When HPC Scheduling Meets Active Learning: Maximizing The Performance with Minimal Data

Abstract

High-performance computing (HPC) systems face complex resource management challenges as they accommodate diverse workloads, from traditional scientific simulations to emerging AI workflows. Accurate workload characterization is essential for optimal resource allocation, yet existing approaches often require extensive labeled data or rely on rigid heuristics that struggle to adapt to new workload patterns. This paper presents RF-AUS (Random Forest Adaptive Uncertainty Sampling), a novel active learning framework designed for HPC workload classification. Our proposed RF-AUS leverages ensemble-based uncertainty measurement, out-of-bag estimates, and feature importance weighting to achieve high classification performance with minimal labeled data. In addition, it employs a diversity-aware sampling strategy to balance exploration and exploitation, ensuring comprehensive coverage of the feature space while prioritizing the most informative instances. Using real-world workload data from NURION, a leading HPC system in South Korea, we demonstrate that RF-AUS achieves a comparable classification accuracy with 47.53% fewer labeled samples than conventional methods. RF-AUS shows robust performance across diverse applications, making it particularly well suited for production HPC environments where efficient resource utilization is crucial.