SMART-MIG framework boosts GPU scheduling efficiency 18% with AI
New RL-based system partitions GPUs dynamically, slashing energy use while meeting deadlines.
SMART-MIG addresses the dual challenge of deciding GPU partitions and scheduling jobs onto heterogeneous slices in modern data centers. The framework leverages Multi-Instance GPU (MIG) technology to split a single GPU into smaller, isolated instances, allowing multiple smaller ML models to run concurrently instead of occupying an entire GPU. To avoid the exponential complexity of joint optimization, SMART-MIG employs Mean-Field Multi-Agent Reinforcement Learning (MF-MARL) for the repartitioning decision — treating each GPU as an agent that learns to adjust its partition configuration over time. The job scheduling component uses tailored heuristics to place incoming tasks onto the available MIG slices, balancing energy consumption and job tardiness (delay beyond deadline).
The team provides theoretical lower bounds on energy and tardiness to benchmark performance, and extensive experiments show SMART-MIG outperforms static-partition baselines by 18% in energy-tardiness efficiency. Perhaps most importantly, the MF-MARL approach scales gracefully: the complexity of the repartitioning decision remains constant as the number of jobs and GPUs grows, making SMART-MIG viable for large-scale deployments. Accepted at IEEE IPDPS 2026, the paper (arXiv:2606.29775) marks a practical step toward sustainable, high-performance GPU scheduling without sacrificing job completion times.
- SMART-MIG uses Mean-Field Multi-Agent RL (MF-MARL) to dynamically repartition GPUs, achieving constant-time complexity regardless of cluster size.
- Achieves 18% better energy-tardiness efficiency compared to static GPU partitioning, and only 27% above the theoretical lower bound on energy consumption.
- Framework combines MF-MARL for repartitioning with heuristic job scheduling, designed for large-scale data centers running multiple small ML models on MIG slices.
Why It Matters
Makes GPU data centers more energy-efficient and scalable, directly reducing operational costs and carbon footprint without sacrificing job performance.