PRISM: Dynamic Primitive-Based Forecasting for Large-Scale GPU Cluster Workloads

A research team led by Xin Wu has introduced PRISM, a novel forecasting framework designed to tackle the notoriously volatile and heterogeneous workloads running on large-scale GPU clusters powering modern AI platforms. The system addresses a critical bottleneck in AI infrastructure: accurately predicting resource demands for jobs that exhibit multiple periodicities and sudden bursts, which traditional predictors struggle with. PRISM's core innovation is a dual-representation approach that combines dictionary-driven temporal decomposition with adaptive spectral refinement. This method breaks down complex workload patterns into stable, interpretable "primitives" or signatures, creating a compositional model that can adapt to diverse GPU job types.

Evaluated against large-scale production traces from real AI training clusters, PRISM demonstrated state-of-the-art forecasting performance. Its key achievement is a significant reduction in errors during burst phases—the moments of peak, unpredictable demand that most strain scheduling systems and cause inefficiencies. By providing more accurate and architecture-aware predictions, PRISM establishes a robust foundation for downstream optimization tasks. This enables data center operators and cloud providers to implement dynamic resource management, including efficient job scheduling, proactive power capping, and optimized GPU allocation, ultimately reducing operational costs and improving cluster utilization for AI workloads.