Précédent : Références Remonter : Projet APACHE

Abstract

Parallel architectures offer a tremendous potential in terms of processing power and memory capacity. Technological hardware advances allow computation on very efficient multiprocessors with different levels of integration: symmetric multiprocessors, multicomputers, networks of workstations, etc. Despite this progress, there is a severe lack of satisfying solutions for exploiting efficiently this potential using automated tools.

In this context, the APACHE research project works in the area of parallel system programming for high performance computing. The goal of the project is to develop a number of techniques, algorithms and tools which provide a beneficial tradeoff between performance and portability and the adaptation to the current load of the machine. This work addresses regular as well as irregular applications and encompasses platforms such as parallel supercomputers as well as networks of workstations.

Part of the research work is dedicated to the design, programming and testing of a parallel executive kernel which allows easy control of the load among the processors. As the user need not be aware of this control, static or dynamic scheduling techniques and load balancing algorithms are studied and implemented in this context. An applicative user interface is designed, which encapsulates the load distribution tools and permits the parallelism of an application to be expressed. This interface induces a programming style and is based on C++.

As a support, adaptive debugging and performance tools are developed. They allow to collect, analyze and filter relevant information. Finally, new algorithms are proposed in various application areas and are implemented on the ATHAPASCAN (the name of the APACHE language) platform.