Process Description
The scheduling process of a scheduler includes job verification, node pre-selection, node scoring, Ascend AI processor selection, and allocation result submission. For details about how to implement the affinity scheduling code of Volcano, see ascend-for-volcano. The code demonstrates how to incorporate the affinity scheduling policy into the scheduler. The following uses the Ascend AI processor of Atlas training product as an example to describe the Volcano scheduling process.
Figure 1 Volcano scheduling workflow
Process Description
- Job verification: Verify the number of Ascend AI processors requested by single-node or distributed jobs.
- Node preselection: Check whether the number of NPUs on a node meets the job requirements.
- Node scoring: Score the preselected nodes based on the affinity policy.
- Ascend AI processor selection: Select Ascend AI processors on the optimal node.
- Allocation result submission: Volcano submits the allocation result to Kubernetes.
Parent topic: Scheduling Algorithm of Ascend AI Processors