Fix resource tracking for migration operations (live, cold, rebuild)¶
Resource tracking for operations that move instances between compute hosts is broken in Nova. The fix requires some refactoring and tweaking of the data model so it’s discussed in a spec. It’s really about fixing of several long standing bugs.
Resource tracking for operations that move instances between compute hosts is broken. Those operations are:
There are 2 problems
In order for resources to be tracked properly in a Nova cloud, whenever a request to build an instance gets to a compute host, a resource claim needs to be done holding a global process-wide lock. Failure to do this can result in wrong resource allocation that will not follow the policy cloud administrator wants, or in some cases, failure to launch an instance. Live-migrate and rebuild code paths currently do not use claims at all.
Some resources like NUMA topology and PCI devices cannot be simply calculated from the flavor in case of a move operations, and must be persisted after a successful claim, as they don’t simply refer to a count of a single uniform resource (like a vCPU) but refer to an actual unique device/resource. 1
Live/cold migrate and rebuild, but it’s essentially a bug fix.
As this relates to resource tracking - this is deemed as a part of the scheduler priority effort.
The changes aim at solving the two issues described above.
First step is to make evacuate and live-migrate code paths do claims before starting the work (and also abort claims on failure). This is really bug-fixing work. It will also move creating migration entries proposed by 2 to happen as part of the claim.
Second, we will add additional data to track resources after being claimed for a migration, we will add a migration_context column to the instance_extra table and store the claimed resources there to be used for tracking migration resources. Part of this work will mean also changing the resource tracker to consider these when doing resource tracking. Initially it will only contain the ‘new’ NUMA topology.
Data model impact¶
Add a single column to the instance_extra table called migration_context. This column is NULL by default, or contains a serialized MigrationContext object, which we will also add as part of this work.
We’ll be accessing this data on the Instance object, in the same fashion other data stored in the instance_extra is accessed.
REST API impact¶
Other end user impact¶
Migrate and evacuate will need to acquire global locks to update tracked resources. It is likely that performance impact of this will be negligible.
Other deployer impact¶
- Primary assignee:
Add a new column to the instance_extra column and related objects code.
Change live_migration to claim resources using a call to ResourceTracker.live_migrate_claim() (which we will add) likely as part of the check_can_live_migrate_destination compute service method. Move the creation of the migration object to this method or additionally flip a flag that lets the resource tracker know to consider it for resource calculations, in case we deem necessary to create migration records elsewhere.
Do the similar as above for rebuild, claim happening in the rebuild_instance compute manager method.
Make sure that the newly added claim methods, persist the newly calculated data (NUMA topology initially).
We rely on 2 to introduce creation of migration objects for live-migrate and rebuild operations.
The scope of this work will focus on solid unit testing of the functionality added/changed. This area is a good target for functional testing, however as with all similar pieces of functionality that need different execution threads to hit interesting edge cases, it is difficult to come up with repeatable automated tests.
Discission on the following patch <https://review.openstack.org/#/c/163440/> offers some more details on this.
Optional section for liberty intended to be used each time the spec is updated to describe new design, API or any database schema updated. Useful to let reader understand what’s happened along the time.