摘要
针对分布内存结构的并行化将串行程序转变为在各处理节点上运行的SPMD并行程序,节点程序包含该节点所执行的运算和与其它节点交换信息的通信操作。讨论了在已知数据分解和计算划分的前提下生成分布内存结构下的消息传递并行程序的算法,以Lam提出的线性不等式基本框架为基础,在Paraguin工作基础上进行了有效的改进:第一,在代码生成算法中引入了数据分布;第二,将处理器空间由一维扩展到多维;第三,将虚拟处理器到物理处理器的映射关系引入代码生成算法,从而减少了节点间通信的数量,提高了生成并行代码的性能。
In parallelisation for distributed_memory architecture,the sequential program is transformed to SPMD program to run on each computing node.Each node program consists of computing and communication code.In this paper,algorithm for automatic message passing code generation is discussed.The algorithm is based on the framework of linear inequalities systems.It improves the work of Paraguin project by introducing data decomposition into the algorithm of code generation,extending processor space from one dimension to multitude dimensions and introducing the map function of virtual processors to physical processors into the linear inequalities system.
出处
《计算机工程与应用》
CSCD
北大核心
2007年第8期74-77,共4页
Computer Engineering and Applications
关键词
分布内存结构
计算划分和数据分布
虚拟处理器
物理处理器
线性不等式系统
distributed_memory architecture
compute and data decomposition
virtual processor
physical processor
linear inequalities system
