[openstack-dev] [nova][scheduler] Updating Our Concept of Resources
ndipanov at redhat.com
Wed Jun 3 14:02:19 UTC 2015
On 06/03/2015 02:13 PM, John Garbutt wrote:
> On 3 June 2015 at 13:53, Ed Leafe <ed at leafe.com> wrote:
>> On Jun 2, 2015, at 5:58 AM, Alexis Lee <alexisl at hp.com> wrote:
>>> If you allocate all the memory of a box to high-mem instances, you may
>>> not be billing for all the CPU and disk which are now unusable. That's
>>> why flavors were introduced, afaik, and it's still a valid need.
>> So we had a very good discussion at the weekly IRC meeting for the Scheduler, and we agreed to follow that up here on the ML. One thing that came up, noted in the quote above, is that I gave the impression in my first email that I thought flavors were useless. I think I did a better job in the original blog post of explaining that flavors are a great way to handle the sane division of a resource like a compute node. The issue I have with flavors is that we seem to be locked into the "everything that can be requested has to fit into the flavor", and that really doesn't make sense.
>> Another concern was from the cloud provider's POV, which makes a flavor a convenient way of packaging cloud resources for sale. The customer can simply say "give me one of these" to specify a complex combination of virtualized resources. That's great, but it means that there has to be a flavor for every possible permutation of resources. If you restricted flavors to only represent the sane ways of dividing up compute nodes, any other features could be add-ons to the request. Something like ordering a pizza: offer the customer a fixed choice of sizes, but then let them specify any toppings in whatever combination they want. That's certainly more sane than presenting them with a menu with hundreds of pizza "flavors", each representing a different size/topping combination.
> I feel there is a lot to be said for treating "consumable" resources
> very separately to "free" options.
> For example grouping the vCPUs into sockets can be "free" in terms of
> capacity planning, so is a valid optional add on (assuming you are not
> doing some level of pinning to match that).
> For things where you are trying to find a specific compute node, that
> kind of attribute has clear capacity planning concerns, and is likely
> to have a specific "cost" associated with it. So we need to make sure
> its clear how that cost concept can be layered on top of the Nova API.
> For example "os_type" often changes the cost, and is implemented on
> top of flavors using a combination of protected image properties on
> glance and the way snapshots inherit image properties.
>>> I totally agree the scheduler doesn't have to know anything about
>>> flavors though. We should push them out to request validation in the
>>> Nova API. This can be considered part of cleaning up the scheduler API.
>> This idea was also discussed and seemed to get a lot of support. Basically, it means that by the time the request hits the scheduler, there is no "flavor" anymore; instead, the scheduler gets a request for so much RAM, so much disk, etc., and these amounts have already been validated at the API layer. So a customer requests a flavor just like they do now, and the API has the responsibility to verify that the flavor is valid, but then "unpacks" the flavor into its components and passes that on to compute. The end result is the same, but there would be no more need to store "flavors" anywhere but the front end. This has the added benefit of eliminating the problem with new flavors being propagated down to cells, since they would no longer need to have to translate what "flavor X" means. Don Dugger volunteered to write up a spec for removing flavors from the scheduler.
> +1 for Nova translating the incoming request to a "resource request"
> the scheduler understands, given the resources it knows about.
> I would look at scoping that to "compute" resources, so its easier to
> add "volume" and "network" into that request at a later date.
I also agree with this pretty much completely. I feel that the single
thing that made some of the scheduler discussions drag on for months is
our lack of willingness to bite of the big chunk that is coming up with
a solid API to the scheduler.
Starting from nouns and verbs - it definitely seems like a good idea to
pass in the _requested_ resources to a scheduler that knows about
_avalible_ resources.  seems like an excellent start.
I seem to remember Jay discussing at one point that not all of the
things we want the scheduler to know about make sense to be modelled as
resources (running instances for example) and it made a lot of sense to
me, but it seems like it's the kind of thing that would be the easiest
to figure out once you see the code (I also don't see it mentioned in
 but I assume Jay dropped it to keep the scope of that BP manageable).
PS. I feel that exactly this type of work such as figuring out an API
for a component, would get done way quicker if we could use feature
branches to iterate quickly with invasive changes, without worrying too
much about rebasing on every single change in master, but still keep the
benefits of having a voting CI.
More information about the OpenStack-dev