Earlier  
Posted Nick Remark
#openstack-nova - 2020-10-23
14:25:17 sean-k-mooney it is not ment to the cpus multiplied by the allocation ration
14:25:25 stephenfin That's wrong
14:25:34 sean-k-mooney no that is how it was ment to work
14:25:38 sean-k-mooney and how it has always worked
14:25:48 sean-k-mooney it has never used the allocation ration
14:25:48 stephenfin Create a load of guests with 'hw:numa_nodes=1' on two node host with 6 CPUs per node
14:26:00 stephenfin How many host CPUs are free if you boot those guests?
14:26:29 stephenfin The os-hypervisors information makes no sense in a world of overcommit
14:26:43 stephenfin which is why we're planning to drop it in Wallaby
14:26:55 sean-k-mooney stephenfin: recyclehero is try ing to boot a 10 vcpu guest on an 8 cpu host
14:27:01 sean-k-mooney which we have never supported
14:27:19 sean-k-mooney because 2 gust cpus cant consome the same host cpu enven if they are floating
14:27:23 stephenfin sean-k-mooney: yes, that's true
14:28:11 stephenfin sean-k-mooney: but we have 'max_unit' to track that
14:29:01 stephenfin recyclehero: Oh, that was an interesting point on the output of e.g. 'openstack resource provider inventory list'
14:29:18 stephenfin I'd have expected the total column to show the total inclusive of allocation_ratio
14:29:41 stephenfin but it doesn't; it shows the total before the allocation ratio is applied
14:29:41 sean-k-mooney no
14:29:50 stephenfin so you have to do that maths yourself :)
14:29:57 sean-k-mooney yes
14:30:06 stephenfin sean-k-mooney: make up your mind, man!
14:30:07 stephenfin :P
14:30:09 sean-k-mooney so that the different values are not depenent on each other
14:30:17 sean-k-mooney you have to do the math yourself
14:30:34 sean-k-mooney multipleying total by allocation ratio - reserved
14:32:13 stephenfin sean-k-mooney: as for that earlier issue, this entire stack is failing with it :( https://review.opendev.org/#/q/topic:bp/compact-db-migrations-wallaby
14:32:29 stephenfin I don't think it's something I've done wrong, but I'm waiting to see if another patch fails first
14:32:55 sean-k-mooney stephenfin: the import issue if its the same one it due to an oom issue causeing the import to fail
14:33:01 sean-k-mooney if not then rbd is not isntalled
14:33:12 sean-k-mooney because ceph-common is not installed
14:33:20 stephenfin oh, no, look down
14:33:22 sean-k-mooney i didnt look too closelon
14:33:32 stephenfin ignore the rbd thing
14:33:37 stephenfin that's lyarwood's work
14:34:59 sean-k-mooney oh line 861
14:35:18 sean-k-mooney it did not scole down fro some reason
14:35:36 sean-k-mooney TypeError: Parameterized generics cannot be used with class or instance checks
14:35:42 stephenfin Yeah
14:35:47 stephenfin I've never seen that before
14:35:52 stephenfin it's buried in eventlet too
14:36:26 sean-k-mooney so its coming from the ytpeing model
14:37:15 stephenfin mypy hasn't been updated recently though, and typing is stdlib now
14:37:18 sean-k-mooney form the libvirt treadpool proxy
14:37:31 sean-k-mooney stephenfin: yes it si from the standardlib
14:37:48 stephenfin crap, new eventlet release yesterday
14:38:13 sean-k-mooney to fix dnspython support?
14:38:23 sean-k-mooney we likely need to cap it untill we figure this out
14:39:39 stephenfin gibi: Looks like we might have a problem with the new release of eventlet ^
14:39:50 stephenfin https://zuul.opendev.org/t/openstack/build/c85b0734fb524eefb67449105528b0a2/log/logs/screen-n-cpu.txt#861
14:40:18 sean-k-mooney https://github.com/eventlet/eventlet/commit/9d731373e2325588a33229b425100ad538eed766
14:40:23 sean-k-mooney this is apparently what changed
14:42:04 sean-k-mooney oh they fixed my ssl bug i think https://github.com/eventlet/eventlet/commit/7b1aa58d77550caabba3ca2e0b1bb79626c83d9c
14:42:35 sean-k-mooney oh no its a different one
14:43:55 sean-k-mooney https://github.com/eventlet/eventlet/issues/632 was our issue
14:44:27 stephenfin yeah, the last working build for https://review.opendev.org/#/c/758396/ used eventlet===0.26.1
14:44:35 stephenfin https://zuul.opendev.org/t/openstack/build/bf1f2b730848408b99c61e82068eb73b/log/job-output.txt#6355
14:44:46 stephenfin the failing one uses eventlet===0.28.0
14:44:50 stephenfin https://zuul.opendev.org/t/openstack/build/00e3a2793a1040b29b944b944c9da795/log/job-output.txt
14:44:59 stephenfin so eventlet has borked us \o/
14:45:04 sean-k-mooney they have done a buch of release in the last month
14:46:02 sean-k-mooney i guess need to file a bug
14:52:43 stephenfin Hmm, we also bumped our version of mypy, greenlet and libvirt-python recently, so it could be those too. This will be fun
15:24:08 recyclehero sean-k-mooney stephenfin thanks guys. I really appreciate ur help. I guess it was okay after we did the 'None' patch in OSA but the data displayed was confusing.
15:24:51 sean-k-mooney it was a design choice that they made very early on in horizon
15:25:10 sean-k-mooney they wanted ti to deisplay your actual resouce avaiablyt not your over commit availablity
15:25:46 sean-k-mooney im sure it could be changed but it would be a change in behavior of what that dashboard shows
15:26:47 recyclehero but anyway IMO 10 out of 8 is not a good thing to show especially in red. its like ur doing something wrong
15:27:14 sean-k-mooney its in red because you are over commiting
15:27:22 sean-k-mooney which in maycases is wrong
15:27:43 sean-k-mooney for disk or memory its potentially dangrous
15:27:53 sean-k-mooney disk can cause data loss and vm hangs
15:28:08 sean-k-mooney memory without swap can cause the vms to be killed
15:28:43 sean-k-mooney cpu if its too high will also eventually cause nova to stop working properly due to to high load
15:29:28 recyclehero BTW I am using kernel shared memory + lots of swap. i hope it works well
15:29:54 sean-k-mooney if you are using non numa guest then it can
15:30:28 recyclehero with this and using same distros for instances I hope I can get near 2.0 for memory
15:30:31 sean-k-mooney recyclehero: you need to have enough swap to fully account for the over commit
15:30:51 recyclehero lots of swap ssd
15:30:54 sean-k-mooney so you have allcoated at least 2 times your memroy for swap
15:32:02 recyclehero equal memory wont work. I have 12GB Ram + 12 GB swap. you are suggesting 24GB swap for allocation ratio of 2.0?
15:32:04 sean-k-mooney the guest uses more memory then in the falvor for the qemu process to use and back device es with
15:32:16 sean-k-mooney recyclehero: yes
15:32:22 sean-k-mooney you shoudl have 24G
15:32:28 sean-k-mooney of swap
15:32:46 sean-k-mooney you could go a little lowwer but you wotn get to 2.0 with only 12G
15:33:03 sean-k-mooney you will hit oom issues
15:33:39 sean-k-mooney you have not accounted for the qemu overhead or nova/os memory usage
15:33:47 recyclehero good to know. I should I make changes while its a fresh deployement.
15:34:05 recyclehero I have counted nova/os memory in the reserverd.
15:34:24 recyclehero but not for the qemu overhead
15:34:51 sean-k-mooney the qemu over head can be small or lager depening on how my cores or devices an instance has
15:35:11 sean-k-mooney it can be as small as ~100MB or well over a GB
15:35:35 sean-k-mooney per instance
15:36:25 sean-k-mooney so you need to caulate what the maxium amount of overhead will be for you smallest and larges flavor by booting them then figure out how many of them will run on each host and set teh reserved value accourdignly
15:36:49 sean-k-mooney or allocate enough swap to copenstate
15:39:18 recyclehero my strategy is reserved for nova+os and swap to conpensate for overcommiting
15:40:00 sean-k-mooney ya but again that is ignoring qemu overhead
15:40:21 sean-k-mooney which is ment to be acounted for in reserved or swap
15:40:45 sean-k-mooney in generall i advise never overcommiting memroy or disks for what its worth
15:40:50 sean-k-mooney only cpus

Earlier   Later