Kubernetes GPU Sharing Advertises Four Pods, Not Guarantees Memory

The author tested three GPU sharing methods in Kubernetes—time-slicing, MPS, and MIG—using a T4 and an RTX PRO 6000, finding that a node GPU count of four is merely a scheduling count rather than a resource guarantee. The tests revealed significant variations in throughput, memory isolation, fault tolerance, and power limits across the different modes. Additionally, repartitioning MIG caused cluster disruption due to cgroup driver mismatches between the node configuration and the GPU Operator. Open article →