| Posted | Nick | Remark | |
|---|---|---|---|
| #openstack-cyborg - 2018-08-30 | |||
| 19:50:10 | Sundar | efried: Please ping me when you can | |
| 19:50:19 | efried | Sundar: Hi! | |
| 19:51:26 | Sundar | Hi, I got tons of feedback on the os-acc spec ;). Can we discuss them now? | |
| 19:51:41 | efried | okay, sure | |
| 19:51:47 | efried | Though I think you got most of it from sean | |
| 19:52:48 | Sundar | One feedback was that the attaching to the VM should be left to Nova virt driver, and os-acc should not do it. That is hypervisor-specific and that's what virt drivers are for. | |
| 19:52:55 | Sundar | Do you agree with that? | |
| 19:53:30 | efried | I... think so, yes. I think that's how neutron plugins operate. But maybe it's not how the os-vif model is set up. | |
| 19:54:20 | efried | anyway, it's certainly hypervisor-specific. No question there. | |
| 19:54:35 | efried | So when you say "os-acc" you really mean "the plugin at the behest of os-acc". | |
| 19:55:30 | Sundar | One main difference between os-vif and os-acc that I am struggling to communicate is this: with os-vif, a VIF gets allocated, port binding happens and then the plug() operation happens. | |
| 19:57:08 | Sundar | With os-acc, the equivalent of binding and plug cannot be separated cleanly. Example: A GPU may need to be reconfigured to create vGPUs of a certain type. Until that point, the accelerators of the right type don't even exist. So, the device may need to be configured to create the right accelerators, or change the inventory of accelerators, before we get the attach handle | |
| 19:58:40 | efried | okay. That stuff is *also* hypervisor-specific. | |
| 19:59:09 | Sundar | Sure, it is also device-specific. The device driver has to do this in a hypervisor-specific way | |
| 19:59:18 | efried | May be the purview of the "driver" rather than the "plugin" but still. | |
| 19:59:44 | Sundar | Yes, I think we already agree that the driver would be hypervisor-specific | |
| 19:59:51 | efried | it may in fact be device *type* specific, not necessarily device specific. | |
| 20:00:02 | efried | Anyway, I think we're in violent agreement on the high points here. | |
| 20:00:36 | Sundar | Well, the actual act of writing to the device and manipulating it to create a new vGPU type or program a bitstream would in fcat be specific to device models or vendors. | |
| 20:02:19 | Sundar | So, the Cyborg and its driver(s) would have to handle the equivalent of port binding (device specific) and plug. That gives us a VAN with the device end configured, ready to be attached to an instance | |
| 20:02:46 | Sundar | After that, Nova virt can take that VAN and do the needful for that hypervisor, in a device-independen way | |
| 20:02:48 | efried | and does that happen in a separate step prior to spawn? | |
| 20:03:18 | Sundar | Yes. I shared a flow diagram with you this morning. | |
| 20:03:31 | Sundar | Before we get to that :), | |
| 20:04:11 | Sundar | the role of os-acc plugin is highly diminished with these aspects considered. The driver does the device end, Nova virt does the instance end | |
| 20:04:48 | efried | "nova virt" by invoking the plugin through os-acc? | |
| 20:04:55 | Sundar | The possible role for os-acc extensions is to handle device-compute interactions, such as NUMA affinity for interrupt vectors. | |
| 20:06:41 | Sundar | Nova virt calls os-acc, and that in turn calls Cyborg in some way. When that call returns, Nova virt has to persist the instance - VAN association in Nova's db. neither os-acc nor Cyborg can do that, so it has to go to Nova virt. The subsequent step of attaching to the VM is already in Nova virt. | |
| 20:06:48 | Sundar | So, what does the plugin do for attach? | |
| 20:07:07 | efried | no idea | |
| 20:07:19 | efried | if it's nothing, it's nothing. | |
| 20:08:03 | Sundar | I understand your need for keeping things hypervisor-specific, but Nova virt is already hypervisor-specific, right? | |
| 20:08:14 | efried | yup | |
| 20:08:39 | efried | I don't really have a stake in getting lots of code into these plugins. | |
| 20:08:48 | efried | I just want to make sure we don't end up with linux-isms in common code paths. | |
| 20:08:54 | efried | linux/libvirt | |
| 20:09:16 | Sundar | OK. Could you look at the flow diagram I shared this morning? | |
| 20:09:26 | efried | If I didn't have to write a separate driver/plugin at all, I would be pretty happy. But I'm not sure that is going to happen. | |
| 20:10:08 | Sundar | Not to spoil your afternoon, but I think you would need PowerVM drivers for GPUs, FPGAs, or whatever you support :) | |
| 20:11:23 | Sundar | Are you looking at Power+KVM hypervisor too? | |
| 20:12:10 | Sundar | brb in 5 min | |
| 20:12:19 | efried | I don't know anything about pkvm | |
| 20:14:00 | Sundar | back. ok | |
| 20:15:43 | Sundar | Back to os-acc plugins. The possible role for os-acc plugins is to handle device-compute interactions, such as NUMA affinity for interrupt vectors. But it may be too much to bring them in this spec. Shall we put os-acc plugins as a placeholder until we get to that? | |
| 20:18:26 | efried | "NUMA affinity for interrupt vectors" <== Greek to me. | |
| 20:18:51 | Sundar | OK, NUMA affinity for devices in general? | |
| 20:19:11 | efried | If you don't have an actual use case for os-acc plugins, then... | |
| 20:19:35 | efried | NUMA affinity is something we want to be able to handle via resource provider structure. | |
| 20:19:48 | efried | we can't yet | |
| 20:19:56 | efried | but it's something we should bring up (again) in Denver. | |
| 20:20:49 | Sundar | Sure. Shall we have a f2f chat at Denver? It'll be good to sync up and drive the specs to a close shortly thereafter | |
| 20:21:14 | efried | definitely | |
| 20:21:30 | efried | If I were you, I would ask for a nova/cyborg cross-project session. | |
| 20:21:52 | efried | Based on what I know of people's schedules, I would say Tuesday would be a good day for it. | |
| 20:22:03 | Sundar | I have entered a slot in Nova etherpad. May be I should ping melwitt? | |
| 20:23:02 | efried | It might be better to schedule something during cyborg's time (M/T) where you can get the "placement cores" to visit the cyborg room and spend a couple of hours hashing stuff out. | |
| 20:23:33 | efried | And keep it on the nova schedule in case we a) don't get everything sorted; and/or b) need a wider audience. | |
| 20:23:48 | Sundar | OK | |
| 20:24:04 | efried | Send something to the dev ML, like what Blazar did, to organize the Tuesday thing. | |
| 20:24:05 | efried | oh | |
| 20:24:14 | efried | Sorry, I forgot, Blazar sent that directly to people. | |
| 20:24:18 | efried | not to the ML. | |
| 20:24:34 | Sundar | Meanwhile, please LMK if you are ok with the flow diagram. Did you get it? | |
| 20:26:34 | efried | I got it. | |
| 20:27:47 | Sundar | Great. I'll grab lunch and get back in 15 min. We can shake it out after that. | |
| 20:29:57 | efried | give me 20 | |
| 21:07:14 | Sundar | efried: Ready any time | |
| 21:08:00 | efried | Sundar: Nova meeting in progress. Not sure how long it will last, but not beyond top of the hour. Will you be around? | |
| 21:09:31 | Sundar | Yes | |
| 21:21:08 | efried | Sundar: ō/ | |
| 21:23:57 | Sundar | That was quick. | |
| 21:24:32 | efried | now where did I put that diagram... | |
| 21:24:56 | efried | got it | |
| 21:25:14 | Sundar | The main thing to note in the flow is that there isn't much that os-acc is doing. | |
| 21:25:59 | Sundar | Most of it is Nova virt or Cyborg. I am still pondering what os-acc can do that is not device-specific (and so in Cyborg) or hypervisor-specific (and so in Nova virt) | |
| 21:26:59 | efried | Sundar: as a touchpoint/router and future extension point for those pieces of the requests, it may be useful. | |
| 21:27:05 | efried | but I'm really not sure. | |
| 21:27:30 | Sundar | Yes, I agree. That's what I was thinking | |
| 21:28:21 | efried | not sure it makes much sense for the calls from os-acc to cyborg API/conductor to be asynchronous if it's just going to poll for them to complete right away... | |
| 21:29:37 | Sundar | The calls may take milliseconds to possibly seconds, depending on whether Glance bitstreams need to fetched, one or more FPGAs need to be programmed, etc | |
| 21:30:20 | efried | so? | |
| 21:30:27 | Sundar | That is why it is async | |
| 21:30:31 | efried | The caller is blocking on their completion anyway | |
| 21:30:39 | efried | So why does it matter if they take "a long time"? | |
| 21:30:48 | Sundar | Ah, the n-cpu could do other things while this is blocked, right? | |
| 21:30:53 | efried | could it? | |
| 21:31:06 | efried | I guess. | |
| 21:31:18 | Sundar | The allocation of other resources -- networking, storage -- could go in parallel | |
| 21:31:21 | efried | Just sounds like it would make things pretty complicated. | |
| 21:31:36 | efried | If n-cpu wanted to parallelize, it could send that request off in its own thread. | |
| 21:31:46 | efried | But having it be async *forces* n-cpu to deal with that async-ness. | |
| 21:31:49 | efried | Anyway, this is really a nit. | |
| 21:31:59 | efried | not a substantive thing. | |
| 21:32:03 | Sundar | The n-cpu to os-acc could be async too. | |
| 21:32:15 | Sundar | So, os-acc deals with the async weirdness. | |
| 21:32:59 | Sundar | Hmm, well, os-acc to Cyborg can be sync if os-acc itself is called in an async way | |
| 21:33:13 | Sundar | Just wondering if a REST API call cna really block for seconds | |
| 21:33:48 | efried | absolutely. | |
| 21:34:10 | efried | I'm sure there's a connection timeout at the HTTP level. But seconds shouldn't be a problem. | |
| 21:34:16 | Sundar | Are there precedents in OpenStack where a REST API blocks for seconds? | |