Dimensioning the heterogeneous multicluster architecture via parallelism analysis and evolutionary computing

In the near future, embedded systems containing hundreds of processing elements running multiple concurrent applications will become a reality. The heterogeneous multicluster architecture enables to cope with the challenging hardware/software requirements presented by such systems. This paper shows principles and optimization of multicluster dimensioning aiming at an appropriate distribution of applications onto clusters containing different types of processing elements. The approach represents an initial exploration phase efficiently finding a suitable multicluster configuration in the large design space. Hence, results should be further refined by more accurate but less time-efficient simulation-based techniques. As the starting point, a parallelism value matrix is analytically extracted describing application mappings independently on the architecture and scheduling. A genetic algorithm (GA) and a mixed-integer linear programming (MILP) approach solving the dimensioning problem are introduced and compared. Both solutions use the parallelism value matrix as input. Scalability results show that the GA generates results faster and with a satisfactory quality relative to the found MILP solutions. Finally, the dimensioning approach is demonstrated for a realistic benchmark scenario.