GPU任务调度研究综述  

Survey of GPU Task Scheduling

作　　者：李来文 胡韬 邓庆绪[1] LI Laiwen;HU Tao;DENG Qingxu(School of Computer Science and Engineering,Northeastern University,Shenyang 110004,China)

机构地区：[1]东北大学计算机科学与工程学院,沈阳110004

出　　处：《小型微型计算机系统》2024年第11期2800-2807,共8页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(62072085)资助.

摘　　要：本文针对运行在GPU上的任务的管理和调度研究进行了综述,并且把讨论重点放在针对单GPU上的相关研究工作.随着人工智能技术的发展以及相关应用的普及,使得GPU成为加速计算的关键工具.本文首先介绍了GPU的架构和编程模型,然后按照调度粒度,从stream级到warp级介绍了多种调度方法的相关研究工作.每个级别的调度方法都旨在提高GPU的性能、资源利用率、可靠性或降低能耗.此外,本文还指出了GPU任务调度面临的挑战以及未来的研究方向,如保障GPU执行时间确定性的软硬件机制研究、结合机器学习的GPU任务调度研究、GPU新架构探索研究以及追求GPU性能和能耗平衡的调度技术研究.本文旨在为研究者们提供一个全面的视角,帮助他们了解GPU任务调度的研究动态和未来的发展方向.This paper provides a survey of research on the management and scheduling of tasks running on Graphics Processor Unit(GPU)and focuses the discussion on related research work targeting a single GPU.The development of artificial intelligence technology and the popularity of related applications have made GPUs a key tool for accelerating computation.In this paper,we first introduce the architecture and programming model of GPUs,and then present related research work on multiple scheduling methods according to scheduling granularity,from stream level to warp level.The scheduling methods at each level aim to improve GPU performance,resource utilization,reliability,or reduce energy consumption.In addition,this paper points out the challenges of GPU task scheduling as well as the future research directions,such as the study of hardware and software mechanisms to guarantee the determinism of GPU execution time,the study of GPU task scheduling combined with machine learning,the study of exploring new GPU architectures,and the study of scheduling techniques pursuing the balance between GPU performance and energy consumption.The purpose of this paper is to provide researchers with a comprehensive perspective to help them understand the research dynamics and future directions of GPU task scheduling.

关 键 词：图形处理单元 CUDA GPU多任务 GPU调度 

分 类 号：TP316[自动化与计算机技术—计算机软件与理论]