摘要
为了优化GPU资源利用率,文章提出了一种新型的GPU资源调度方案。该方案结合了最小剩余时间算法与SR-IOV(SingleRootI/OVirtualization)技术,可优化多用户、多任务环境下的GPU资源利用率和系统性能。传统的GPU调度方法往往面临资源利用不足、任务等待时间长和系统吞吐量受限等问题。为了应对这些挑战,该方案通过动态分析任务的剩余执行时间,利用SR-IOV技术实现了GPU资源的细粒度隔离与共享,可为更高效的资源分配和任务调度提供支持。实验结果表明,相较于传统的无调度、容器调度和常见机器学习调度方案,该方案在均值准确率、GPU利用率、系统吞吐量和任务执行时间等方面均具有一定的优势,可为多用户多任务场景下的GPU资源管理提供有益的参考。
In order to optimize the utilization of GPU resources,the article proposes a new type of GPU resource scheduling scheme.This scheme combines the minimum remaining time algorithm with SR-IOV(Single Root I/O Virtualization)technology to optimize GPU resource utilization and system performance in multi-user and multitasking environments.Traditional GPU scheduling methods often face problems such as insufficient resource utilization,long task waiting times,and limited system throughput.To address these challenges,this solution dynamically analyzes the remaining execution time of tasks and utilizes SR-IOV technology to achieve fine-grained isolation and sharing of GPU resources,providing support for more efficient resource allocation and task scheduling.The experimental results show that compared to traditional unscheduled,container scheduling,and common machine learning scheduling schemes,this scheme has certain advantages in mean accuracy,GPU utilization,system throughput,and task execution time.It can provide useful reference for GPU resource management in multi-user and multitasking scenarios.
作者
梁桂才
何现海
马梓钧
陆富业
LIANG Guicai;HE Xianhai;MA Zijun;LU Fuye(Information Management Center,Guangxi Technological College of Machinery and Electricity,Nanning 530007,China;Guangxi Changji Electronic Technology Co.,Ltd.,Nanning 530022,China;Information Center,Guangxi Mechanical and Electrical Engineering School,Nanning 530007,China)
出处
《计算机应用文摘》
2024年第9期140-145,共6页
Chinese Journal of Computer Application
基金
2023年广西统筹支持工业振兴资金(工业互联网发展)项目:基于GPU的AI算力一体化资源池的构建(2209-450000-07-04-451566)。
