Skip to content

GPU Sharing#

conceptstub

Time-slicing, MPS, MIG, HAMi, and DRA: different ways to share a GPU, with different isolation guarantees.

  • Concepts

    How it works: the mechanism, the moving parts, and where it sits in the stack.

  • Design

    Trade-offs, decision records, and when to choose which option.

  • Implementation

    Manifests, Helm values, and scripts. Every script states whether it is read-only or changes the cluster.

  • Testing

    How to prove it works: test procedures, expected output, and performance baselines.

  • Troubleshooting

    Failure modes, symptoms, and the diagnostic path from symptom to root cause.

Planned content

  • Time-slicing: oversubscription, no memory isolation
  • MPS: concurrent kernels, limited fault isolation
  • MIG: hardware partitions with dedicated SMs, L2, and memory
  • HAMi: memory caps and soft compute limits via CUDA API interception
  • DRA: structured device claims and sharing