Workshops Archive

Data-Driven Static MIG Planning for an H200 Research Cluster


Workshop: Workshop on Operational Data Analytics in HPC (HPC-ODA)

Authors: Arsalan Akhter (Northeastern University)

Abstract: Research computing clusters support GPU workloads ranging from small interactive jobs to large full-device training runs. Exposing each GPU as a monolithic resource wastes capacity because many jobs require only a fraction of a GPU. Modern GPUs, such as the NVIDIA H200, support Multi-Instance GPU (MIG) partitioning, but real-time reconfiguration is impractical for heterogeneous, continuously arriving HPC workloads.

We frame MIG planning as an operational data analytics problem, using long-term production telemetry to make a one-time provisioning decision. We construct a slot-level demand matrix and formulate layout selection as an exact mixed-integer linear program (MILP), evaluating planners by weighted fulfillment. Our optimal layout improves in-sample fulfillment from 66.8% to 98.1%, reaches 97.6% on held-out data, and 99.7% on raw demand. A simple greedy planner achieves the same optimum due to a near-zero LP integrality gap (<0.001%). The layout is robust to demand inflation, and deploys directly with Slurm GRES.


Back to Workshop on Operational Data Analytics in HPC (HPC-ODA) Archive Listing

Back to Full Workshop Archive Listing