| Posted | Nick | Remark | |
|---|---|---|---|
| #openstack-nova - 2018-01-18 | |||
| 09:57:14 | kashyap | mdbooth: Hehe | |
| 09:57:19 | kashyap | I love the bike | |
| 09:57:21 | mdbooth | Cold is fine, wet is... sometimes ok. Ice and wind are never ok. | |
| 09:57:27 | kashyap | Aside: http://knmi.nl/waarschuwingen (About the weather alarm; use translate) | |
| 10:11:19 | mdbooth | kashyap: So I just finished reading and digesting your commit message, and the linked article | |
| 10:12:34 | mdbooth | kashyap: FWIW, assuming libvirt/qemu aren't planning on handling this at that layer, we should probably switch this on *by default* in Nova, with a tuning knob to switch it back off if somebody has a weird reason to need that. | |
| 10:14:21 | kashyap | mdbooth: So you raise a good point | |
| 10:15:04 | mdbooth | kashyap: That wouldn't be a something for this change, though | |
| 10:15:04 | kashyap | mdbooth: About being handled at QEMU level. I was involved on the qemu-devel list thread where there was a debate about adding a specific CPU model | |
| 10:15:16 | kashyap | That will include the PCID feature by *default* | |
| 10:15:34 | kashyap | However. It's soon getting ridiculous, because of this exhibit: | |
| 10:16:03 | kashyap | E.g. one CPU model name will be: 'Haswell-noTSX-IBRS' | |
| 10:16:08 | kashyap | See anything odd? | |
| 10:16:23 | kashyap | How about this: 'Haswell-noTSX-IBRS-nowThisBug-andAlsoThusBug-AndMoreBugs' | |
| 10:16:33 | mdbooth | Yeah, that's pretty silly | |
| 10:17:14 | kashyap | So that was the thread, please read DanPB's response on context of Nova: https://lists.nongnu.org/archive/html/qemu-devel/2018-01/msg03689.html | |
| 10:17:24 | kashyap | mdbooth: Thanks for taking time to read this! I appreciate the concentration | |
| 10:17:52 | kashyap | 'this' as in: the Nova change and all the related info | |
| 10:18:18 | kashyap | mdbooth: But the thing is...about switching on by _default_, it's complicated. | |
| 10:18:23 | mdbooth | From reading that post, it sounds like this is something (?:pretty much)? all our users will want. | |
| 10:18:47 | kashyap | As this perf impact will only occur when someone chooses one of the named CPU models that doesn't have the PCID feature in libvirt's notion of the model. | |
| 10:18:52 | kashyap | (But, it exists in the hardware) | |
| 10:18:59 | kashyap | mdbooth: It won't affect users | |
| 10:19:06 | kashyap | ...*if* they chose 'host-model' | |
| 10:19:12 | mdbooth | users==operators | |
| 10:19:16 | kashyap | In that case, libvirt will do the grump work for you | |
| 10:19:23 | kashyap | mdbooth: (Yeah, I translated as much) | |
| 10:19:29 | kashyap | s/grump/grunt/ | |
| 10:20:27 | kashyap | mdbooth: One thing is for *sure*. We (or I) have to do some clear write-up about what is the operator impact in all three cases: | |
| 10:20:31 | kashyap | (1) 'host-model' | |
| 10:20:35 | kashyap | (2) 'host-passthrough' | |
| 10:20:51 | mdbooth | kashyap: Is that something we can detect at startup? Host-model = X, real cpu is Y, so we should change host-model to X + PCID | |
| 10:20:51 | kashyap | (3) Custom CPU model <-- this is where my change comes in | |
| 10:22:30 | mdbooth | A reason *not* to do this would be the upgrade impact of not being able to live-migrate across the boundary? | |
| 10:22:45 | mdbooth | i.e. X != X+PCID, so you can't live migrate | |
| 10:23:04 | Spazmotic | You mean for the CPU mismatch? | |
| 10:23:04 | kashyap | mdbooth: Yes, live migration is a strict requirement | |
| 10:23:08 | mdbooth | Yeah | |
| 10:23:10 | kashyap | Yes, mismatch, indeed | |
| 10:23:13 | mdbooth | Spazmotic: Yeah | |
| 10:23:31 | kashyap | mdbooth: I still need to think more clearly about detecting it at guest start-up, where Nova generates the guest XML | |
| 10:23:41 | mdbooth | So if we could maintain host model strictly across live migrate, but reset it on restart... | |
| 10:23:41 | kashyap | What you're saying is --- | |
| 10:24:00 | mdbooth | s/guest/host/ startup | |
| 10:24:01 | kashyap | "If the guest is using a CPU without the PCID feature, add it." | |
| 10:24:22 | kashyap | (Assuming we detect that cleanly.) | |
| 10:26:32 | kashyap | mdbooth: So, from #virt, OFTC: | |
| 10:26:33 | kashyap | 11:21 < kashyap> Hi folks, can we detect at boot / start-up time what CPU model a guest is using? | |
| 10:26:36 | kashyap | 11:22 < kashyap> E.g. 'host-model' / 'host-passthrough' / or some custom model | |
| 10:26:39 | kashyap | 11:25 < danpb> in general you can't | |
| 10:26:41 | kashyap | 11:25 < danpb> you just get to see a model + flags in /proc/cpuinfo | |
| 10:27:58 | mdbooth | kashyap: I don't understand why we're interested in what the guest sees | |
| 10:28:17 | mdbooth | I mean, the guest is interested in what the guest sees, but we're just interested in what we present | |
| 10:28:36 | mdbooth | And we can absolutely determine that | |
| 10:30:11 | kashyap | mdbooth: Err, you're right. We should detect before the start-up | |
| 10:30:49 | mdbooth | kashyap: My thinking was that the host configuration for how guests are started is static, so we'd do it at host startup. | |
| 10:30:51 | kashyap | Using the baselineCPU() API | |
| 10:30:58 | kashyap | And VIR_CONNECT_BASELINE_CPU_EXPAND_FEATURES. | |
| 10:31:56 | mdbooth | Does that require a running domain? Or does it just query the xml file you referenced? | |
| 10:34:17 | kashyap | No, not a running domain | |
| 10:35:29 | kashyap | mdbooth: Near as I recall, libvirt caches the capabilities info it got from QEMU, via a QMP command. | |
| 10:35:48 | kashyap | (Before guest start up, obviously) | |
| 10:35:55 | kashyap | We'd be querying that cached XML, IIUC. | |
| 10:38:16 | Spazmotic | Checking my Xen boxes out of curiousity now | |
| 10:39:32 | kashyap | mdbooth: I'll stop hand-waving here, and get back with clearer details on the review. | |
| 10:39:39 | openstackgerrit | Stephen Finucane proposed openstack/nova master: Fix openstackdocstheme options for api-ref https://review.openstack.org/534712 | |
| 10:39:39 | openstackgerrit | Stephen Finucane proposed openstack/nova master: trivial: Remove crud from 'conf.py' https://review.openstack.org/534713 | |
| 10:40:55 | Spazmotic | Yeah interesting... i may need to send out an email to some peoples heh | |
| 10:41:47 | bauzas | gentle reminder for gibi and stephenfin about https://review.openstack.org/#/c/528832/ ;) | |
| 10:41:51 | lyarwood | kashyap: before you go, have you seen `Live Migration failure: invalid argument: monitor must not be NULL: libvirtError: invalid argument: monitor must not be NULL | |
| 10:42:05 | lyarwood | kashyap: urgh sorry, paste fail, ^ LM failure on f26 with the virt-preview repo | |
| 10:43:03 | kashyap | lyarwood: Got a link? | |
| 10:43:12 | lyarwood | ah wait that isn't the virt-preview repo | |
| 10:43:19 | lyarwood | kashyap: nope sorry local failure at the moment | |
| 10:44:29 | kashyap | Phew, okay. One less live migration issue to worry about, then | |
| 10:46:41 | lyarwood | cool, virt-preview works anyway | |
| 10:46:43 | lyarwood | sorry about that | |
| 10:51:30 | kashyap | No problem | |
| 11:02:37 | ameeda | jaypipes: good morning :) what If I use glance API to check the instance_metadata from nova side ? where is the wrong in that ? | |
| 11:04:27 | ameeda | jaypipes: I think that glance store image properties to nova db table "instance_metadata" | |
| 11:09:21 | mdbooth | kashyap: Reviewed, thanks. | |
| 11:09:52 | kashyap | mdbooth: Thank _you_ for spending time on this. I will read this on the train | |
| 11:14:58 | kashyap | mdbooth: A quick 'parser error', what do you mean by: "I don't think it lives [?] in this commit" | |
| 11:20:35 | kashyap | mdbooth: Disregard me, I got it. | |
| 11:34:30 | ameeda | jaypipes: I will update the patch with current functional test adding new comments. | |
| 11:35:45 | ameeda | jaypipes: I tried to access Nova DB directly, but I faced many issues. so I think the current scenario . | |
| 11:36:48 | ameeda | jaypipes: create new image with properties greater than 255 chars, then create server, great image from the created server, | |
| 11:37:30 | ameeda | jaypipes: new created image contains new image properties which is fetched from nova db. | |
| 11:38:04 | ameeda | jaypipes: then I check the image properties for new created image, so if it truncated it will raise error ] | |
| 11:40:41 | openstackgerrit | Ameed Ashour proposed openstack/nova master: change instance_system_metadata column type https://review.openstack.org/526900 | |
| 12:03:51 | bhujay | hello everybody , is there any config for niva so that multiple compute nodes can share a Fibre Channel Storage Array ? | |
| 12:09:11 | lyarwood | bhujay: for block storage? Nova would just use Cinder to provide volumes from the array and use os-brick to connect to these. | |
| 12:11:47 | bhujay | lyarwood: I am looking for VM to boot from the shared storage , just the way it does for rbd . This for ease of live migration | |
| 12:13:37 | lyarwood | bhujay: Yup, you can boot from volumes provided by cinder, the only imagebackend you could use for this would be LVM but IMHO using a cinder volume driver for the FC array is the way to go. | |
| 12:21:06 | bhujay | lyarwood: Just to make sure I got it correctly . Cinder to be configured for FC block storage , create bootable image volumes and while creating instance select that volume for boot. Right ? | |
| 12:26:16 | bhujay | lyarwood: But if we use boot from a image which is there in glance with glance backend as ceph will the instance OS ephemeral be placed on FC array by nova ? | |
| 12:28:06 | lyarwood | bhujay: yeah, create the volume from the image with `cinder create --image $UUID $SIZE ; cinder set-bootable $VOLUME_ID true` and then boot from it with `nova boot --boot-volume $VOLUME_ID [..] etc` | |
| 12:28:28 | lyarwood | bhujay: cinder copies the image into a volume with the above | |
| 12:28:44 | lyarwood | bhujay: but if you're already using rbd I don't see why you need the FC array :) | |
| 12:30:57 | openstackgerrit | Ameed Ashour proposed openstack/nova master: change instance_system_metadata column type https://review.openstack.org/526900 | |