| Posted | Nick | Remark | |
|---|---|---|---|
| #openstack-nova - 2022-06-14 | |||
| 09:35:44 | gibi | do you have the PGPU RP in the tree of in_tree=adbdb144-d84b-4b80-b03b-a0bc520d91ba ? | |
| 09:36:29 | gibi | so you requests two vgpus from two different pGPUs (isolate) | |
| 09:36:42 | gibi | I mean one vgpu from each pgpu | |
| 09:37:53 | gibi | that should work if you have to pGPU RPs | |
| 09:39:04 | gibi | could you paste the output of https://github.com/gibizer/osc-placement-tree ? (or publish the test you are running so I can reproduce it?) | |
| 09:41:33 | bauzas | sorry was doing other thing | |
| 09:42:10 | bauzas | gibi: I can install it in the func venv | |
| 09:42:37 | bauzas | oh, you said it in the README :) | |
| 09:42:45 | bauzas | lemme pip it | |
| 09:48:11 | bauzas | gibi: I need to disappear for gym reasons but I'll work on it this afternoon | |
| 09:48:29 | bauzas | I just installed, I just need to add the placement api module in my test class | |
| 09:48:39 | gibi | bauzas: have a nice workout. feel free to ping me later | |
| 09:51:18 | bauzas | gibi: <3 | |
| 09:51:21 | bauzas | gibi: found my problem | |
| 09:51:30 | bauzas | I was only having one child RP | |
| 09:51:50 | gibi | ack, good to hear that | |
| 09:55:14 | gibi | bauzas, sean-k-mooney: an interesting performance bug https://bugs.launchpad.net/nova/+bug/1978372 | |
| 09:59:49 | sean-k-mooney | first tought is they dont know how to use the hw:numa_* extra specs | |
| 10:00:38 | sean-k-mooney | its very bad pratcit to use hw:numa_cpus or hw:numa_mem with symetric numa toplogies | |
| 10:00:46 | sean-k-mooney | its explictly an anti pattern | |
| 10:00:59 | sean-k-mooney | and should not be doen | |
| 10:01:38 | sean-k-mooney | it wont affect the perfrmace but the also mis understand the relation ship betweeh hw:cpu_threads and hw:cpu_max_threads | |
| 10:01:52 | sean-k-mooney | if hw:cpu_treads is specifed then max is ignored | |
| 10:02:03 | sean-k-mooney | sam for sockets | |
| 10:02:32 | sean-k-mooney | so that flavor defintion annoys me on a fundimental level :) | |
| 10:03:45 | sean-k-mooney | gibi: hoststly im not that suprised we know that this is a very slow algorthim | |
| 10:03:50 | gibi | then look at the minimal reproduction that has no falvor :) | |
| 10:04:17 | sean-k-mooney | i did some experiments a few year ago repimlementing it form scratch | |
| 10:04:39 | sean-k-mooney | but the reason i did not continue with that was we were going to implement numa in placement any day now | |
| 10:04:55 | gibi | also we are emitting 18G debug logs during that run which is pretty extreme | |
| 10:05:35 | sean-k-mooney | peopel complaied there was not enough logging in the hardware module :P | |
| 10:06:06 | sean-k-mooney | https://github.com/SeanMooney/cpu-pinning | |
| 10:06:13 | sean-k-mooney | this was my experimal approch | |
| 10:08:31 | gibi | as far as I understand this call does not call out to any extrnal system (no fs, db, rabbit calls) so it runs in pure python. So taking minutes indicate that we are doing something crazy | |
| 10:10:59 | sean-k-mooney | we are try to validate cpu (floating and pinned),ram ,pci, and pmem affinity based on the rules set in teh falvor with regards to thread affintiy, device affintiy and any semetic or asmetirc numa requirements | |
| 10:16:32 | sean-k-mooney | gibi: by the way we have know about this basically from the start that its qudratic or wrose as you scale the number of host/guest numa nodes | |
| 10:16:50 | sean-k-mooney | its why the numa toploty filter shoudl always be last in your filter list if you enable it | |
| 10:19:28 | sean-k-mooney | i am not sure if ic an repodcue that simpel example in my ohter implemntaion but im going to give it a try quickly | |
| 10:28:30 | sean-k-mooney | [11:27:50]❯ ./run.sh | |
| 10:28:32 | sean-k-mooney | Courtesy Notice: Pipenv found itself running within a virtual environment, so it will automatically use that environment, instead of creating its own for any project. You can set PIPENV_IGNORE_VIRTUALENVS=1 to force pipenv to ignore that environment and create its own instead. You can set PIPENV_VERBOSITY=-1 to suppress this warning. | |
| 10:28:34 | sean-k-mooney | ||
| 10:28:36 | sean-k-mooney | 0:0,1:32,2:512,3:544 | |
| 10:28:38 | sean-k-mooney | 0:1,1:2,2:3,3:4 | |
| 10:28:40 | sean-k-mooney | 0:5,1:69,2:133,3:197 | |
| 10:28:42 | sean-k-mooney | 0:33,1:513,2:34,3:514,4:35,5:515,6:36,7:6 | |
| 10:28:44 | sean-k-mooney | 0:2047,1:1983,2:1919,3:1855,4:2015,5:1951,6:1887,7:1823 | |
| 10:28:46 | sean-k-mooney | bug repoducer | |
| 10:28:48 | sean-k-mooney | ---------------------------------------- | |
| 10:28:50 | sean-k-mooney | 0:0,1:3,2:6,3:9,4:12,5:15,6:18,7:21,8:48,9:51,10:54,11:57,12:60,13:63,14:66,15:69,16:1,17:4,18:7,19:10,20:13,21:16,22:19,23:22,24:49,25:52,26:55,27:58,28:61,29:64,30:67,31:70,32:2,33:5,34:8,35:11,36:14,37:17,38:20,39:23,40:50,41:53,42:56,43:59,44:62,45:65,46:68,47:71 | |
| 10:28:52 | sean-k-mooney | real 0m1.004s | |
| 10:28:54 | sean-k-mooney | user 0m0.893s | |
| 10:28:56 | sean-k-mooney | sys 0m0.094s | |
| 10:28:58 | sean-k-mooney | gibi: is ^ better | |
| 10:29:18 | sean-k-mooney | or | |
| 10:29:20 | sean-k-mooney | real 0m0.986s | |
| 10:29:22 | sean-k-mooney | user 0m0.887s | |
| 10:29:24 | sean-k-mooney | sys 0m0.078s | |
| 10:29:26 | sean-k-mooney | if i jus trun the repoducer | |
| 10:32:24 | gibi | the reproducer took 6 mins to run on my laptop | |
| 10:32:36 | gibi | so yours seems to be a loooot faster | |
| 10:34:17 | sean-k-mooney | https://github.com/SeanMooney/cpu-pinning/commit/018a1d0a9abeeec40d3e3ccbfd640363158433c5 | |
| 10:34:28 | sean-k-mooney | that is the same case right | |
| 10:35:07 | sean-k-mooney | 1 socket 16 numa nodes and 48 cores with 2 thread pre cpu for 96 theads | |
| 10:35:29 | sean-k-mooney | then booth a 48 core cpu with pinning and the prefer thread polciy with 1G hugepages | |
| 10:35:42 | sean-k-mooney | and 488GB of ram | |
| 10:35:42 | gibi | there are some usage too in the reproducer | |
| 10:36:32 | gibi | the first 7 host numa cell is used | |
| 10:36:55 | sean-k-mooney | ack i can add that and see if it makes any differnce | |
| 10:50:50 | sean-k-mooney | basically the same | |
| 10:50:52 | sean-k-mooney | real 0m1.074s | |
| 10:50:54 | sean-k-mooney | user 0m0.936s | |
| 10:50:56 | sean-k-mooney | sys 0m0.115s | |
| 10:51:30 | sean-k-mooney | https://github.com/SeanMooney/cpu-pinning/blob/master/pinning.py#L378-L415= | |
| 10:52:25 | sean-k-mooney | the issue with my alternitiv implemntation is it only supprots hugepages and cpu pinning | |
| 10:52:36 | sean-k-mooney | it is missing pci supprot and mixed cpu support | |
| 10:53:48 | sean-k-mooney | ill run there repodcuer to have a comparison on the same hardware | |
| 10:59:21 | sean-k-mooney | im going to make coffee whiel this runs... | |
| 11:03:42 | gibi | dansmith, bauzas: what do you think do we want to / can do something about this? https://bugs.launchpad.net/nova/+bug/1978549 It feels like a bug but I'm not sure I want to go back and add a db migration to stein in nova or touch the placement db init script. https://bugs.launchpad.net/nova/+bug/1978549 | |
| 11:06:20 | sean-k-mooney | gibi: dont we tend to not drop the columes right away | |
| 11:06:29 | sean-k-mooney | to cater for rolling upgrades | |
| 11:06:48 | gibi | sure | |
| 11:06:52 | sean-k-mooney | we stop the usage and then a few release later we can drop the column | |
| 11:07:01 | gibi | but then we suddenly forget to add it to the placement db init script | |
| 11:07:27 | gibi | so if you have a DB that was created before placemetn was moved out of nova repor then you have can_host | |
| 11:07:35 | sean-k-mooney | well that woudl only be an issue if you were initing on a version that used it right | |
| 11:08:23 | gibi | we removed the can_host column but without a DB migration | |
| 11:08:38 | gibi | the removal happen when the new db init was created for the moved out placement | |
| 11:08:46 | sean-k-mooney | i see | |
| 11:10:18 | sean-k-mooney | gibi: real 12m32.814s | |
| 11:10:38 | sean-k-mooney | maybe my 4 year old laptop is starting to show its age | |
| 11:11:55 | sean-k-mooney | gibi: we could just document this as a workaround for the db issue | |
| 11:12:27 | sean-k-mooney | the fix is only requried if movign form in nova to out of tree placement right | |
| 11:12:40 | sean-k-mooney | if you are doign a db restore | |
| 11:12:48 | sean-k-mooney | and even then only if you are migrating db backends | |
| 11:13:16 | sean-k-mooney | if you were just doing a db export and import in mysql it would be fien because the improt would create the tabels | |
| 11:13:46 | sean-k-mooney | in there case they likely are just migrating the data and using placemnt-mange to init an empty db | |
| 11:16:08 | gibi | I agree to only document the workaround. I'm not sure where to put that documentation though. As adding a reno now in zed in placement for an issue introduced in stein in nova does not feels right | |
| 11:16:49 | sean-k-mooney | we can do a stable only reno | |
| 11:16:59 | sean-k-mooney | on stine | |
| 11:17:48 | sean-k-mooney | or update the existing one for speliting out placment which i assuem exists somewhere | |