Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2019-11-20
14:59:51 gibi efried: I'm sort on context. Is it about provider config?
15:00:05 efried if we can detect it at startup, we can halt the service and force you to fix it. But we don't want to kill the service if it creeps in after we've already started running.
15:00:07 efried gibi: yes
15:00:16 dansmith efried: if it's hard/impossible to do at start, is clearly wrong config, and we can log an error later I think that's reasonable
15:00:29 dansmith efried: no, don't halt.. log an error periodically
15:00:36 efried dansmith: ack
15:00:41 gibi efried: so in case of a compute RP the name is known before the compute was ever started but the uuid only known after the compute creates the RP
15:00:44 dansmith we should never kill nova-compute after it's hit steady-state
15:00:48 efried right
15:00:57 dansmith gibi: we have $COMPUTE_NODE for that
15:01:05 efried gibi: Right, in that case, we can use the special ``$COMPUTE_NODE`` ... yeah ^
15:01:07 gibi dansmith: I see.
15:01:09 dansmith they don't need to know either the name or the uuid for the compute node
15:01:33 gibi so instead of allowing RP identification by name in general, we add a specific case for compute node
15:01:50 dansmith we have that already
15:02:03 dansmith uuid can be a uuid or the special string "$COMPUTE_NODE"
15:02:40 gibi so far I'm OK with this. I guess if other RPs needs to be identified by name then we will add other symboles for that or we re-think the identify-by-name feature
15:02:41 dansmith efried: note that I'm okay with either removing name, or just detecting and logging this case later
15:02:50 bauzas gibi: dansmith: I haven't paid attention yet to this convo as I'm dragged into some internal all-hands but when I was looking on how to use the new COMPUTE_NODE trait, I think that we miss the world of nested RPs
15:03:11 bauzas unless we have some way to have some traits to be cascaded into child RPs
15:03:24 bauzas (and I really want to avoid saying 'inheritance')
15:03:59 bauzas but for example, if you lookup some RP, you need to call the RP and check its trait to know whether the original RP is still related to a compute node
15:04:02 efried bauzas: The only rp we're supporting this "wildcard" for right now is the root. If we find we need a way to generically identify its children later, we'll have more designing to do.
15:04:23 bauzas you need to call the *root* RP
15:04:27 bauzas (sorry for confusing)
15:04:40 dansmith is bauzas talking about this same thing?
15:04:46 efried almost :P
15:04:53 gibi hm, the neutron device RPs (provising bandwidth) also has a stable name generated from the device name. But as I don't have a use case for provider config related to the bandwidth feature I rest my case.
15:05:01 bauzas efried: sure, I'm just saying this is only convenient when you're in a flat world
15:05:26 efried I agree. Predictably-named nested RPs make a case for identifying by name.
15:06:08 efried though I think we've talked about the brittleness of relying on names before.
15:06:32 gibi efried: we have predicatable named nested RPs | 1110cf59-cabf-526c-bacc-08baabbac692 | aio:Open vSwitch agent:br-test
15:06:55 mnaser sean-k-mooney: sorry, was in a meeting too then, but no that wasnt enabled, but they are much older and also running ddr3 memory for comparision
15:07:02 mnaser much lower latency but lower bandwidth
15:07:12 gibi efried: but as I said I have no direct use case so I don't want to push for the naming support in provider config
15:07:14 sean-k-mooney so that was cross socket latency
15:07:16 bauzas efried: FWIW, we stecked a few things at the PTG between gibi, stephenfin and me that were tied to naming conventions :)
15:07:23 bauzas we sketched*
15:07:24 efried I know, what I'm saying is that we shouldn't have architected anything to rely on naming conventions, and we shouldn't perpetuate such things.
15:07:35 efried but that ship has probably sailed.
15:07:37 mnaser sean-k-mooney: yep!
15:07:37 sean-k-mooney mnaser: i was trying to figure out if the intel number were between socket or two numa nodes in the same socket
15:07:43 bauzas efried: indeed
15:07:49 sean-k-mooney ya those latency numbers for amd are bad
15:07:54 mnaser nope, between two sockets, which is interesting in just how much faster cross socket latency is low
15:08:21 efried In any case, if gibi doesn't have any real, current bw use cases, and bauzas doesn't have any real, current vgpu use cases, I'm fine making the decision to remove identify-by-name to limit scope/complexity.
15:08:23 gibi efried: there are things already architectured that way. like the hostname of the compute node is an id neutron use. and nova user the name of the device RP from neutron
15:08:24 mnaser sean-k-mooney: im looking at the bios for the speed between the two sockets, i learned something about the amd architecture that doesnt seem ideal
15:08:41 efried we've removed a lot of bells and whistles for that reason, just to make this go.
15:08:41 gibi efried: agree with the limited scope
15:08:52 gibi efried: go for it
15:08:56 bauzas efried: it's not really a vgpu thingy
15:08:56 sean-k-mooney so that lend more crednece to enableing numa per ccx and usign muiti numa guests
15:08:59 gibi efried: we can always add that later
15:09:05 bauzas efried: it's more what gibi said
15:09:08 efried otoh, where I started this morning actually handles things including name pretty well IMO
15:09:19 bauzas efried: some other projets create their own RPs and we need to know them
15:09:19 sean-k-mooney mnaser: the fact the infity fabric speed used to be tied to the meory speed
15:09:20 efried and dustinc has mostly already coded it up
15:09:26 efried so it's really just a question of test surface
15:09:34 mnaser sean-k-mooney: "By default, the BIOS for EPYC 7002 Series processors will run at the maximum allowable clock frequency by the platform. This configuration results in the maximum memory bandwidth for the processor, but in some cases, it may not be the lowest latency. The Infinity Fabric will have a maximum speed of 1467 MHz (lower in some platforms), resulting in a single clock penalty to transfer data from the
15:09:34 mnaser memory channels onto the Infinity Fabric to progress through the SoC. To achieve the lowest latency, you can set the memory frequency to be equal to the Infinity Fabric speed. Lowering the memory clock speed also results in power savings in the memory controller, thus allowing the rest of the SoC to consume more power potentially resulting in a performance boost elsewhere, depending on the workload."
15:09:50 bauzas efried: for the moment, we infer their existences based on naming convention but I'm all up for other ideas :-)
15:10:26 gibi bauzas: a counter example is shared disk. there we said use the uuid of the placement aggregate instead of the name of the sharing RP in the nova conf.
15:10:34 sean-k-mooney mnaser: ya so in gen 1 they were tied of the same clock
15:10:48 gibi bauzas: so we are getting better at this
15:10:49 bauzas gibi: aaaaaand we discussed 30 mins of any potential issues with it :)
15:10:56 dansmith efried: reminder: I'm fine with logging the conflict late when the admin configured it wrong
15:11:02 sean-k-mooney so your memroy bandwith was limited by the speed at which the infinity frabric could run
15:11:04 bauzas the chicken-and-egg problem
15:11:10 dansmith maybe nobody else is :)
15:11:13 gibi bauzas: that is what PTG for. isn't it? ;)
15:11:21 sean-k-mooney mnaser: they split them to allwo you to tune them seperatly
15:11:34 bauzas gibi: that's surely some justification for my expense report, for sure :p
15:11:38 sean-k-mooney mnaser: so you can optimise for latency or througput
15:12:05 efried dansmith: That wfm too. The only question is, in addition to logging the error, we have to *pick one* to use.
15:12:10 bauzas "look ! this isn't a very expensive dinner, this was just a brainstorming session in a crowdy area !"
15:12:13 efried or, I suppose, refuse to use either.
15:12:19 dansmith efried: refuse to use either
15:12:33 bauzas "and we ordered snacks meanwhile"
15:12:39 dansmith IMHO
15:12:46 efried dansmith: ack. I agree. We shall make it so.
15:12:51 efried dansmith: Since you're being so nice already, and I've obv got you swapped out of whatever else you would be doing, can we talk about shelve-offloading an instance with a vTPM?
15:13:11 dansmith efried: yeah
15:13:37 efried Do you have a better idea than "create a glance (or maybe swift?) object for the vdev file"?
15:13:42 gibi bauzas: there wasn't too much snack provided during the PTG so _we had to_ order some :)
15:14:11 efried dansmith: seems like this is a natural for swift, but that adds yet another required component.
15:14:27 bauzas "good luck, expenses team, for understanding my bills !"
15:14:41 dansmith efried: that would be one place I guess, but yeah, can't really depend on swift being there
15:15:04 mnaser sean-k-mooney: interestingly enough looking at the bios the 4-link xGMI max speed is set to 10.667Gbps
15:15:28 dansmith efried: I dunno how safe that would really be, not that it's less secure than the alternatives, but..
15:15:50 mnaser "Setting this to a lower speed can save uncore power that can be used to increase core frequency or reduce overall power. It will also decrease cross socket bandwidth and increase cross socket latency."
15:16:00 kashyap coreycb: One more on that thing you asked earlier: QEMU upstream has this notion of "versioned CPU models", and we'll be getting new named models with the affected flags turned off.
15:16:08 mnaser i wonder why SMC would ship a box with with it set to the minimum speed for a server where i dont really care about cpu performance, ill bump that to 18Gbps..
15:16:25 efried dansmith: Well, we're relying on the encrypted-ness of the vdev file already. A main use case is "someone walks away with the disk and we're still okay".
15:16:28 kashyap coreycb: I'll update the Ubuntu bug you pointed.
15:16:43 coreycb kashyap: that sounds nice. great, thanks
15:17:06 efried dansmith: though since you bring it up, I wonder if it would be possible to (ab)use barbican...
15:17:36 dansmith efried: sure, but access to that enables an offline attack, and not everyone may have the same security policy for their swift cluster as they do for their hypervisor nodes
15:17:59 dansmith efried: switft may be a best effort uuber cheap storage bucket for them, whereas hypervisor and glance storage is high security at a premium

Earlier   Later