DRA
Dynamic Resource Allocation (DRA) is a Kubernetes mechanism for requesting and allocating devices through resource claim objects. It is not a GPU-only capability.
In ACP, DRA is used through a product driver. A DRA driver is an implementation of the DRA mechanism, not the mechanism itself.
How it affects workloads
DRA is relevant when:
- a workload uses
ResourceClaimorResourceClaimTemplateinstead of only extended resource names; - the accelerator product documents mention
DeviceClass,ResourceSlice, or claim-based allocation; - the cluster has both traditional Device Plugin based products and DRA based products, and you need to identify which allocation path a workload uses.
Product documentation
Use this page to understand what DRA means for workload resource requests and how to distinguish DRA from traditional Device Plugin based allocation. Use the product documentation for DRA driver installation, Kubernetes version requirements, feature gates, node labels, claim examples, validation, and troubleshooting.
Related products
Only document a DRA driver in a product site after the corresponding Alauda product or package explicitly supports it.
- A vendor DRA driver belongs to the corresponding vendor integration.
- A HAMi DRA driver belongs to the HAMi sharing and scheduling path because HAMi owns the sharing, virtualization, scheduling, and resource semantics.
- Future GPU, NPU, or other accelerator products may also use DRA.