Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Wu, Xin, Teng, Fei, Li, Xingwang, Zheng, Bin, Duan, Qiang
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2603.25378
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866911546169360384
author Wu, Xin
Teng, Fei
Li, Xingwang
Zheng, Bin
Duan, Qiang
author_facet Wu, Xin
Teng, Fei
Li, Xingwang
Zheng, Bin
Duan, Qiang
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2603_25378
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle PRISM: Dynamic Primitive-Based Forecasting for Large-Scale GPU Cluster Workloads
Wu, Xin
Teng, Fei
Li, Xingwang
Zheng, Bin
Duan, Qiang
Distributed, Parallel, and Cluster Computing
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.
title PRISM: Dynamic Primitive-Based Forecasting for Large-Scale GPU Cluster Workloads
topic Distributed, Parallel, and Cluster Computing
url https://arxiv.org/abs/2603.25378