位置:成果数据库 > 期刊 > 期刊详情页
Godson-T众核体系结构上的Broadcast性能优化
  • 期刊名称:计算机研究与发展, 2009.
  • 时间:0
  • 分类:TP302[自动化与计算机技术—计算机系统结构;自动化与计算机技术—计算机科学与技术]
  • 作者机构:[1]北京交通大学软件学院,北京100044, [2]中国科学院计算技术研究所计算机系统结构重点实验室,北京100190, [3]北京交通大学计算机与信息技术学院,北京100044
  • 相关基金:国家“九七三”重点基础研究发展规划项目基金(2005CB321600);国家自然科学基金重点项目(60736012);中国科学院计算技术研究所计算机系统结构重点实验室开放课题
  • 相关项目:高性能片上存储系统
中文摘要:

Godson—T是中国科学院计算技术研究所计算机系统结构重点实验室先进微系统组正在研制开发的适合于超深亚微米工艺实现的大规模片上众核系统.Godson—T片上存储的单端口结构节省了芯片面积但制约了共享数据的读取效率.直接在Godson—T上实现传统的Broadcast算法需要大量的同步互斥开销,无法达到很好的性能提升.基于Godson-T体系结构,对数据共享的重要并行算法Broadcast进行优化,提高了Godson-T体系结构下的数据共读的效率.主要采取了以下3项技术:消除大规模的线程同步,建立源地址到目的地址的映射表和用汇编语言实现Broadcast的核心部分.优化后Broadcast在小核数为32时即可达到5.8倍加速比.

英文摘要:

Godson-T is a large scale many-core system architecture to be implemented by ultra-deep submicron MOS technology under development by the Group of Advanced Microsystem in the Key Laboratory of Computer System and Architecture of the Institute of Computer Technology, Chinese Academy of Sciences. The single port design of Godson-T on-chip memory saves the chip's total area but limits the efficiency of data sharing. Broadcast is a basic parallel algorithm used to accelerate data sharing process, but implementing the traditional algorithm on Godson-T requires a large amount of synchronization and mutual exclusion expenses and therefore could not bring a good performance. Based on Godson-T system architecture, the authors optimize the important parallel algorithm Broadcast and enhance the efficiency of concurrent read. Three techniques are proposed for the optimization: eliminating bulk synchronization among threads, establishing mapping table between source addresses and destination addresses, and rearranging assembly instructions in Broadcast kernel. The first one reduces expenses of synchronizing a large amount of threads, the second one provides a quicker method for destination address search, and the last one fully makes use of the advantage of Godson-T architecture. The optimized Broadcast algorithm on Godson-T system architecture performs well; especially when core number is 32, the speedup of the algorithm can reach 5.8.

同期刊论文项目
期刊论文 118 会议论文 65
同项目期刊论文