17:33:27 sean-k-mooney the new object im suggesting here https://etherpad.openstack.org/p/numa_migration
17:33:38 artom Oh, your thing from the etherpad
17:33:44 sean-k-mooney ya
17:34:06 openstackgerrit Matt Riedemann proposed openstack/nova master: Move resize.prep.start/end notifications to helper method https://review.openstack.org/633298
17:34:06 openstackgerrit Matt Riedemann proposed openstack/nova master: Isolate cell-targeting code in MigrationTask https://review.openstack.org/621310
17:34:07 openstackgerrit Matt Riedemann proposed openstack/nova master: Extract compute API _create_image to compute.utils https://review.openstack.org/621311
17:34:07 openstackgerrit Matt Riedemann proposed openstack/nova master: Move resize.(start|end) notification sending to helper method https://review.openstack.org/634831
17:34:08 openstackgerrit Matt Riedemann proposed openstack/nova master: Add Destination.allow_cross_cell_move field https://review.openstack.org/614035
17:34:08 openstackgerrit Matt Riedemann proposed openstack/nova master: Add Migration.cross_cell_move and get_by_uuid https://review.openstack.org/614012
17:34:09 openstackgerrit Matt Riedemann proposed openstack/nova master: Change HostManager to allow scheduling to other cells https://review.openstack.org/614037
17:34:09 openstackgerrit Matt Riedemann proposed openstack/nova master: Add InstanceAction/Event create() method https://review.openstack.org/614036
17:34:10 openstackgerrit Matt Riedemann proposed openstack/nova master: Add Instance.hidden field https://review.openstack.org/631123
17:34:10 openstackgerrit Matt Riedemann proposed openstack/nova master: Add CrossCellWeigher https://review.openstack.org/614353
17:34:11 openstackgerrit Matt Riedemann proposed openstack/nova master: Add CrossCellMigrationTask https://review.openstack.org/631581
17:34:11 openstackgerrit Matt Riedemann proposed openstack/nova master: Add TargetDBSetupTask https://review.openstack.org/627892
17:34:12 openstackgerrit Matt Riedemann proposed openstack/nova master: Add can_connect_volume() compute driver method https://review.openstack.org/621313
17:34:12 openstackgerrit Matt Riedemann proposed openstack/nova master: Execute TargetDBSetupTask https://review.openstack.org/633853
17:34:13 openstackgerrit Matt Riedemann proposed openstack/nova master: WIP: Add PrepResizeAtDestTask https://review.openstack.org/627890
17:34:13 openstackgerrit Matt Riedemann proposed openstack/nova master: Add prep_snapshot_based_resize_at_dest compute method https://review.openstack.org/633293
17:34:14 openstackgerrit Matt Riedemann proposed openstack/nova master: WIP: Add PrepResizeAtSourceTask https://review.openstack.org/627891
17:34:14 openstackgerrit Matt Riedemann proposed openstack/nova master: Add prep_snapshot_based_resize_at_source compute method https://review.openstack.org/634832
17:34:16 sean-k-mooney so the idea is that object is going to be passed back form the dest to source
17:34:28 artom sean-k-mooney, yeah, no I'm before that
17:34:31 artom When picking a host
17:34:39 artom We need to pick a host with the same pagesize as the source
17:34:40 sean-k-mooney but there is no reason it cant be passed form the source to dest
17:35:18 sean-k-mooney artom: the numa toplogy filter could also maybe look at the instance_numa_toplogy bolb and see if we recored the pagesize
17:35:45 sean-k-mooney if we do then it can pretent that page size was requested explcitly if hw:mem_page_size=large

Earlier   Later