Techniques et outils pour les communications et la répartition dynamique de charge dans les réseaux de stations de travail

摘要

In this thesis, we focus on technics and tools that help in building and optimizing parallel and distributed applications for networks and clusters of workstations. The first problem we study is dynamic load balancing in a strongly heterogeneous environment: an efficient dynamic load balancing strategy requires first to evaluate and compare the availability of each of the workstations and second to find a good matching between this availability and the resources requirements of the applications. For this purpose, we propose an empirical method whose goal is to model the behavior of the compo- nents of a network of workstations according to the workload levels being observed. This method allows the construction of multi-dimensional and multi-criteria workload indices. To put this method in practice, we have developed LoadBuilder, a distributed environment especially intended to drive distributed experiments. The second problem we study is accessing reliable ordered multi-point communication mechanisms and protocols in UNIX operating systems. In order to deal with the communications requirements of distributed applications, we propose an original solution, the MPCFS virtual file system. This file system provides its users a way of (i) creating and managing process groups and (ii) exchanging multi-points messages in these groups. These operations can be achieved easily, through the usual UNIX file and directory operations. We have developed a dynamically loadable module for the Linux 2.0 operating system, that implements a prototype of this file system.