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

2026-03-26

AI Summary

GPU workloads exhibit high volatility, multi-scale periodicity, and heterogeneity, posing significant challenges for traditional prediction methods and thereby limiting the scheduling efficiency of AI infrastructure. To address this, this work proposes PRISM, a novel framework that uniquely integrates primitive composition modeling with adaptive spectral refinement. By leveraging dictionary-driven time-series decomposition, PRISM extracts stable and interpretable multi-scale features, effectively capturing both periodic patterns and bursty behaviors inherent in complex GPU workloads. The framework establishes an architecture-aware dynamic forecasting system that substantially improves prediction accuracy on large-scale production traces, particularly reducing errors during burst phases. This advancement provides robust support for resource scheduling in GPU clusters, enhancing overall system responsiveness and efficiency.

Technology Category

• Machine Learning: Hardware-aware ML
• Planning, Routing, and Scheduling: Model-Based Reasoning
• Humans and AI: Human-Aware Planning and Behavior Prediction

Application Category

• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Accurately forecasting GPU workloads is essential for AI infrastructure, enabling efficient scheduling, resource allocation, and power management. Modern workloads are highly volatile, multiple periodicity, and heterogeneous, making them challenging for traditional predictors. We propose PRISM, a primitive-based compositional forecasting framework combining dictionary-driven temporal decomposition with adaptive spectral refinement. This dual representation extracts stable, interpretable workload signatures across diverse GPU jobs. Evaluated on large-scale production traces, PRISM achieves state-of-the-art results. It significantly reduces burst-phase errors, providing a robust, architecture-aware foundation for dynamic resource management in GPU-powered AI platforms.

Problem

• GPU workload forecasting
• large-scale AI infrastructure
• workload volatility
• heterogeneous workloads
• multiple periodicity

Innovation

• primitive-based forecasting
• temporal decomposition
• spectral refinement
• GPU workload prediction
• dynamic resource management