About fault-tolerant allocation of tasks in multiprocessor architectures and system dimensioning

This paper deals with dimensioning a multiprocessor system and allocating tasks on it: easy reconfiguration and reallocation after hardware faults are the main goals. Computation is modelled as a set of tasks sequentially executed. A compact representation of the computation is obtained by means of Timed Petri Nets: transitions represent the tasks, while places contain data exchanged between tasks. Arcs give the execution ordering. Transitions are labelled by execution time of correspondent tasks. Parallel execution is allowed: tasks belonging to different branches may be performed concurrently. A heuristic algorithm for dimensioning of hardware architecture is proposed to find a configuration allowing a task allocation which minimizes costs and satisfies predefined performance constraints. Task allocation, when the architecture of the system is given, is also discussed.