| Posted | Nick | Remark | |
|---|---|---|---|
| #openstack-cyborg - 2019-04-24 | |||
| 03:01:40 | Sundar | Good to have many of us right at the beginning | |
| 03:02:23 | Sundar | Agenda: PTG session time allocation, Summit project update slides | |
| 03:02:27 | Sundar | Anything else? | |
| 03:03:08 | Li_Liu | Have we decided the team dinner date yet? | |
| 03:03:39 | Sundar | It is Thu evening. But we haven't decided a place yet. I'll leave that to all of you | |
| 03:04:00 | Li_Liu | ok | |
| 03:04:08 | yikun | 2th May, right? | |
| 03:04:21 | Sundar | yikun: Yes | |
| 03:04:26 | yikun | Sundar: ok | |
| 03:04:30 | Coco_gao | OK, maybe we can keep the old place. | |
| 03:04:33 | Sundar | #topic PTG session planning | |
| 03:05:20 | Sundar | We have 3.5 hours on Thu afternoon (1:30-5) and 2.5 hours on Fri afternoon (1:30-4) because I need to leave by 4 to catch my flight | |
| 03:05:33 | Sundar | Not much time! We need to plan the sessions tightly | |
| 03:05:48 | Sundar | You can all continue on Fri eve and on Sat too, when I am gone | |
| 03:06:48 | Sundar | BTW, Fri morn are the cross-projects with Nova and ironic | |
| 03:06:52 | wangzhh | #info wangzhh | |
| 03:07:25 | Sundar | Shall we start allocating times to the topics in Looking at: https://etherpad.openstack.org/p/cyborg-ptg-train ? | |
| 03:07:51 | Sundar | s/Looking at// | |
| 03:08:01 | xinranwang | Hi all | |
| 03:08:07 | yikun | xinranwang: hi | |
| 03:08:16 | Sundar | Hi xinranwang, wangzhh and Yumeng | |
| 03:08:28 | xinranwang | #info xinranwang | |
| 03:08:42 | xinranwang | hi Sundar yikun | |
| 03:08:45 | Yumeng | Hi Sundar, and all | |
| 03:08:50 | Yumeng | # info Yumeng | |
| 03:09:00 | wangzhh | Hi Sundar, hi all. | |
| 03:09:04 | Sundar | Retrospective: keep it to 10 min? | |
| 03:09:51 | Sundar | Too short? | |
| 03:10:16 | yikun | Sundar: I think we should classify these topic in etherpad, such as, driver, api, common/design...and then do the allocation tiems | |
| 03:10:27 | yikun | * times | |
| 03:11:11 | Coco_gao | 10min if no other people join, because we all familar with the project status. | |
| 03:11:26 | Coco_gao | Sundar, agree with yikun | |
| 03:11:38 | Sundar | yikun: ok. The main topics in my mind are: Train Goals, Networked FPGAs, mapping names to UUIDs, new drivers/devices (incl. Ascend) | |
| 03:11:39 | Coco_gao | Better to classify | |
| 03:12:15 | wangzhh | Yikun, do you need more time for topic #line 58? | |
| 03:12:16 | Sundar | Some of them are cross-cutting across these categories, esp. first 2 | |
| 03:12:52 | yikun | wangzhh: checking... | |
| 03:13:56 | yikun | wangzhh: bitstream management? So what's the "more time" mean? you mean the more time to discuss in PTG? | |
| 03:14:42 | wangzhh | Yep, I saw you leave comment 'expand the bitstream to more generic level'. So... | |
| 03:14:50 | Sundar | yikun, all: In my mind, bitstream mgmt means Cyborg provides APIs to parse metadata for bitstreams, validate and upload in Glance. What else do you have in mind? | |
| 03:15:37 | Sundar | I think firmware update etc. are usually done as part of host config, with pupper/ansible/... | |
| 03:15:54 | Sundar | *puppet | |
| 03:16:12 | yikun | wangzhh: ok, make sense, it is just some idea for "Do not provide specific API for specific type driver." | |
| 03:16:49 | wangzhh | Got it. | |
| 03:16:51 | ikuo_o | Sundar: I'm thinking bitstream md5 checksum to check integrity. | |
| 03:16:55 | Coco_gao | I thought there were a generic Driver in code, but we don't use that or inherit from that. | |
| 03:17:51 | Sundar | ikuo_o: That is up to the vendor stack, because they may use different checksums, authentication/decryption schemes, or even do it in hardware in some cases | |
| 03:18:13 | Coco_gao | About the bug, do we need specific time for that? | |
| 03:18:14 | wangzhh | Coco, Yep, but, actually, it is out of date. Should be improved as our new design. | |
| 03:18:22 | yikun | Coco_gao: yes, https://github.com/openstack/cyborg/blob/master/cyborg/accelerator/drivers/generic_driver.py | |
| 03:18:49 | Coco_gao | Thanks, yikun. The generic driver is too old. | |
| 03:18:55 | Sundar | Have you all looked at the driver API in Line 396 https://review.opendev.org/#/c/608624/4/specs/stein/approved/cyborg-api-wflows-for-instance-ops.rst ? The can_handle API can be dropped but can you review the rest? | |
| 03:20:07 | yikun | So these APIs is belong to the generic_driver.py? | |
| 03:20:55 | Sundar | Anyways, I'll take a first stab like this: 30 min each for Train Goals, Networked FPGAs, mapping names to UUIDs, new drivers/devices, NTT's discussion. That is 2.5 hours. Now we can add more time to some of these topics and fit more on Friday | |
| 03:20:56 | Coco_gao | Sundar, OK | |
| 03:21:32 | yikun | Sundar: sounds good | |
| 03:21:33 | Sundar | May be new devices needs more time | |
| 03:22:13 | ikuo_o | Thanks, Sundar! | |
| 03:22:53 | Sundar | ikuo_o: Welcome :) | |
| 03:23:03 | Coco_gao | I need some time in Friday to dicuss the dirty data caused by conductor diff. | |
| 03:23:29 | Sundar | Coco_gao: Good. Conductor diff: 15-30 min? | |
| 03:23:37 | Coco_gao | That's good | |
| 03:23:51 | Sundar | PCI slot change on reboot -- why is it any different than Nova's PCI subsystem? | |
| 03:25:34 | Coco_gao | I will make sure with shaohe, and see if we need to discuss that. | |
| 03:25:51 | Sundar | wangzhh, Coco_gao: any performance countres planned for GPUs? I am thinking just report driver-reported perf counters from agent to a collectd plugin. From there, it can be consumed by telemetry/ceilometer, prometheus, .... | |
| 03:25:53 | Coco_gao | Right now , we can skip that | |
| 03:26:04 | Sundar | Coco_gao: ok | |
| 03:27:45 | yikun | Who is the owner of L42 idea? I plus some info on it | |
| 03:28:24 | Sundar | yikun: do you mean SmartNIC? | |
| 03:28:33 | yikun | "Performance/health monitoring for GPUs or any other devices?" | |
| 03:30:01 | wangzhh | Whuuu, difficult to monitor it now, because when we passthrough the GPU in the vm. Host will lost control of this device. I'll consider about that. | |
| 03:30:36 | Sundar | yikun: I responded too. I would like to find an owner. | |
| 03:31:00 | ikuo_o | I'm not the owner but the function seems useful. | |
| 03:31:08 | wangzhh | Does FPGA or other devices have the same scenario? | |
| 03:31:10 | Sundar | wangzhh: yea, same problem for FPGAs, but we should have the PF in the host from where we can get shell counters. | |
| 03:31:45 | Sundar | Besides, we also have dependence on the vendor stack | |
| 03:32:04 | Sundar | yikun: Does Ascend support SR-IOV? | |
| 03:32:27 | wangzhh | Sundar, Yep. Depend on vendor stack. | |
| 03:33:58 | Sundar | If Ascend supports SR-IOV, and we can get performance counters from its PF, we have a potential use case. Otherwise, we'll focus on other practical things and then come to this. | |
| 03:35:14 | yikun | Sundar: ascend 310 don't support sr-iov yet | |
| 03:35:25 | Sundar | Then shall we deprioritize this topic for PTG? | |
| 03:36:00 | xinranwang | ascend support entire pci passthrough right | |
| 03:36:05 | yikun | yes | |
| 03:36:26 | yikun | passthrough is supported. | |
| 03:36:44 | wangzhh | Sundar, Depend on the priorities and our time. | |
| 03:36:46 | yikun | Sundar: I think we still can add this topic in PTG | |
| 03:37:16 | Sundar | OK, I only said de-prioritized. So, if there's time after topics, sure | |
| 03:37:46 | Sundar | Driver support for Movidius - who owns this? | |
| 03:38:05 | yikun | Shaohe? | |
| 03:38:06 | xinranwang | Sundar: i think is Shaohe | |
| 03:38:08 | Coco_gao | Coco | |
| 03:38:37 | Sundar | RISC-V? | |
| 03:39:06 | Coco_gao | We use Movidius in the Edge. | |
| 03:39:32 | yikun | Movidius is a USB device, right? | |
| 03:39:33 | Sundar | Coco_gao: Cool. Maybe both you and Shaohe can drive it? | |
| 03:39:42 | Coco_gao | Yes | |
| 03:40:38 | Coco_gao | It's USB right now, in order to use it, we need to install openvino | |
| 03:40:44 | Sundar | I should a topic for post-PTG plans. That's important. Merge pilot with master, add FPGA CI, etc. are important for Nova intgration | |
| 03:41:03 | Coco_gao | Sundar, agree | |
| 03:41:17 | Coco_gao | the most important thing is nova integration | |