Scheduler Evolution

Evolving the scheduler has been a priority item over several releases:

The scheduler has become tightly coupled with the rest of nova, limiting its capabilities, accuracy, flexibility and maintainability. The goal of scheduler evolution is to bring about a better separation of concerns between scheduling functionality and the rest of nova.

Once this effort has completed, its conceivable that the nova-scheduler could become a separate git repo, outside of nova but within the compute project. This is not the current focus.

Problem Use Cases

Many users are wanting to do more advanced things with the scheduler, but the current architecture is not ready to support those use cases in a maintainable way. A few examples will help to illustrate where the scheduler falls short:

Cross Project Affinity

It can be desirable, when booting from a volume, to use a compute node that is close to the shared storage where that volume is. Similarly, for the sake of performance, it can be desirable to use a compute node that is in a particular location in relation to a pre-created port.

Accessing Aggregates in Filters and Weights

Any DB access in a filter or weight slows down the scheduler. Until the end of kilo, there was no way to deal with the scheduler accessing information about aggregates without querying the DB in every call to host_passes() in a filter.

Filter Scheduler Alternatives

For certain use cases, radically different schedulers may perform much better than the filter scheduler. We should not block this innovation. It is unreasonable to assume a single scheduler will work for all use cases.

However, to enable this kind of innovation in a maintainable way, a single strong scheduler interface is required.

Project Scale issues

There are many interesting ideas for new schedulers, like the solver scheduler, and frequent requests to add new filters and weights to the scheduling system. The current nova team does not have the bandwidth to deal with all these requests. A dedicated scheduler team could work on these items independently of the rest of nova.

The tight coupling that currently exists makes it impossible to work on the scheduler in isolation. A stable interface is required before the code can be split out.

Key areas we are evolving

Here we discuss, at a high level, areas that are being addressed as part of the scheduler evolution work.

Fixing the Scheduler DB model

We need the nova and scheduler data models to be independent of each other.

The first step is breaking the link between the ComputeNode and Service DB tables. In theory where the Service information is stored should be pluggable through the service group API, and should be independent of the scheduler service. For example, it could be managed via zookeeper rather than polling the nova DB.

There are also places where filters and weights call into the nova DB to find out information about aggregates. This needs to be sent to the scheduler, rather than reading directly from the nova database.

Versioning Scheduler Placement Interfaces

At the start of kilo, the scheduler is passed a set of dictionaries across a versioned RPC interface. The dictionaries can create problems with the backwards compatibility needed for live-upgrades.

Luckily we already have the oslo.versionedobjects infrastructure we can use to model this data in a way that can be versioned across releases.

This effort is mostly focusing around the request_spec. See, for example, this spec.

Sending host and node stats to the scheduler

Periodically nova-compute updates the scheduler state stored in the database.

We need a good way to model the data that is being sent from the compute nodes into the scheduler, so over time, the scheduler can move to having its own database.

This is linked to the work on the resource tracker.

Updating the Scheduler about other data

For things like host aggregates, we need the scheduler to cache information about those, and know when there are changes so it can update its cache.

Over time, its possible that we need to send cinder and neutron data, so the scheduler can use that data to help pick a nova-compute host.

Resource Tracker

The recent work to add support for NUMA and PCI pass through have shown we have no good pattern to extend the resource tracker. Ideally we want to keep the innovation inside the nova tree, but we also need it to be easier.

This is very related to the effort to re-think how we model resources, as covered by discussion about resource providers.

Parallelism and Concurrency

The current design of the nova-scheduler is very racy, and can lead to excessive numbers of build retries before the correct host is found. The recent NUMA features are particularly impacted by how the scheduler works. All this has lead to many people running only a single nova-scheduler process configured to use a very small greenthread pool.

The work on cells v2 will mean that we soon need the scheduler to scale for much larger problems. The current scheduler works best with less than 1k nodes but we will need the scheduler to work with at least 10k nodes.

Various ideas have been discussed to reduce races when running multiple nova-scheduler processes. One idea is to use two-phase commit “style” resource tracker claims. Another idea involves using incremental updates so it is more efficient to keep the scheduler’s state up to date, potentially using Kafka.

For more details, see the backlog spec that describes more of the details around this problem.