Design Methodologies for Developing Multiprocessor Super- computers

本项目研究了两种设计方法，动态收缩化和宏观数据流，在开发多处理器超级计算机中很有用。多处理器单元收缩的概念是由克雷X-MP实现的流水线链推广而来的。该技术为设计可重构收缩阵列提供了一种新的网络方法，用于快速计算复合函数和特殊算法。宏数据流是指一种事件驱动的方法来开发数据流超级计算机。宏数据流可以有效地支持可变粒度的多任务处理。这些研究包括动态收缩和宏观数据流的原理、功能机制的发展、多处理软件的支持、系统重构技术、通过仿真实验的多任务基准研究以及超级计算机性能评估。主要应用领域强调偏微分方程（PDE）问题的求解。

This project investigates two design methodologies, dynamic systolization and macro-dataflowing, useful in developing multiprocessor supercomputers. The concept of multiprocessor element systolization is generalized from pipeline chaining implemented in Cray X-MP. This technique offers a new network approach to designing reconfigurable systolic arrays for fast evaluation of compound functions and special algorithms. Macro-dataflowing refers to an event-driven approach to developing dataflow supercomputers. Macro-dataflow is effective to support multitasking with variable granularity. The studies include principles of dynamic systolization and macro-dataflowing, development of functional mechanisms, multiprocessing software supports, system reconfiguration techniques, multitasking benchmark studies through simulation experiments, and supercomputer performance evaluations. The major application area emphasizes the solution of Partial Differential Equation (PDE) problems.