Integration with Kueue

[kannon92]

We have discussed some comparisons of other schedulers (#29).

I think it would be worth describing how a kueue integration would work.

KAI could support Jobs/Jobset/Pytorch jobs without much effort for Kueue.

For KAI support of services I think #63 is needed.

To expand on batch jobs, I think one needs to investigate if it is possible to use Kueue's ClusterQueues/LocalQueues in place of KAI Queues. To put it simple, Kueue integration (sans Topology Aware Scheduling) could be that Kueue handles queueing and resuming workloads once their is capacity in the cluster (queueing) and KAI can handle scheduling.

For KAI maintainers, the main request would be to figure out what would be lost if KAI's queueing logic was folded into Kueue. Is there anything missing in Kueue that would not allow KAI to utilize Kueue for queueing while leaving scheduling for KAI?

[enoodle]

Hi @kannon92

On the features side I think that Kueue is missing the concept of over quota weight or the differentiation between deserved quota and quota limits and instead it has other ways to describe quota
https://kueue.sigs.k8s.io/docs/concepts/cluster_queue/#flavors-and-resources
https://github.com/NVIDIA/KAI-Scheduler/blob/main/docs/queues/README.md#queue-resources

Another feature in KAI is that jobs can be defined as non preemptible, and we will want to refine the way it is used in the future.

KAI is pretty modular so you could run it with many actions/plugins turned off or configured differently.
For example if you want to integrate them today some of the actions (reclaim, preempt) and plugins (proportion) can be turned off and you could duplicate all the Kueue queues into KAI queues with infinite quota which I think will get Kueue to control the pods creation and KAI to try and schedule everything that is created.

[kannon92]

Another feature in KAI is that jobs can be defined as non preemptible, and we will want to refine the way it is used in the future.

https://github.com/kubernetes-sigs/kueue/pull/6141/files

Looks like someone is interesting in adding this feature to Kueue.

Maybe this will be interesting to you.