| Posted | Nick | Remark | |
|---|---|---|---|
| #openstack-cyborg - 2026-06-09 | |||
| 14:08:52 | sean-k-mooney | the order you have them in gerrit is not | |
| 14:08:58 | sean-k-mooney | jgilaber: right | |
| 14:09:07 | sean-k-mooney | we shoudl not need https://review.opendev.org/c/openstack/cyborg/+/991579/2 | |
| 14:09:15 | chandankumar | https://review.opendev.org/c/openstack/cyborg/+/991598 | |
| 14:09:34 | chandankumar | I will do that | |
| 14:09:56 | sean-k-mooney | the main think we need to make it work is https://review.opendev.org/c/openstack/cyborg/+/991598/2/cyborg/agent/manager.py | |
| 14:10:09 | sean-k-mooney | and oblviy uypdatign the fpga dirver to use update internally | |
| 14:10:30 | sean-k-mooney | im sure there are other places in teh manger we need to fix | |
| 14:10:42 | sean-k-mooney | but evenutlly we shoudl nto call any driver spcific method | |
| 14:10:51 | sean-k-mooney | only the methods on the base driver | |
| 14:11:06 | sean-k-mooney | anyway the ci is green | |
| 14:11:08 | jgilaber | the last patch looks ok from a quick glance | |
| 14:11:13 | sean-k-mooney | and you updated tempest as well | |
| 14:11:26 | sean-k-mooney | so if you remove the midel patch i think we are good directionally | |
| 14:11:30 | sean-k-mooney | and can pick this up in gerrit | |
| 14:12:03 | jgilaber | +1, since update is already in the base driver, this looks correct | |
| 14:12:21 | chandankumar | https://github.com/openstack/cyborg/blob/master/cyborg/accelerator/drivers/modules/generic.py#L82 | |
| 14:12:22 | sean-k-mooney | if you read the docsting/comment for it | |
| 14:12:32 | sean-k-mooney | its pretty clear that it was alwasy inteded for progrming the device | |
| 14:12:48 | chandankumar | one more thing https://github.com/openstack/cyborg/blob/master/cyborg/accelerator/drivers/modules/generic.py#L82 and https://github.com/openstack/cyborg/blob/master/cyborg/accelerator/drivers/driver.py also | |
| 14:13:07 | chandankumar | Do we want to keep both of the files? or keep one | |
| 14:13:08 | sean-k-mooney | yes so https://github.com/openstack/cyborg/blob/master/cyborg/accelerator/drivers/driver.py | |
| 14:13:11 | sean-k-mooney | defiens the interface | |
| 14:13:31 | chandankumar | yes | |
| 14:13:40 | sean-k-mooney | we may combine them into one file | |
| 14:13:54 | sean-k-mooney | but lets do that eperatly | |
| 14:13:58 | sean-k-mooney | *seperatly | |
| 14:14:03 | chandankumar | yes that would be much better | |
| 14:14:25 | chandankumar | Ok I will propose a patch after the meeting | |
| 14:14:55 | sean-k-mooney | well we need to decied on the direction | |
| 14:15:02 | sean-k-mooney | my linclition is to delet this entirly | |
| 14:15:08 | sean-k-mooney | but i want to look a the history and usage | |
| 14:15:15 | sean-k-mooney | before we make a code change for it | |
| 14:15:57 | sean-k-mooney | id ont think https://github.com/openstack/cyborg/blob/master/cyborg/accelerator/drivers/modules/generic.py#L82 | |
| 14:15:59 | sean-k-mooney | is currently used | |
| 14:16:19 | sean-k-mooney | so we will liekly delete the enfire https://github.com/openstack/cyborg/tree/master/cyborg/accelerator/drivers/modules directory | |
| 14:16:36 | sean-k-mooney | but as i said we need to investigate if this is infact dead code or not | |
| 14:16:39 | chandankumar | https://review.opendev.org/c/openstack/cyborg/+/473186 - added the generic driver | |
| 14:17:30 | jgilaber | it looks like not all drivers inherit from the base one e.g https://github.com/openstack/cyborg/blob/233f2a5b7e396c24bf15ae22f8cf64a480666f5c/cyborg/accelerator/drivers/gpu/base.py#L30 | |
| 14:17:51 | sean-k-mooney | chandankumar: yes but that generic dirver was lated remvoed | |
| 14:18:46 | sean-k-mooney | jgilaber: ya that also a problem kind of | |
| 14:18:59 | sean-k-mooney | it will work but id dont want to chagne thie picemeal | |
| 14:19:00 | chandankumar | I will look into the history | |
| 14:19:32 | chandankumar | I can see two specs also https://github.com/openstack/cyborg-specs/blob/981405a8938ed786feaaf63ea1955aa7174a367b/specs/pike/implemented/cyborg-driver-proposal.rst#L8 and https://github.com/openstack/cyborg-specs/blob/981405a8938ed786feaaf63ea1955aa7174a367b/specs/train/implemented/cyborg-accelerator-driver.rst#L8 around generic driver | |
| 14:19:44 | sean-k-mooney | chandankumar: ok | |
| 14:20:03 | sean-k-mooney | so i have been hining at the fact that this might need a spec or mor thought then a quick bugfix | |
| 14:20:21 | sean-k-mooney | i.e. tha that you shoudl proably not focus on this right now | |
| 14:20:26 | sean-k-mooney | let me know what you find | |
| 14:20:27 | chandankumar | yes correct | |
| 14:20:43 | chandankumar | sure, I will dig into the history and open a bug around that | |
| 14:20:47 | chandankumar | we can take it from there | |
| 14:20:53 | sean-k-mooney | this was somithing i was plannign to look into closer to the end of the cycle ebfor ethe next ptg | |
| 14:21:12 | sean-k-mooney | i dont really expect use to make large changes to this ebfore then | |
| 14:22:26 | sean-k-mooney | basiclly i want to look at propsoign a new generic driver framework for next chcel | |
| 14:22:39 | sean-k-mooney | and prot the exitign fucntionaltiy as part fo that work | |
| 14:22:44 | chandankumar | yup, I will keep it to finding, you can take it from there | |
| 14:23:19 | sean-k-mooney | cool so i think we can move onto jgilaber spec | |
| 14:23:25 | chandankumar | yes | |
| 14:23:38 | chandankumar | #link https://review.opendev.org/c/openstack/cyborg-specs/+/982276: Add generic mdev driver spec for 2026.2 | |
| 14:23:56 | chandankumar | jgilaber: I have added your spec since you made changes to it | |
| 14:24:05 | chandankumar | Anything you want to highlight from last revision | |
| 14:24:18 | jgilaber | thanks for the reviews on the spec | |
| 14:24:35 | jgilaber | there is one thing worth pointing out I think | |
| 14:25:10 | jgilaber | while going through sean-k-mooney comments there is one that made me doubt, let me find the link | |
| 14:25:26 | jgilaber | #link https://review.opendev.org/c/openstack/cyborg-specs/+/982276/comment/b5f2b851_3cfdee76/ | |
| 14:25:31 | jgilaber | #undo | |
| 14:25:48 | jgilaber | sorry, I though I pasted the wrong link | |
| 14:26:02 | sean-k-mooney | the deplicate device coment | |
| 14:26:11 | jgilaber | yes | |
| 14:26:18 | sean-k-mooney | ya so in the medium term the conducrot shoud nto eb talkign to palcement at all | |
| 14:26:36 | sean-k-mooney | the cybrog compute agents should | |
| 14:26:39 | jgilaber | it looks to me that currently in cyborg we cannot currently have a deployable with more than one resource class | |
| 14:26:55 | jgilaber | which we might want to do to use multiple mdev classes from one device | |
| 14:27:16 | sean-k-mooney | well that proably ok | |
| 14:27:29 | sean-k-mooney | what we need to model is as follows | |
| 14:27:44 | sean-k-mooney | we need 1 placement resouce provider per PF | |
| 14:28:00 | sean-k-mooney | and we need an inventory of mdev benetat it | |
| 14:28:38 | sean-k-mooney | so each mdev will be a deployable right but there wotn be a 1:1 mapping between deployabels and resouce providers | |
| 14:29:13 | sean-k-mooney | i need to look into that in more deail myself to see how this is currently working | |
| 14:29:41 | sean-k-mooney | one thing that would help is some ascii diagrams in the spec | |
| 14:29:46 | sean-k-mooney | to visually show how this works | |
| 14:30:24 | sean-k-mooney | we have 2 related concepts in cybrogs api devices and deployables | |
| 14:30:32 | jgilaber | good point I'll add some diagrams, it was confusing for me too | |
| 14:30:54 | jgilaber | but if I understood correctly, in cyborg the resrouce provider is created from a DriverDeployable | |
| 14:31:10 | jgilaber | not from a DriverDevice | |
| 14:32:19 | sean-k-mooney | so the only way we woudl supprot more then one mdev type per phsyica device | |
| 14:32:27 | sean-k-mooney | is if we were addign the VFs | |
| 14:33:00 | sean-k-mooney | with that said we coudl aslo suffix the pci adress in the rp name | |
| 14:33:02 | sean-k-mooney | with the type | |
| 14:33:09 | sean-k-mooney | so i think we coudl workaroudn that limiation | |
| 14:33:20 | sean-k-mooney | but we may want to think about the modelign more | |
| 14:33:31 | sean-k-mooney | in nova we started with 1 type per device | |
| 14:33:38 | sean-k-mooney | so that woudl eb a reasonable limiation fro v1 | |
| 14:33:56 | sean-k-mooney | the nvidia driver didnt supprotr multipel asfar as im aware either | |
| 14:34:06 | jgilaber | that is my understanding yes, the simplest workaround would be one DriverDeployable for each mdev type | |
| 14:34:31 | sean-k-mooney | can you write up some of the options in the spec | |
| 14:34:39 | sean-k-mooney | and we can discuss the tradeoffs | |
| 14:34:43 | jgilaber | the nvidia driver just picks the first vgpu type it finds and uses that to create the rp | |
| 14:34:55 | sean-k-mooney | i see | |
| 14:34:57 | jgilaber | sure, I already have some of this but I'll expand with the diagrams | |