Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2023-11-21
17:22:59 bauzas if we magically make the feature opt-out, then we leave operators responsible of ensuring that their kernel can support cpu offline
17:23:03 gibi I don't see how that implies that the feature cannot be enabled with zero PCPUs and expect to do nothing
17:23:24 gibi bauzas: I'm not changing the default of the flag. It is still opt in
17:23:40 gibi the operator needs to explicitly enable the feature
17:23:59 gibi with [libvirt]cpu_power_management
17:24:03 gibi with [libvirt]cpu_power_management = True
17:25:20 bauzas touché
17:25:50 gibi what changes is that the operator can enable [libvirt]cpu_power_management = True before decides about the PCPU list of the given host
17:26:24 bauzas I mixed different conversations, one being not public
17:26:30 gibi this allows to enable [libvirt]cpu_power_management globally in a cloud in a deployment engine, and set PCPU config on the hypervisor independetly later
17:27:32 bauzas gibi: given the latter requires a nova-compute restart, I don't really see the benefits of enabling the former before, except if you have a deployment tool and you don't want your users to change that option
17:27:35 sean-k-mooney bauzas: are there specific risks or concerns you have with this change?
17:28:04 sean-k-mooney bauzas: i would eventully like to make [libvirt]cpu_power_management = True the default in nova by the way
17:28:09 bauzas sean-k-mooney: none for the master patch itself
17:28:17 gibi bauzas: there will be a single restart but there will be indedpendent decisions i) I need power management ii) I have a list of PCPUs used in a given hypervisor
17:28:18 bauzas hence my +1
17:28:29 sean-k-mooney but i dont want to require cpu_dedicated_set
17:28:55 sean-k-mooney ok so your concern is really just with framing this as a bug
17:29:07 gibi and if the decision #ii) result in an empty cpu_dedicated_set then I don't want to go back to decision i) as that is still valid
17:29:09 bauzas well, power management on non-dedicated cores is currently not something existing
17:29:28 sean-k-mooney right but this has nothign to do with that
17:29:29 bauzas sean-k-mooney: this is correct, as I wrote in both gerrit and here
17:29:37 gibi true, but power management on zero PCPUs actually work
17:29:48 gibi and help keeping the two configuration decision independent
17:30:15 bauzas if we were enabling power management on shared cores, surely we would make those options independent
17:30:20 gibi take my space heater example
17:30:43 gibi the on/off button and the temp knob are independent decisions
17:30:45 sean-k-mooney if we were to enabel it on shared cpus we woudl have to only supprot the governer policy
17:30:55 sean-k-mooney we coudl not supprot the cpu state policy
17:31:09 sean-k-mooney but even then i would expect new config optiosn for the cpu share set
17:31:17 gibi I don't want to alway use the on/off button when my room becomes cold. I want to rely on the combination of the two config to dyanmically heat my room
17:31:21 sean-k-mooney im not sure we will add that funcitonaltiy
17:31:31 sean-k-mooney at least right now i dont see a uescase to do that
17:32:51 bauzas you know what ? I'll stop
17:34:26 gibi I stop too. See you tomorrow
17:35:14 bauzas I sent it to... wait, actually nothing
17:35:20 bauzas the check pipeline failed
17:35:40 bauzas but I added a label called 'Workflow +1'
17:35:42 bauzas fwiw
17:36:02 gibi bauzas: thanks
17:40:29 bauzas (fwiw, the analogy with a room heater seems wrong to me : the on/off button seems to me the opt-in flag about power management, and the temp knob is the strategy)
17:41:21 bauzas and when I wrote the feature, I thought it was errorprone to tell "I want my cores to be power managed" if I wasn't providing dedicated cores
17:42:00 bauzas but I don't want to debate for so long about that, so let's ship it
17:42:02 sean-k-mooney the temp knob is hte cpu_dedicated_set (the quantiy of cpus) but i get your perspective too
17:44:21 bauzas if some deployment tool automatically enables the power management feature, it still needs to properly document the fact that the strategy can be modifed
17:44:42 bauzas maybe some operators would be surprised to see their cores offlined
17:44:59 bauzas (at least I would)
17:45:24 sean-k-mooney well that kind of raise the question of what the defautl behavior shoudl be upstream
17:45:25 bauzas and maybe it wouldn't hit me that it was nova which offlined them for my usage
17:45:34 sean-k-mooney i think we eventually woudl want to make that the default
17:45:47 sean-k-mooney perhasp even this cycle although that should be a seperate patch
17:46:10 bauzas sean-k-mooney: the problem here is that I don't want us to buy a feature that depends on a kernel support
17:46:42 sean-k-mooney this has been supported in the kernel for a very very long time, although you ar ecorrect that it can be compiled out
17:46:46 bauzas I don't know for exotic archs and kernels, but I don't wanna bet that all of them have the sysfs calls
17:46:55 sean-k-mooney or disabeld such as when in a vm
17:47:20 sean-k-mooney so you dont think we shoudl enabel by default in the long term then?
17:47:52 sean-k-mooney its certnely not somethign we have to rush into
17:47:53 bauzas I have some concerns about it and I need to be convinced that nova wouldn't stop supporting exotic kernels just because of that
17:48:17 sean-k-mooney kind of like being secure by default i feel we shoudl try to be energy effiect by default
17:48:36 bauzas sure, and that makes sense
17:48:39 sean-k-mooney bauzas: well you could alwasy trun it off or we coudl try and detect it
17:48:49 bauzas we should even
17:48:58 sean-k-mooney prehasp we woudl intoduce an auto value in the future
17:49:10 bauzas detecting whether you can disable them or not at startup
17:49:21 sean-k-mooney ya
17:49:23 bauzas yeah, this sounds a right way
17:49:26 bauzas the auto thing
17:49:39 sean-k-mooney so if mede it a tri state true,false,auto
17:49:44 sean-k-mooney and then did the detection
17:49:51 sean-k-mooney that would be the best of both worlds
17:50:08 sean-k-mooney and it solves the ux problem but i also agree that would not be backportable
17:50:10 bauzas yeah, and this would convince me of opting-out this
17:50:25 bauzas provided the autodetect
17:51:14 sean-k-mooney do we ant to wait to disucss this with gibi tomorrow before proceeding with the current patch
17:54:33 bauzas this is additive
17:54:40 bauzas and as you said, not backportable
19:37:56 opendevreview Andriy Kurilin proposed openstack/python-novaclient master: Disable NEUTRON_ENFORCE_SCOPE at function job https://review.opendev.org/c/openstack/python-novaclient/+/901621
20:41:56 opendevreview Elod Illes proposed openstack/nova stable/yoga: [stable-only] Remove nova-emulation from check pipeline of Yoga https://review.opendev.org/c/openstack/nova/+/901604
20:44:05 elodilles sean-k-mooney: i forgot about the [stable-only] tag /o\ so now i've moved the job to experimental pipeline as you originally wanted ^^^
21:06:06 opendevreview Artom Lifshitz proposed openstack/nova master: Allow live migrate paused instance when post copy is enabled https://review.opendev.org/c/openstack/nova/+/444517
21:09:41 artom melwitt, since you looked at it previously and sean-k-mooney now has a +2, can I bother you for https://review.opendev.org/c/openstack/nova/+/883682?
21:11:20 andreykurilin Hi folks! python-novaclient-functional job became broken after neutron-related change to devstack https://review.opendev.org/c/openstack/devstack/+/899306 . I have a fix to unblock novaclient’s CI - https://review.opendev.org/c/openstack/python-novaclient/+/901621 but IDK whether the issue should be addressed differently. The review is appreciated.
21:16:28 melwitt artom: sure, will look
#openstack-nova - 2023-11-22
01:14:20 opendevreview Artom Lifshitz proposed openstack/nova master: Allow live migrate paused instance when post copy is enabled https://review.opendev.org/c/openstack/nova/+/444517
07:55:33 opendevreview Merged openstack/nova master: Allow enabling cpu_power_management with 0 dedicated CPUs https://review.opendev.org/c/openstack/nova/+/901188
08:29:31 opendevreview likui proposed openstack/nova master: Add py311 in tox file https://review.opendev.org/c/openstack/nova/+/901587
08:38:10 opendevreview likui proposed openstack/nova master: Add py311 in tox file https://review.opendev.org/c/openstack/nova/+/901587
09:11:28 opendevreview Pierre Libeau proposed openstack/nova master: nova-condutor put instance in error state https://review.opendev.org/c/openstack/nova/+/901655
09:22:16 opendevreview Balazs Gibizer proposed openstack/nova stable/2023.2: Do not manage CPU0's state https://review.opendev.org/c/openstack/nova/+/901566
09:22:39 opendevreview Balazs Gibizer proposed openstack/nova stable/2023.1: Do not manage CPU0's state https://review.opendev.org/c/openstack/nova/+/901567
09:24:21 gibi elodilles bauzas: ^^ power management bugfix backport ^^
09:25:14 gibi bauzas: this is a differnt bugfix
09:25:28 bauzas oh shit, you're tight
09:25:31 bauzas right*
09:26:37 bauzas ETOOMANYTHINGSINMIND
09:26:41 bauzas +2d
09:26:48 bauzas (the bobcat one)
09:28:31 gibi thanks!
09:29:59 opendevreview Balazs Gibizer proposed openstack/nova stable/2023.2: Allow enabling cpu_power_management with 0 dedicated CPUs https://review.opendev.org/c/openstack/nova/+/901656
09:30:13 gibi bauzas: now it is the time to look away ^^ ;)

Earlier   Later