| Posted | Nick | Remark | |
|---|---|---|---|
| #openstack-nova - 2020-12-10 | |||
| 17:34:18 | sean-k-mooney | bauzas: but im not sure we want it there by default | |
| 17:35:14 | sean-k-mooney | https://github.com/openstack/nova/blob/master/nova/tests/fixtures.py#L1891-L1905 | |
| 17:35:43 | sean-k-mooney | personally i would not add it there since most deployment wont have it and just add it in your test class | |
| 17:35:56 | bauzas | ack thanks | |
| 17:36:14 | bauzas | sorry I had to drop because of an Amazon delivery for my new headset \o/ | |
| 17:36:25 | sean-k-mooney | oh nice | |
| 17:38:15 | JamesBenson | sean-k-mooney: originally I had tried `custom` and posted `cpu_models = nehalem,westmere` & `cpu_model_extra_flags = pcid, vmx, pdpe1gb` but then I couldn't boot up VM's. Those three flags are the ones of interest, but aren't in both models. I thought that the way mentioned above was the right way, but couldn't get it working. Maybe the flags need to be there for all of them? | |
| 17:45:18 | sean-k-mooney | JamesBenson: that should be valid | |
| 17:45:29 | sean-k-mooney | pdep1GB is for 1G hugepages | |
| 17:45:49 | sean-k-mooney | nehalem does not actully support that if i remeber correctly | |
| 17:46:01 | sean-k-mooney | you need host supprot to use it | |
| 17:46:06 | JamesBenson | correct, nehalem doesn't support any of those three. | |
| 17:46:15 | JamesBenson | but westmere does | |
| 17:46:19 | sean-k-mooney | wel it support vmx | |
| 17:46:34 | sean-k-mooney | not in the model defieniotn but in the cpu | |
| 17:47:02 | sean-k-mooney | wehn using multiple models the flags must be valid for all models in the list | |
| 17:47:59 | sean-k-mooney | on your westmer host that should work | |
| 17:48:01 | JamesBenson | ah, okay, so that's the trick then. And just verifying, to check for the flags, I just issued `cat /proc/cpuinfo` | |
| 17:48:07 | sean-k-mooney | but on your nehalem host it will fail | |
| 17:48:24 | JamesBenson | let me double check the flags on my hosts. | |
| 17:48:27 | sean-k-mooney | yep cat /proc/cpuinfo or use lscpu | |
| 17:49:05 | sean-k-mooney | for vmx you need to enable nested virt too in the hosts kernel | |
| 17:50:08 | JamesBenson | okay, yeah only vmx on the nehalem, westmere has all three. | |
| 17:50:16 | sean-k-mooney | cat /sys/module/kvm_intel/parameters/nested | |
| 17:50:18 | sean-k-mooney | Y | |
| 17:50:34 | JamesBenson | just change that to Y? | |
| 17:50:44 | sean-k-mooney | Y or 1 | |
| 17:51:18 | sean-k-mooney | that is showing the current nested virt status | |
| 17:51:25 | sean-k-mooney | if that is N or 0 then you dont have it enabled | |
| 17:51:57 | JamesBenson | okay, cool. So which is better: cpu_models = nehalem,westmere with the vmx flag or host-model? Or should I get a list of flags that are on both and put all of the flags in there? | |
| 17:52:11 | JamesBenson | It is an `N` | |
| 17:52:31 | JamesBenson | Thanks for the help & insight be the way. Super helpful :-) | |
| 17:52:34 | sean-k-mooney | ya so you need to drop a file into /etc/modeprobe.d | |
| 17:53:41 | sean-k-mooney | what i would do is as follows | |
| 17:54:20 | sean-k-mooney | on the nehelm host set cpu_models = nehalem and cpu_model_extra_flags = vmx | |
| 17:54:53 | sean-k-mooney | on the westmear host do cpu_model_extra_flags = pcid, vmx, pdpe1gb cpu_models = nehalem,westmere | |
| 17:55:22 | sean-k-mooney | in the flavor then add a required trait for pdep1gb on flavor that use 1G hugepages | |
| 17:56:12 | sean-k-mooney | JamesBenson: the feature flags are reported as traits to placement so you can express what you need in both the flavor and image | |
| 17:56:22 | JamesBenson | okay. Is this the recommended path for mixed architectures to get the most out of them? | |
| 17:56:38 | JamesBenson | great, we have windows VM's so that's what I'm hoping to speed up with using those flags. | |
| 17:57:55 | sean-k-mooney | normaly you would use something like host aggreates. PCID should acclerate gust where you have mitigation enabled somewhat | |
| 17:58:32 | sean-k-mooney | vmx is only useful if those windows instnace have hyperv or are using docker with linux instance | |
| 17:59:18 | sean-k-mooney | on host without pdep1gb you can still use 2MB hugepages | |
| 17:59:33 | sean-k-mooney | so that is what i advise on nehelem | |
| 18:01:03 | sean-k-mooney | JamesBenson: http://paste.openstack.org/show/800947/ | |
| 18:01:18 | sean-k-mooney | you shoudl have a file that looks like that or you can create one | |
| 18:01:43 | sean-k-mooney | options kvm_intel nested=1 is what you need to uncomment unless you have a kernel newwer then 4.19 | |
| 18:01:50 | sean-k-mooney | in which case its the default value | |
| 18:02:15 | sean-k-mooney | although your proably want a kernel newer then 5.1 unless it has backports | |
| 18:02:36 | JamesBenson | freshly deployed ubuntu 18.04 with 4.15.0 | |
| 18:03:04 | sean-k-mooney | i would sugges using the 5.3 or 5.4 kernel they provide form focal instead | |
| 18:03:21 | sean-k-mooney | its the hwe kernel | |
| 18:03:52 | sean-k-mooney | https://wiki.ubuntu.com/Kernel/LTSEnablementStack | |
| 18:05:12 | JamesBenson | okay :-) | |
| 18:05:39 | sean-k-mooney | out of interst is tehre a reason you are not usin g2.04 | |
| 18:05:45 | sean-k-mooney | *20.04 | |
| 18:05:59 | sean-k-mooney | if its a fresh install | |
| 18:06:43 | JamesBenson | kolla doesn't support it I think fully | |
| 18:06:52 | JamesBenson | That's my deployment method. | |
| 18:07:16 | sean-k-mooney | yep i use kolla too. althouhg im currentoy using centos kind of | |
| 18:07:42 | sean-k-mooney | JamesBenson: depending on the release they have different levels of supprot for 20.04 | |
| 18:08:20 | JamesBenson | yeah, I'm on ussuri, I thought victoria was the first one to support 20 fully. | |
| 18:09:15 | sean-k-mooney | JamesBenson: yes it is | |
| 18:09:48 | sean-k-mooney | JamesBenson: i guess they have not had there offical release yet technically | |
| 18:09:49 | JamesBenson | ah, well, then I can definitely try that out. We do it all, including the OS deploy, in a CI/CD environment, so easy enough ;-) | |
| 18:10:15 | JamesBenson | yeah, not yet, soon I think | |
| 18:10:27 | JamesBenson | but for victoria, we do plan on pushing to ubuntu 20.04 | |
| 18:10:55 | JamesBenson | but I'll see about bumping the kernel, I've just been using what it deploys with, having tried to get it to a newer version. | |
| 18:11:32 | sean-k-mooney | the nested virt support signifcantly improve after kernel 4.19 so if you use it i always recommend having 4.19 and preferbly 5.1+ as you min kernel | |
| 18:12:11 | sean-k-mooney | JamesBenson: the hwe kernel works well for nested virt without needing the os bump | |
| 18:12:17 | sean-k-mooney | its 5.3 i belive | |
| 18:12:33 | JamesBenson | cool, have you ever tried low-latency kernels in openstack? | |
| 18:13:41 | sean-k-mooney | i have but my experice was tainted. i deployed with the low-latency kernel but i later found out that my system hangs which i attibuted to it was due to 3 failing dimms | |
| 18:13:45 | ganso | elod, bauzas, lyarwood: Hi! do you think those patches could be allowed to be backported from rocky to queens: https://review.opendev.org/c/openstack/nova/+/580720 and https://review.opendev.org/c/openstack/nova/+/583715 ? | |
| 18:14:05 | sean-k-mooney | JamesBenson: so i have not really looked at it fairly. | |
| 18:14:32 | sean-k-mooney | JamesBenson: my deployemtn is just my home test/dev cloud so its currently runing the mainline 5.9 kernel | |
| 18:15:25 | sean-k-mooney | i was hoping to move to the 5.10 upstream lts and do a ussuri to victoria upgrade over the holiday break but that is still tbd | |
| 18:15:40 | JamesBenson | okay, looks like linux-lowlatency-hwe-18.04 is 5.4.0.53.59 | |
| 18:15:49 | JamesBenson | so maybe I'll try that out :-) | |
| 18:17:03 | sean-k-mooney | cool with that you wont need t o set the modeprobe paramter eitehr since its enabel by default | |
| 18:17:09 | ganso | elod, bauzas, lyarwood, nevermind I kind of got the answer myself by cherry-picking and seeing that queens is missing 2 other RequestSpec updates | |
| 18:17:46 | sean-k-mooney | ganso: we cant backport modifcation to any nova objects | |
| 18:18:07 | ganso | sean-k-mooney: thanks! | |
| 18:18:10 | sean-k-mooney | ganso: well we cant backport new fields | |
| 18:18:32 | sean-k-mooney | we can backport change that dont modify the serialsed format | |
| 18:20:12 | sean-k-mooney | ganso: you wanted this in queens | |
| 18:20:30 | ganso | sean-k-mooney: yes, in queens | |
| 18:20:57 | ganso | sean-k-mooney: since the fix requires the new field, we wouldn't be able to adapt it to queens, as I understand | |
| 18:22:09 | sean-k-mooney | that is correct unfortunetly however we coudl do something else possibly | |
| 18:22:33 | sean-k-mooney | if the places that need is_bfv hae a context and instnace tehy could use is_volume_backed_instance(ctxt, instance) instead directly | |
| 18:24:28 | ganso | sean-k-mooney: hmmm I will investigate this in the code. Thanks! Although, I'm aware it would make the backport with new code with extra risk, and then the stable maintainers may disagree with merging it | |
| 18:24:56 | sean-k-mooney | yes you would basically have to modeify https://review.opendev.org/c/openstack/nova/+/580720/6/nova/scheduler/utils.py | |
| 18:25:26 | sean-k-mooney | unfortunetly that does not have th instance or a keystone context | |
| 18:25:40 | ganso | oh I see | |
| 18:25:44 | ganso | yea, bummer | |
| 18:25:56 | ganso | thanks again sean-k-mooney ! =) | |
| 19:44:54 | melwitt | gibi, lyarwood: yes I can take over babysitting those patches once yall are gone | |
| 20:10:25 | openstackgerrit | Lance Bragstad proposed openstack/placement master: Implement secure RBAC for resource providers https://review.opendev.org/c/openstack/placement/+/760240 | |
| 21:21:33 | JamesBenson | sean-k-mooney, one follow up question. Is there a setting (or benefit) to passing all of the CPU flags in the nova.conf? | |
| 21:50:29 | lyarwood | elod: https://review.opendev.org/q/I8f24b839bf42e2fb9803dc7df3a30ae20cf264eb lets see what that breaks | |