摘要
数据流计算机解决了多机系统存储器延时和同步２个不易解决的基本问题，可以充分利用程序的各级并行成分．这种计算机在算术运算等方面性能优越，但在结构数据（如向量）处理方面开销大、效率低．将向量处理技术引入数据流计算机，能有效地解决这个问题．通过在类似Ｍａｎｃｈｅｓｔｅｒ的数据流计算机上增加向量处理功能，并进行了矩阵乘运算的模拟．结果表明，增加向量处理后，矩阵乘运算的效率比近几年提出的Ｍｏｎｓｏｏｎ和Ｅｐｓｉｌｏｎ等数据流计算机提高一倍，甚至２倍以上．
The dataflow computer explicitly addresses the two main issues: memory latency and synchronization. In MIMD computer, it can make use of every level of parallelism in programs. This kind of computer showed the ability to surpass the control flow computer in arithmetic computation, etc.. But it is of low efficiency in processing structure data. This problem could be solved by adding vector processing to the dataflow computer. The simulation was carried out for matrix multiply on an analogue Manchester dataflow computer to which the function for vector processing is added. The result shows that the efficiency is greatly raised and is 1 or 2 times higher than that of Monsoon and Epsilon dataflow computer.
出处
《西安交通大学学报》
EI
CAS
CSCD
北大核心
1997年第6期38-44,共7页
Journal of Xi'an Jiaotong University
