Earlier  
Posted Nick Remark
#openstack-nova - 2018-01-11
13:45:39 mdbooth Second is I suspect we're going to want to have different ways to create this list.
13:45:41 bauzas it's not only the GPU thing, it's all the devices we create
13:45:45 openstackgerrit Stephen Finucane proposed openstack/nova master: trivial: Resolve 'oslo.context' deprecation warnings https://review.openstack.org/532822
13:45:52 mdbooth e.g. creating new mdevs vs re-using existing mdevs
13:46:16 mdbooth s/existing/already allocated to this instance/
13:46:25 bauzas mdbooth: that is precisely something that should kept hidden from the other libvirt places
13:46:33 bauzas for tech debt reasons
13:46:34 mdbooth Where creating == create or reallocate
13:46:54 bauzas how the vGPUs are implemented in libvirt should be kept very isolated from the rest
13:47:16 mdbooth I'm talking about the code in libvirt/driver.py here
13:47:20 bauzas sure
13:47:30 mdbooth So it shouldn't affect the decisions of other drivers.
13:47:37 bauzas I still think we need to limit the driver knowledge about what is a vGPU
13:48:16 mdbooth I think the guest needs to end up with a list of mdevs it should add
13:48:37 bauzas mdbooth: that's just basically splitting the method in twice
13:48:53 bauzas and introducing a new level of indirection, but okay
13:49:52 bauzas mdbooth: I still don't see why we really need to split in twice but okay
13:50:03 mdbooth Well I think the split is helpful to the reader, because it calls out a different type of operation in a different place. It also allows us to generate the list in different ways based on caller context without having to pass the context all the way down into _get_guest_xml
13:50:05 bauzas it just makes the interfaces a bit cleaner, but that's it
13:50:38 mdbooth e.g. If we're creating a new guest, allocating new instances, vs re-generating our existing xml with our current mdevs.
13:50:52 bauzas mdbooth: a-ha, that's different
13:51:01 bauzas mdbooth: I see your thoughts
13:51:36 bauzas mdbooth: ie. not passing allocated all the way down to _get_guest_xml() but rather just call _give_me_vgpus() in the caller itself ?
13:51:57 mdbooth bauzas: Something like that would be my preference.
13:51:57 bauzas and then the caller attaching those vgpus *after* it got the XML ?
13:52:08 mdbooth No, you have to do it before
13:52:39 bauzas mdbooth: you understand that it's mostly a convenience problem ?
13:52:41 mdbooth So you have to do the host config first, then you can generate the guest xml based on how you did it
13:53:04 mdbooth It's the same as the volume attachment problem. You can't generate the guest xml until you can tell libvirt where you mounted the volume.
13:53:09 bauzas mdbooth: so you would pass the extra bits down _get_guest_xml, right?
13:53:22 mdbooth Yes.
13:54:52 bauzas mdbooth: and then the whole _give_me_mdevs() would be a synchronized section?
13:55:15 bauzas gibi: chairing the nova meeting, right?
13:55:33 mdbooth bauzas: Yes, it would have to be.
13:55:55 bauzas mdbooth: I still consider the synchronized section to be in a follow-up
13:56:02 bauzas mdbooth: timeline wise
13:56:25 bauzas because we're 2 weeks away from the FF, and I don't want that feature to be punted
13:56:58 mdbooth bauzas: Adding that would be way simpler than moving code around out of _get_guest_xml
13:57:18 mdbooth Just a simple host-wide local lock
13:57:26 bauzas I agree
13:57:49 bauzas here are my thoughts
13:58:07 bauzas based on what I see important to be merged before end of Queens
13:58:16 bauzas #1 the patch itself
13:58:32 bauzas (with the synchronized lock)
13:58:50 bauzas #2 the possibility to disallow nova to create mdevs
13:59:10 bauzas #3 the fix for power-on and others
13:59:28 bauzas #4 the possibility to recreate mdevs after a reboot
13:59:33 bauzas and the rest looks optional to me
13:59:44 bauzas we can refine the interfaces in Rocky
14:00:23 bauzas besides the fact that the problem you mention isn't particular to the VGPU feature, but rather for a shit ton of extra devices and features we add to the guest
14:00:36 bauzas mdbooth: agree ?
14:00:38 mdbooth Only volumes, I think.
14:00:45 mdbooth I agree with your prioritisation.
14:00:55 gibi bauzas: I did not get the memo that I'm the chair
14:00:56 bauzas mdbooth: no, we litterally generate all the devices at the same place
14:01:10 gibi bauzas: but I can jump in
14:01:13 efried gibi bauzas Somebody oughtta do it. The time is now :)
14:03:01 mdbooth bauzas: I really think it's only volumes.
14:03:30 mdbooth The others we're just adding devices to the guest. The different with volumes and now vgpus is that we're reconfiguring the host.
14:03:46 mdbooth So we're doing _get_guest_xml() and it has surprising side-effects
14:03:48 bauzas mdbooth: https://github.com/openstack/nova/blob/master/nova/virt/libvirt/driver.py#L4882 ?
14:04:09 mdbooth bauzas: Yeah, that's just pulling them out of the instance.
14:04:16 mdbooth IIUC they've already been allocated elsewhere.
14:04:17 bauzas mdbooth: what's the difference between attaching a PCI device or a mediated device ?
14:04:40 mdbooth Because you're not creating the pci device, or doing any kind of 'scheduling'
14:05:13 mdbooth You're not calling _get_guest_xml() and suddenly you've created a new pci device
14:05:59 mdbooth Anyway, I agree with your prioritisation. That said, I really don't like it about volumes, and I'm uncomfortable adding more of it.
14:06:03 bauzas mdbooth: but the allocation was made way earlier :)
14:06:12 bauzas for the VGPU
14:06:35 bauzas if you go to that compute, that's because it has free space for a vGPU
14:06:45 bauzas because placement told you
14:06:58 bauzas exactly like NUMA things
14:07:12 bauzas the scheduler or placement won't know which specific device you're assigned
14:07:19 bauzas but it will know there is room for you
14:07:39 bauzas in theory, you should never raise the exception I added
14:08:47 mdbooth bauzas: There's also the case of recreating the domain xml with existing mdevs.
14:09:35 mdbooth Although perhaps that doesn't matter as long as we're absolutely certain the domain isn't running.
14:09:58 mdbooth But again that's an embedded assumption which makes everything harder to understand.
14:10:32 bauzas mdbooth: the problem is that the XML isn't persisted when you stop the instance
14:10:38 bauzas I mean in Nova
14:10:46 bauzas that's surprising to me
14:10:57 bauzas we regenerate the XML when we start again the instance
14:11:02 bauzas while it's already been defined !
14:11:23 bauzas mdbooth: any idea you know why ?
14:11:25 mdbooth Yeah, that's... suboptimal.
14:11:34 bauzas that's rather errorprone
14:11:52 mdbooth A legacy design choice that Nova is the ultimate source of truth. It means we lose stuff like device ordering and address assignments.
14:13:04 bauzas I can understand the reasoning
14:13:12 bauzas but then, you somehow need to persist that
14:13:19 bauzas and rather not rebuilding it
14:13:23 bauzas but that's MHO
14:13:34 bauzas and I'm possibly missing some design choices
14:14:28 artom It may have come from the need to support other drivers?
14:14:35 artom Who don't necessarily keep state like libvirt does?
14:14:50 mdbooth artom: This is an internal libvirt driver detail.
14:15:10 bauzas anyway
14:15:36 bauzas I have some stupid formatting to do, because pci names are different from filesystem
14:15:47 bauzas I mean pci names in libvirt
14:15:53 bauzas grrrr
14:15:56 mdbooth bauzas: I agree that moving that stuff is less important, though.

Earlier   Later