Support Kubernetes Dynamic Resource Allocation API #320
Description
Summary
Research and implement PoC for DRA, allocate Device ResourceSlice in none-discrete way and allow users write complex requirement expression.
Motivation
There is a limitation of Kubernetes CDI or Device Plugin, it can not retrieve Pod info on secondary device allocation after primary centralized kube-scheduler allocation, TensorFusion has to implement GPU allocation logic in very hack way, and so as for other GPU schedulers.
Reference
- https://github.com/kubernetes/enhancements/blob/master/keps%2Fsig-node%2F4381-dra-structured-parameters%2FREADME.md
- Kubelet device plugin to pass additional information about Pods to Allocate call kubernetes/kubernetes#59109
- Karpenter - Dynamic Resource Allocation (DRA) kubernetes-sigs/karpenter#1231
Another dependency is Karpenter, its provisioning and disruption modules don't support DRA yet, it may cause Karpenter integration part broken.