| Posted | Nick | Remark | |
|---|---|---|---|
| #openstack-nova - 2023-02-10 | |||
| 10:40:50 | bauzas | Ogame, got it | |
| 10:41:14 | bauzas | that one stole too much of my free time | |
| 10:49:25 | bauzas | gibi: about https://review.opendev.org/c/openstack/nova/+/821228/6/nova/virt/libvirt/host.py#745 I wonder whether we really need to call *again* getCPUMap() | |
| 10:49:43 | bauzas | gibi: the existing logic just considers to return the cpu blindless | |
| 10:49:48 | bauzas | blindly | |
| 10:50:27 | bauzas | but I see your point | |
| 10:50:32 | bauzas | nevermind | |
| 10:51:24 | gibi | bauzas: you probably still need to differentiate between non existent cpu ids in the dedicated_set and existing but offined cpus | |
| 10:52:00 | bauzas | yeah, I'm about adding a get_available_cpus() which will return all CPUs for the map | |
| 10:53:23 | gibi | yeah that will work | |
| 10:59:45 | kashyap | bauzas: What exactly does getCPUMap() fetch? The docs only say "Get node CPU information" | |
| 11:00:12 | kashyap | Ah, it maps to `virsh cpu-stats` | |
| 11:00:57 | bauzas | yup | |
| 11:01:01 | kashyap | Oh, interesting. When I run `virsh cpu-stats` on my Fedora 36 VM for a guest, it gives me: | |
| 11:01:05 | bauzas | https://libvirt.org/html/libvirt-libvirt-host.html#virNodeGetCPUMap | |
| 11:01:12 | kashyap | $> sudo virsh cpu-stats 1 | |
| 11:01:13 | kashyap | error: Operation not supported: operation 'getCpuacctPercpuUsage' not supported for backend 'cgroup V2' | |
| 11:01:13 | kashyap | error: Failed to retrieve CPU statistics for domain 'el8-vm1' | |
| 11:01:25 | bauzas | [sbauza@sbauza temp]$ python | |
| 11:01:26 | kashyap | bauzas: Yeah, was reading. It looks like there's some accounting to be done w.r.t CGroups version | |
| 11:01:26 | bauzas | >>> conn = libvirt.open('qemu:///system') | |
| 11:01:26 | bauzas | >>> import libvirt | |
| 11:01:26 | bauzas | Type "help", "copyright", "credits" or "license" for more information. | |
| 11:01:26 | bauzas | Python 3.11.1 (main, Jan 6 2023, 00:00:00) [GCC 12.2.1 20221121 (Red Hat 12.2.1-4)] on linux | |
| 11:01:27 | bauzas | >>> conn.getCPUMap() | |
| 11:01:29 | bauzas | (8, [True, True, True, True, True, True, True, True], 8) | |
| 11:01:55 | kashyap | Yep | |
| 11:10:42 | sean-k-mooney | you can also just get the online cores form sysfs | |
| 11:11:18 | sean-k-mooney | if this is an issue but im fine with using libvirt's api for this | |
| 11:16:12 | bauzas | I'm done with the new rev, just updating the upper patch now | |
| 11:24:22 | sean-k-mooney | my steamdeck is charging so im currently still at my work laptop.(playing factorio) so gibi if your happy to review bauzas serise and it looks good to you ping me when your done and i can then do a final pass over it quickly and we can likely merge that today. with that said i have a doctors apointmen in a littel over 3 hours so ill be away after that. | |
| 11:24:38 | bauzas | that's a love | |
| 11:24:52 | auniyal | O/ | |
| 11:24:57 | gibi | sean-k-mooney: ack | |
| 11:25:04 | auniyal | in devstack is there a way to see nova-manage logs | |
| 11:25:20 | bauzas | I'm honestly torn, I don't know whether we should really pay attention to ignoring errors | |
| 11:25:36 | bauzas | and whether we should be cautious | |
| 11:25:55 | bauzas | honestly, we create a temp dir, so I don't except problems besides the full disk problem | |
| 11:26:11 | sean-k-mooney | you should not get any now with the way the fixture works | |
| 11:26:17 | bauzas | so I'll turn into using the .cleanup() method | |
| 11:26:22 | bauzas | and meh | |
| 11:26:30 | bauzas | meh to ignoring errors | |
| 11:26:42 | gibi | I'm fine ignoring errors | |
| 11:26:49 | gibi | during delete of a temp dir | |
| 11:27:00 | gibi | we will never reuse the temp dir as it has a random postfix | |
| 11:27:05 | sean-k-mooney | the flag is only there in 3.10 | |
| 11:27:09 | gibi | if my /tmp fills up that is on me | |
| 11:27:10 | sean-k-mooney | and we need to support 3.8 | |
| 11:27:17 | gibi | aah | |
| 11:27:21 | bauzas | gibi: that's the problem | |
| 11:27:26 | sean-k-mooney | so its fine for it to error | |
| 11:27:28 | sean-k-mooney | it wont | |
| 11:27:29 | gibi | ahh | |
| 11:27:30 | bauzas | we can't say 'ignore errors' if we use cleanup | |
| 11:27:38 | bauzas | hence me torn | |
| 11:27:41 | gibi | I'm fine both ways | |
| 11:27:41 | sean-k-mooney | i could have errored when we had the copy of sysfs because of some permision | |
| 11:27:51 | sean-k-mooney | but now its just normal files owned by us | |
| 11:27:54 | bauzas | correct | |
| 11:28:00 | gibi | if it starts failing on cleanup then we will switch to shutil | |
| 11:28:01 | sean-k-mooney | so it wont error unless there is a disk issue which si out of scope | |
| 11:28:08 | bauzas | yup, this ^ | |
| 11:28:09 | sean-k-mooney | +1 | |
| 11:28:44 | bauzas | I'll add a comment explaining the risk and how to mitigate it if we see it in CI | |
| 11:29:39 | gibi | cool | |
| 11:39:48 | bauzas | sean-k-mooney: gibi: I haven't yet written the docs patch as it requires a bit of effort, so I'll keep your comments on it unresolved but don't misunderstand me, surely I'll do it in a subsequent patch once I'm done (monday morning hopefully) | |
| 11:39:58 | bauzas | I just wanna give chance to other people to get reviewed | |
| 11:40:14 | gibi | bauzas: sure, doc is OK after FF | |
| 11:40:22 | sean-k-mooney | ya it can be a seperate patch | |
| 11:40:47 | bauzas | the core #0 note is actually very important | |
| 11:40:49 | bauzas | TIL about it | |
| 11:41:13 | bauzas | but yeah that makes sense from an OS perspective | |
| 11:41:21 | bauzas | you always rely on that core to be available | |
| 11:41:50 | bauzas | that's the most portable assumption | |
| 11:42:04 | sean-k-mooney | ya it generally need one core that can always be used to handel interupts and you know turn on the others | |
| 11:43:14 | sean-k-mooney | bauzas: that is why there is no online file in /sys/bus/cpu/devices/cpu0/ | |
| 11:43:24 | sean-k-mooney | but there is in all the rest | |
| 11:44:07 | bauzas | ooooooooh | |
| 11:44:39 | bauzas | but you can write an online file and set 0 into it for core #0, right ? | |
| 11:45:13 | bauzas | that's quite a destructive CPU equivalent of disk's rm -rf / | |
| 11:45:25 | bauzas | except it's stateless | |
| 11:49:40 | sean-k-mooney | it will be ignored | |
| 11:49:55 | sean-k-mooney | you can create that file but it wont do anything as far as i am aware | |
| 11:50:52 | sean-k-mooney | by the way https://github.com/SeanMooney/arbiterd/blob/master/src/arbiterd/common/cpu.py#L58-L62 | |
| 11:51:02 | sean-k-mooney | is how i got the aviable cpus | |
| 11:52:02 | sean-k-mooney | gibi: bauzas actully also for context | |
| 11:52:04 | sean-k-mooney | https://github.com/SeanMooney/arbiterd/blob/master/src/arbiterd/common/cpu.py#L106-L107 | |
| 11:52:14 | sean-k-mooney | that default 1 was because of this | |
| 11:52:19 | bauzas | gtk | |
| 11:52:24 | sean-k-mooney | https://github.com/SeanMooney/arbiterd/blob/master/src/arbiterd/common/cpu.py#L106-L114 | |
| 11:52:35 | sean-k-mooney | thats also why i did the get_online check in set online | |
| 11:52:46 | sean-k-mooney | to not do the write | |
| 11:52:59 | sean-k-mooney | i proably should have left a code comment for that... | |
| 11:53:26 | sean-k-mooney | ye removed that optimisation but that was actully the real reason i did that in the poc. it was not an optimisation | |
| 11:53:37 | sean-k-mooney | it was to workaround the cpu0 weridness | |
| 11:54:28 | sean-k-mooney | although to be fair i didnt fix that on the set offline path | |
| 11:54:30 | sean-k-mooney | so meh | |
| 12:04:01 | opendevreview | Sylvain Bauza proposed openstack/nova master: libvirt: let CPUs be power managed https://review.opendev.org/c/openstack/nova/+/821228 | |
| 12:04:02 | opendevreview | Sylvain Bauza proposed openstack/nova master: Enable cpus when an instance is spawning https://review.opendev.org/c/openstack/nova/+/868237 | |
| 12:04:06 | bauzas | gibi: sean-k-mooney: ^ | |