Setting up the scheduling service on your cluster
The scheduling service is a cluster-wide pod scheduling service that you can install on your IBM® Cloud Pak for Data cluster.
Important: If you plan to install the Watson™ Machine Learning
Accelerator service on your cluster, you must install the
scheduling service.
Overview of the scheduling service
The scheduling service offers enhancements
over the default Kubernetes scheduler, including:
- Quota enforcement
- Refresh 2 or
later This feature enables you
to programmatically enforce the quotas that you set for Cloud Pak for Data or for various Cloud Pak for Data services.
For details on quota enforcement, see Managing the platform.
- Co-scheduling of pods
- Refresh 2 or
later This feature is provided
for the Watson Machine Learning
Accelerator service.
Parallel and AI workloads can co-schedule pods to
- Guarantee that all pods can start
- Remove resource deadlock
- Enable workloads to grow and shrink
- Support reclaiming pods in the event of resource contention
- GPU sharing
- Refresh 2 or
later This feature is provided
for the Watson Machine Learning
Accelerator service.
The scheduling service allows competing groups to share GPUs, which improves GPU utilization. Sharing policies govern how to resolve resource contention.