On 10/14/2015 06:38 AM, Dmitry Tantsur wrote: > Hi OoO'ers :) > > It's going to be a long letter, fasten your seat-belts (and excuse my > bad, as usual, English)! > > In RDO Manager we used to have a feature called advanced profiles > matching. It's still there in the documentation at > http://docs.openstack.org/developer/tripleo-docs/advanced_deployment/profile_matching.html > > but the related code needed reworking and didn't quite make it upstream > yet. This mail is an attempt to restart the discussion on this topic. > > Short explanation for those unaware of this feature: we used detailed > data from introspection (acquired using hardware-detect utility [1]) to > provide scheduling hints, which we called profiles. A profile is > essentially a flavor, but calculated using much more data. E.g. you > could sat that a profile "foo" will be assigned to nodes with 1024 <= > RAM <= 4096 and with GPU devices present (an artificial example). > Profile was put on an Ironic as a capability as a result of > introspection. Please read the documentation linked above for more details. > > This feature had a bunch of problems with it, to name a few: > 1. It didn't have an API > 2. It required a user to modify files by hand to use it > 3. It was tied to a pretty specific syntax of the hardware [1] library > > So we decided to split this thing into 2 parts, which are of value one > their own: > > 1. Pluggable introspection ramdisk - so that we don't force dependency > on hardware-detect on everyone. > 2. User-defined introspection rules - some DSL that will allow a user to > define something like a specs file (see link above) via an API. The > outcome would be something, probably capabilit(y|ies) set on a node. > 3. Scheduler helper - an utility that will take capabilities set by the > previous step, and turn them into exactly one profile to use. > > Long story short, we got 1 and 2 implemented in appropriate projects > (ironic-python-agent and ironic-inspector) during the Liberty time > frame. Now it's time to figure out what we do in TripleO about this, namely: > > 1. Do we need some standard way to define introspection rules for > TripleO? E.g. a JSON file like we have for ironic nodes?
Yes, please. > > 2. Do we need a scheduler helper at all? We could use only capabilities > for scheduling, but then we can end up with the following situation: > node1 has capabilities C1 and C2, node2 has capability C1. First we > deploy a flavor with capability C1, it goes to node1. Then we deploy a > flavor with capability C2 and it fails, despite us having 2 correct > nodes initially. This is what state files were solving in [1] (again, > please refer to the documentation). It sounds like the answer is yes. If the existing scheduler can't handle a valid use case then we need some sort of solution. > > 3. If we need, where does it go? tripleo-common? Do we need an HTTP API > for it, or do we just do it in place where we need it? After all, it's a > pretty trivial manipulation with ironic nodes.. I think that would depend on what the helper ends up being. I can't see it needing a REST API, but presumably it will have to plug into Nova somehow. If it's something that would be generally useful (which it sounds like it might be - Ironic capabilities aren't a TripleO-specific thing), then it belongs in Nova itself IMHO. > > 4. Finally, we need an option to tell introspection to use > python-hardware. I don't think it should be on by default, but it will > require rebuilding of IPA (due to a new dependency). Can we not just build it in always, but only use it when desired? Is the one extra dependency that much of a burden? > > Looking forward to your opinions. > Dmitry. > > [1] https://github.com/redhat-cip/hardware > > __________________________________________________________________________ > OpenStack Development Mailing List (not for usage questions) > Unsubscribe: [email protected]?subject:unsubscribe > http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev > __________________________________________________________________________ OpenStack Development Mailing List (not for usage questions) Unsubscribe: [email protected]?subject:unsubscribe http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
