Les riches heures de l'ordonnancement

Abstract

Parallelizing means executing simultaneously multiple operations. The efficient management of the assignment of tasks on available computing resources is a challenging problem. Whatever the considered execution support, this problem is crucial at any level from basic instructions, library components to full applications. Moreover, it can be studied from various points of view (users, system administrators). Scheduling is the main part for efficient resource management. It consists in determining where and when to execute the tasks while optimizing one or several objectives under hardware constraints. This is an old problem which arised with the first vector machines in the early seventies. Today, it is still a challenging problem since many new constraints should be considered. The purpose of this paper is to review the main constraints and to show how to take them into account in the scheduling policies. We defend a position of well-founded approaches for designing scheduling algorithms that allow to obtain guaranteed performances at a reasonable cost.