-
Feature
-
Resolution: Unresolved
-
Undefined
-
None
-
None
-
Product / Portfolio Work
-
-
False
-
-
False
-
None
-
None
-
None
-
None
-
-
None
-
None
-
None
-
None
Feature Overview (aka. Goal Summary)
Background:
In OpenShift, Dynamic Resource Allocation (DRA) is used to manage the allocation of specialized hardware resources (e.g., GPUs, NICs) beyond CPU and memory. However, traditional DRA supports only whole-device allocation, which can lead to underutilization of powerful hardware.
Enhancement Summary:
Kubernetes Enhancement Proposal #4815 introduces support for partitionable devices within the DRA framework. This enhancement enables OpenShift administrators to allocate logical slices of a single device to multiple pods—such as GPU slices (e.g., NVIDIA MIG), virtual functions of SR-IOV NICs, or other virtualizable hardware.
Use Cases in OpenShift
| Use Case | Description |
|---|---|
| GPU Sharing for AI/ML | Allow multiple AI workloads to share slices of the same GPU, maximizing resource usage. |
| High-density Scheduling | Improve bin-packing efficiency for workloads requiring only partial device capacity. |
| Virtual NICs | Enable SR-IOV virtual NICs to be exposed to pods through partitionable allocation. |
| Cost Optimization | Reduce hardware costs by minimizing the number of physical devices needed per node. |