Research Initiation Award: A Communication Library for Parallel Computers

通信效率是大规模并行计算取得广泛成功的关键。通过查看并行计算的成功案例可以看出这一点，这些成功案例目前仅限于具有较小通信需求或使用少量处理器的应用程序。为了在更广泛的应用中使用大规模并行计算，必须开发有效的算法来执行底层处理器间通信。本研究的目的是创建一个算法库，以在几种多处理器拓扑结构中执行一些通用通信任务。将检查应用程序中经常出现的通信任务和流量模式。重点是寻找易于实现的通信算法，以便在最优或接近最优的时间内执行这些任务。这些算法可以被多处理器计算机编译器的程序员作为通信原语调用，就像在传统计算机的函数库中调用实现标准函数的子程序一样。＊＊＊

9309098 Varvarigos Communication efficiency is the key to the broad success of massively parallel computation. One can see that by looking at the success stories of parallel computation, which are currently limited to applications that have small communication requirements, or use a small number of processors. In order to use massively parallel computation for a broader range of applications efficient algorithms to execute the underlying inter processor communications have to be developed. The objective of this research is to create a library of algorithms to execute a number of generic communication tasks in several multiprocessor topologies. Communication tasks and traffic patterns that arise often in applications will be examined. Emphasis will be given to finding easily implementable communication algorithms to execute these tasks in optimal or near-optimal time. These algorithms can be called as communication primitives by the programmer of the compiler of a multiprocessor computer in the same way that subroutines implementing standard functions are called from a library of functions in a conventional computer. ***