CNCF: Multi-tenant GPU Kubernetes on simulated hardware with Kamaji and HAMi
The author built a single-machine lab where Kamaji runs two tenant Kubernetes control planes inside one management cluster and HAMi shares simulated GPUs between pods using a mock device plugin. The post reports that three fractional-GPU pods were scheduled onto one simulated A100 card, that HAMi refused pods exceeding the remaining cores or memory, and that eight tests passed. The source text provided is cut off before the tenant isolation test results are shown.
What changes for platform teams
Platform teams evaluating per-team Kubernetes isolation with GPU sharing can see a reproducible pattern for putting control planes in a shared cluster while scheduling fractional GPUs through HAMi. Results come from simulated hardware, so they do not show real GPU performance or driver behavior.
Read the original at CNCF Blog ↗
written by claude-haiku-5-5, prompt 7cbe154, checked by publish rules