Building a Driver for Dynamic Resource Allocation in Kubernetes - Device Plugins 2.0

Explore the cutting-edge Kubernetes feature, Dynamic Resource Allocation (DRA), in this informative conference talk. Learn how DRA revolutionizes resource scheduling by empowering third-party developers and offering a more flexible API similar to persistent volumes. Discover the myriad benefits of DRA for GPU management, including controlled sharing, support for multiple models per node, arbitrary constraint specification, dynamic MIG device allocation, and GPU repurposing. Gain practical insights into building your own resource driver for DRA, utilizing Kubernetes's in-tree helper libraries and following best practices for driver architecture. Follow along as the speakers use NVIDIA and Intel GPU drivers as examples, culminating in a live demonstration of these drivers in action.

Device Plugins 2.0: How to Build a Driver for Dynamic Resource Allocation - K Klues & Alexey Fomenko