Implementation of parallel programming aid on hypercube multicomputer

Abstract

A parallel programming aid, KAPPA, which provides a paralled programming environment and tools for message passing multicomputer system are implemented in this thesis. KAPPA is designed for KAICUBE II, a message passing multicomputer system of hypercube topology. The programming model for KAPPA allows the user to develop parallel programs more easily through the augmented programming language based on the C programming language. KAPPA detects dependence relations in the source program so as to generate a Macro Dataflow Graph (MDG) which encapsulates the parallel nature of a given program. Then, KAPAA schedules this graph to reduce the execution time within given number of processors. Finally, a parallel code for the target computer system is generated, which is efficient and suitable for the target system. For the performance estimation, a hypercube computer simulator is embedded in KAPPA. A graphical user interface using X-window system is provided to help users get the information easily. KAPPA relieves the large amount of the tedious and routine tasks of the programmer so that they can devote themselves to the more creative jobs like the development of the more efficient algorithms. The efficiency of the generated parallel code is high enough to be comparable with that of the code designed by the experienced programmer. Moreover, as the size and complexity of the program increases, it outperforms that of manually written code in performance. In consequence, KAPPA can be used as a workbench in which the programmer is able to develop a parallel program, translate it into the conventional parallel program of the KAICUBE II, and finally pre-execute to get the information about the performance estimates of a given program.