Introduction

Most of resources in a grid are parallel, either clusters of workstations or parallel machines. Computational grids are even considered as hierachical sets of parallel resources, as we can see in ongoing project like the french research grid project, Grid'5000 [2] (for the moment, 9 sites are involved), or like the EGEE^6.1project (Enabling Grids for E-science in Europe), composed of more than a hundred centers in 48 countries. Then, in order to provide transparent access to resources, grid middleware must supply efficient mechanisms to provide parallel services.

Because parallel resources are managed differently on each site, it is neither the purpose of DIET to deal with the deployment of parallel tasks inside the site, nor manage copies of data which can possibly be on NFS. DIET implements mechanisms for a SeD programmer to easily provide a service that can be portable on different sites; for clients to request services which can be explicitly sequential, parallel or solved in the real transparent and efficient metacomputing way: only the name of the service is given and DIET chooses the best resource where to solve the problem.