计算机工程 ›› 2012, Vol. 38 ›› Issue (17): 268-271.doi: 10.3969/j.issn.1000-3428.2012.17.072

• 开发研究与设计技术 • 上一篇    下一篇

一种GPU集群的动态任务映射策略

陈庆奎,曹欢欢,庄松林   

  1. (上海理工大学光电信息与计算机工程学院,上海 200093)
  • 收稿日期:2011-10-31 修回日期:2011-12-23 出版日期:2012-09-05 发布日期:2012-09-03
  • 作者简介:陈庆奎(1966-),男,教授、博士生导师,主研方向:视频解码,并行处理;曹欢欢,硕士研究生;庄松林,教授
  • 基金项目:
    国家自然科学基金资助项目(60970012, 61003031);教育部高等学校博士学科点专项科研基金资助项目(20113120110008);上海信息技术领域重点科技攻关计划基金资助项目(09511501000, 09220502800);上海市教育科学研究基金资助项目(B11042)

A Dynamic Task Mapping Strategy of GPU Cluster

CHEN Qing-kui, CAO Huan-huan, ZHUANG Song-lin   

  1. (School of Optical-electrical and Computer Engineering, University of Shanghai for Science & Technology, Shanghai 200093, China)
  • Received:2011-10-31 Revised:2011-12-23 Online:2012-09-05 Published:2012-09-03

摘要: 在3G网络的视屏计算中,需要解决逻辑复杂和数值运算密集的难题。为此,提出一种集中式动态任务映射策略。采用消息驱动方式,将任务初始分配,根据消息的内容进行实时优化分配,利用GPU集群的任务特征,设计对应的效益函数,以衡量任务映射策略的优劣。实验结果表明,该策略可以较好地协同各部分资源完成计算任务,并能适应集群的变化,在集群规模扩大或是有节点故障时均有比较稳定高效的性能。

关键词: 图像处理单元, 动态任务映射, 效益函数, 标准因子, 可扩展性, 任务池

Abstract: The calculation of the screen in the 3G network, needs to solve the complex logic and numerical computation intensive problem, this paper proposes a centralized dynamic task mapping strategy, which uses message-driven approach to dynamically update the task queues. It can adapt to the scalability of clusters well. For the purpose of measuring the pros and cons of the strategy, this paper proposes a suitable benefit function according to the characteristics of the cluster’s task. Experimental results show that this strategy can be well adapted to the scalability of the cluster and have a good performance on practical task mapping with high revenue and throughput.

Key words: Graphics Processing Unit(GPU), dynamic task mapping, benefit function, standard factor, scalability, task pool

中图分类号: