Topology-Aware Workload Scheduling

Topology-Aware Scheduling (TAS) is a feature of the Workload API that optimizes the placement of pods within the cluster.

TAS ensures that all pods within a PodGroup are co-located into a specific topology domain, such as a single server rack or zone. This minimizes inter-pod communication latency and prevents workload fragmentation across the cluster infrastructure.

Topology-aware scheduling with gang scheduling policy

When applied to PodGroups with gang scheduling policy, TAS simulates the potential assignment (placement) of the full group of pods at once. It guarantees that at least the specified minCount pods can fit together into the same topology domain before committing resources. If no feasible placement is found, the entire PodGroup becomes unschedulable.

This is the recommended approach for workloads like distributed AI and ML training that strictly require proximity to minimize inter-pod communication latency.

If new pods are added to the PodGroup where some pods are already scheduled (for example, if pods are recreated), the scheduler will force all new incoming pods to land on the exact same topology domain where the existing pods currently reside. If that specific domain lacks sufficient capacity for the new pods, the pods will remain pending - even if it means that less than minCount pods are scheduled at this point.

Topology-aware scheduling with basic scheduling policy

Using TAS with basic scheduling policy may exhibit inconsistent behavior. The scheduler may only observe a subset of pods when entering the PodGroup scheduling cycle - therefore placement feasibility is only evaluated for the observed pods, rather than the entire PodGroup. To partially mitigate this limitation, you can use scheduling gates to hold off PodGroup scheduling until all pods within the PodGroup are in the scheduling queue.

If no feasible placement is found for the entire PodGroup, only a subset of pods may be scheduled, and they are guaranteed to meet the scheduling constraints.

If new pods are added to the PodGroup where some pods are already scheduled, the scheduler will act the same as in case of gang policy - forcing the new pods into the same domain, unless there is insufficient capacity (in which case the new pods will remain pending).

API configuration: scheduling constraints

Every PodGroup (or PodGroupTemplate) may optionally declare the schedulingConstraints field, which is interpreted by the placement-based PodGroup scheduling algorithm. If constraints are defined in PodGroupTemplate, they will be copied to referencing PodGroups.

As of Kubernetes v1.36, the API supports topology constraints.

Topology constraint

To define a topology constraint for a PodGroup you need to set a key, which corresponds to a Kubernetes node label, representing the target topology domain (for example, a rack or a zone). The scheduler strictly enforces that all pods within the PodGroup are placed onto nodes that share the exact same value for this specified label.

Here is an example of a PodGroup configured with a topology constraint:

apiVersion: scheduling.k8s.io/v1alpha2
kind: PodGroup
metadata:
  name: example-podgroup
spec:
  schedulingPolicy:
    gang:
      minCount: 4
  schedulingConstraints:
    topology:
      - key: topology.example.com/rack

What's next

• Learn about pod group policies.
• Learn about plugins related Topology-aware Scheduling
• Read about gang scheduling algorithm.