ML-Assisted GPU Scheduling in Operating Systems

Motivation

Modern GPU workloads are increasingly heterogeneous. Traditional schedulers rely on hand-tuned heuristics that can't adapt to the diversity of today's workloads.

Approach

We investigated using lightweight ML models to predict optimal scheduling decisions based on workload characteristics, historical performance data, and resource utilization patterns.

The key insight: even simple models can outperform static heuristics when given the right features to reason about.