| Posted | Nick | Remark | |
|---|---|---|---|
| #openstack-cyborg - 2019-08-29 | |||
| 03:15:26 | Coco_gao_ | agree, Sundar | |
| 03:15:49 | Sundar | As 2nd priority, I'd say devices -- that will give an inventory of accelerator devices in the cluster | |
| 03:16:57 | Sundar | IMHO, when devices are asked for, we can return the components like deployables and attributes, so the client gets a full picture | |
| 03:16:59 | shaohe_feng | yes, client if more friendly than curl | |
| 03:17:25 | s_shogo | As related the client, the deadline for openstackSDK's commit seems to be near, so would like to begin commit to that, prior to the merge of APIv2 patches. | |
| 03:17:34 | Sundar | Yes, makes sense | |
| 03:17:48 | Sundar | Thanks, s_shogo! | |
| 03:18:28 | Sundar | The main thing that is holding me back is that I am testing P5-P9 with the notification and Placement report patches. Plus, Nova code changes to create a merge conflict for me. | |
| 03:18:46 | Sundar | Once those are resolved, hope we can merge the P5-P9 patches | |
| 03:19:01 | Sundar | ANy other comments on the client, anybody? | |
| 03:19:13 | shaohe_feng | yes, async job depends on P5-P9 | |
| 03:19:25 | s_shogo | In my assumption,python-cyborg client and openstacksdk could to be completed before the Train release, | |
| 03:19:38 | shaohe_feng | great | |
| 03:19:38 | s_shogo | but I'm anxious of sufficiency in my test codes, thus please review that in following patches, and help that if necessary. | |
| 03:19:53 | Sundar | shaohe_feng: Agreed. I'll expedite as much as I can. | |
| 03:20:06 | Sundar | s_shogo: Agreed, we'll help for sure | |
| 03:20:23 | s_shogo | Thanks , Sundar | |
| 03:20:29 | shaohe_feng | maybe the test codes can be add later. | |
| 03:20:35 | Sundar | wangzhh: Thanks for proposing the RBAC patch. I had some concerns/questions in the patch. Please take a look. | |
| 03:20:50 | shaohe_feng | firstly let the client can work. | |
| 03:21:11 | Coco_gao_ | s_shogo, thank you . We will review the code. | |
| 03:21:28 | wangzhh | Yep. I have updated my code. May commit after meeting. | |
| 03:21:39 | Sundar | Thanks, wangzhh | |
| 03:22:03 | s_shogo | shaohe_feng : OK, I'll do that preferentially. | |
| 03:22:26 | Sundar | shaohe_feng: Part of the issue is that some Nova developers want to test Cyborg code with Nova code in theor env. Also, we need to show tempest working end-to-end. | |
| 03:23:41 | Sundar | Anybody else trying out the Placement report? With GPUs, AI chip, etc.? | |
| 03:23:42 | Coco_gao_ | What's the remaining work for tempest? | |
| 03:23:46 | shaohe_feng | yes, tempest can eliminate their concerns | |
| 03:24:07 | Sundar | Coco_gao_: It is mostly to get the patches to work together, I think | |
| 03:24:34 | Sundar | Xinran's patches look good IMO. Trying to make sure they work with P5-P9 | |
| 03:24:51 | Yumeng | I have tried the Placement report With GPUs | |
| 03:25:20 | Sundar | Yumeng: Good to know | |
| 03:25:42 | Sundar | #topic Nova functional tests | |
| 03:26:40 | Sundar | There was talk at the PTG that we should propose functional tests for Nova, which mock CYborg API in a test fixture, and use that to test Nova patches | |
| 03:27:08 | Sundar | They seem to cover a few more scenarios than unit tests and tempest | |
| 03:27:52 | Coco_gao_ | mock cyborg API's return? | |
| 03:27:53 | chenke | I agree we need to import functional test for nova. | |
| 03:28:09 | Sundar | Coco_gao_: Yes | |
| 03:28:44 | Sundar | We have an entry in the Storyboard too. I have not any comments of late, but there is concern that it may come up at the last moment | |
| 03:29:11 | Sundar | Since there is lots of stuff in Nova runway, it can be tough to get a 2nd look if this issue comes up | |
| 03:29:38 | Sundar | DO we have any volunteers for writing Nova functional tests? I'll help as much as I can | |
| 03:31:35 | Sundar | Please think it over and LMK if you can. | |
| 03:32:33 | Sundar | shaohe_feng: Do you want to bring up the discussion about ARQ states and transitions, as followup? Or is it settled? | |
| 03:33:12 | shaohe_feng | yes | |
| 03:33:45 | shaohe_feng | one things is that, who delete the ARQ | |
| 03:34:09 | shaohe_feng | when delete API tag the state as delete_pending? | |
| 03:34:22 | Sundar | There is Nova code to delete the ARQ in some error cases and when VM is terminated | |
| 03:35:03 | shaohe_feng | maybe it is still in bind process | |
| 03:35:47 | shaohe_feng | the bind process to delete it when it find the state is delete_pending? | |
| 03:36:06 | Sundar | Yes. In that case, IMHO, it is best to let the bind complete and the traits get updated in Placement, and then unbind/delete the ARQ | |
| 03:36:26 | Sundar | If we try to interrupt FPGA progamming, bad things can happen | |
| 03:36:46 | shaohe_feng | we will not add any rollback this release for bind. just go through the whole process even deleting. | |
| 03:36:54 | Sundar | Agreed | |
| 03:37:05 | shaohe_feng | OK. | |
| 03:37:09 | Coco_gao_ | OK | |
| 03:37:48 | shaohe_feng | any state transform should be transaction. | |
| 03:38:39 | Sundar | Yes, db transaction | |
| 03:39:21 | shaohe_feng | seems there is a state machine in oslo lib | |
| 03:39:26 | Sundar | Any other issue, shaohe_feng? | |
| 03:39:43 | shaohe_feng | we will not introduce it release | |
| 03:40:12 | Sundar | Ok by me. What are the benefits of using that? | |
| 03:40:13 | shaohe_feng | for I need time to read up it. | |
| 03:40:31 | shaohe_feng | do not look into it at present. | |
| 03:40:35 | Sundar | ok | |
| 03:40:55 | shaohe_feng | maybe after the whole flow code are finished | |
| 03:41:07 | shaohe_feng | we can have a look for cons and pros | |
| 03:41:32 | Sundar | Sure. We'll trust your judgement on this :) | |
| 03:41:48 | shaohe_feng | another things, should the async job timeout? | |
| 03:42:01 | Sundar | On a different note, I am seeing this issue for allocating attach handles: https://opendev.org/openstack/cyborg/src/branch/master/cyborg/db/sqlalchemy/api.py#L269 The in_use field does not get written to db | |
| 03:42:37 | shaohe_feng | but there's still a problem. | |
| 03:42:40 | Sundar | The timeout should correspond to default Nova timeout | |
| 03:43:10 | shaohe_feng | maybe it is in programming or other critical job | |
| 03:43:34 | Sundar | The programming typically takes a few seconds, so default of 300 seconds (I think) is good enough | |
| 03:43:40 | shaohe_feng | timeout can be disaster | |
| 03:44:21 | shaohe_feng | another things | |
| 03:44:43 | shaohe_feng | currently the bind process is specify for FPGA | |
| 03:45:40 | Sundar | Umm, bind if for all accelerators. Only programming is for FPGA. the bind means the ARQ is associated with a host and deployable in Cyborg's db, and the device is ready to use | |
| 03:45:44 | Sundar | *is for | |
| 03:45:55 | shaohe_feng | there should be good extension for other kinds | |
| 03:46:05 | shaohe_feng | I means: | |
| 03:46:22 | shaohe_feng | 1. get the resource type. | |
| 03:46:50 | shaohe_feng | every resource type should has its own extend bind action | |
| 03:46:56 | shaohe_feng | for FPGA it is program. | |
| 03:47:08 | shaohe_feng | other's maybe evn setup, not sure. | |
| 03:47:43 | shaohe_feng | 2. every resource should be has its own placement report. | |
| 03:48:20 | shaohe_feng | the report info maybe different | |
| 03:48:35 | shaohe_feng | so the code should be: | |
| 03:49:09 | shaohe_feng | type, num = arq.group_get_resource() | |
| 03:49:17 | shaohe_feng | for n in num: | |
| 03:50:05 | shaohe_feng | action = get_accelerator_action(type) # fpga is program | |
| 03:50:08 | shaohe_feng | action() | |
| 03:50:15 | shaohe_feng | somethings like this | |
| 03:50:38 | shaohe_feng | and these code should be split from the arq object file | |
| 03:51:36 | Sundar | In general, the process should be generic for all accelerators. The current code looks at the device profile request group to see if it has function_id or bitstream_id entries, which are specific to FPGA, to decide if programming is needed | |
| 03:52:32 | shaohe_feng | we maybe add other spec in | |
| 03:52:36 | Sundar | AFAIK, for non-FPGA devices in this release, there is nothing required to prepare the device, right? | |
| 03:53:10 | shaohe_feng | devices profile for different acclerations | |
| 03:53:14 | shaohe_feng | such as HDDL | |
| 03:53:19 | shaohe_feng | we can add | |
| 03:54:18 | shaohe_feng | "accel:affinity": true | |