Section: Scientific Foundations
Transparent Resource Management
We think of the future Grid as of a medium to access resources. This access has to be as transparent as possible to a user of such a Grid and the management of these resources has not to be imposed to him/her, but entirely done by a ``system '', so called middleware . This middleware has to be able to manage all resources in a satisfactory way. Currently, numerous algorithmic problems hinder such an efficient resource management and thus the transparent use of the Grid.
By their nature, distributed applications use different types of resources; the most important being these of computing power and network connections. The management and optimization of those resources is essential for networking and computing on Grids. This optimization may be necessary at the level of the computation of the application, of the organization of the underlying interconnection network or for the organization of the messages between the different parts of the application. Managing these resources relates to a set of policies to optimize their use and allow an application to be executed under favorable circumstances.
Our approach consists of the tuning of techniques and algorithms for a transparent management of resources, be they data, computations, networks, ...This approach has to be clearly distinguished from others which are more focused on applications and middlewares. We aim at proposing new algorithms (or improve the exiting ones) for the resource management in middlewares. Our objective is to provide these algorithms in libraries so that they may be easily integrated. For instance we will propose algorithms to efficiently transfer data (data compression, distribution or redistribution of data) or schedule sequential or parallel tasks.
The problems that we are aiming at solving are quite complex. Therefore they often translate into combinatorial or graph theoretical problems where the identification of an optimal solution is known to be hard. But, the classical measures of complexity (polynomial versus NP-hard) are not very satisfactory for really large problems: even if a problem has a polynomial solution it is often infeasible in reality whereas on the other hand NP-hard problems may allow a quite efficient resolution with results close to optimality.
Consequently it is mandatory to study approximation techniques where the objective is not to impose global optimality constraints but to relax them in favor of a compromise. Thereby we hope to find good solutions at a reasonable price. But, these can only be useful if we know how to analyze and evaluate them.