数据密集型计算作业在GPU集群上的共享调度算法
作者:
作者单位:

作者简介:

通讯作者:

汤小春,E-mail:tangxc@nwpu.edu.cn

基金项目:

国家重点研发计划(2018YFB1003400)


Algorithm of Share Scheduling for Data-intensive Computing Operations onto GPU Cluster
Author:
Affiliation:

Fund Project:

National Key Research & Development Program of China(2018YFB1003400)

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
    摘要:

    数据密集型作业包含大量的任务,使用GPU设备来提高任务的性能是目前的主要手段,但是,在解决数据密集型作业之间的GPU资源公平共享以及降低任务所需数据在网络间的传输代价方面,现有的研究方法没有综合考虑资源公平与数据传输代价的矛盾.论文分析了GPU集群资源调度的特点,提出了一种基于最小代价最大任务数的GPU集群资源调度算法,解决了GPU资源的公平分配与数据传输代价较高的矛盾,将调度过程分为两个阶段,第一阶段为各个作业按照数据传输代价给出自己的最优方案,第二阶段为资源分配器合并各个作业的方案,按照公平性给出全局的最优方案.首先,论文给出了GPU集群资源调度框架的总体结构,各个作业给出自己的最优方案,资源分配进行全局优化;第二,给出了网络带宽估计策略以及计算任务的数据传输代价的方法;第三,给出了基于GPU数量的资源公平分配的基本算法;第四,提出了最小代价最大任务数的资源调度算法,描述了资源非抢夺,抢夺以及不考虑资源公平策略的实现策略;最后,设计了六种数据密集型计算作业,对论文中提出的算法进行了实验,通过实验验证,最小代价最大任务数的资源调度算法,对于资源公平性能够达到90%左右,同时亦能保证作业并行运行时间最小.

    Abstract:

    Data-intensive tasks include a large number of tasks. Using GPU devices to improve the performance of tasks is the main method currently. However, in the case of solving the fair sharing of GPU resources between data-intensive tasks and reducing the cost of data network transmission, the existing research methods do not comprehensively consider the contradiction between resource fairness and data transmission costs. The paper analyzes the characteristics of GPU cluster resource scheduling, and proposes an algorithm based on the minimum cost and the maximum number of tasks in GPU cluster resource scheduling. The method can solve the contradiction between the fair allocation of GPU resources and the high cost of data transmission. The scheduling process is divided into two stages. In the first stage, each job gives its own optimal plan according to the data transmission costs, and in the second stage, the resource allocator merges the plan of each job. Firstly, the paper gives the overall structure of the framework, and the source allocator works globally after each job giving its own optimal plan. Secondly, the network bandwidth estimation strategy and the method of computing the data transmission cost of the task are given. Thirdly, the basic algorithm for the fair allocation of resources based on the number of GPUs is given. Fourthly, the scheduling algorithm with the smallest cost and the largest number of tasks is proposed, which describing the implementation strategies of resource non-grabbing, robbing and resource fairness strategies. Finally, six data-intensive computing tasks are designed, and the algorithm proposed in the paper is tested, and the experiments verifies the scheduling algorithm can achieve about 90% of resource fairness, while also ensuring that the parallel operation time of jobs is minimized.

    参考文献
    相似文献
    引证文献
引用本文

汤小春,朱紫钰,毛安琪,符莹,李战怀.数据密集型计算作业在GPU集群上的共享调度算法.软件学报,,():0

复制
分享
文章指标
  • 点击次数:
  • 下载次数:
历史
  • 收稿日期:2020-05-10
  • 最后修改日期:2020-11-30
  • 录用日期:
  • 在线发布日期: 2021-05-21
您是第位访问者
版权所有:中国科学院软件研究所 京ICP备05046678号-3
地址:北京市海淀区中关村南四街4号,邮政编码:100190
电话:010-62562563 传真:010-62562533 Email:jos@iscas.ac.cn
技术支持:北京勤云科技发展有限公司

京公网安备 11040202500063号