Towards an Algorithm and Communication Cost Model for the Parallel Particle Swarm Optimization

Inhalt:
The adaptation of sequential algorithms for High Performance Computing (HPC) systems is determined by a tradeoff between algorithmic effectiveness (software) and communication frequency (hardware) of the parallel implementation (efficiency). To get a better understanding of the correlation, we define simple models for both, software and hardware, in order to dynamically find the best mapping parameters for the execution of the algorithm on the parallel system. For the evaluation of our method we look at population-based algorithms like the Particle Swarm Optimization Algorithm (PSO) for the solution of optimization problems. Different goals like best quality of the solution of the optimization problem for a given execution time or best execution time to find the optimum are defined by the user. Our method enables us to find the best parameters for the mapping which then results in an efficient and effective parallel implementation to achieve the user-defined goals on a High Performance Computing Cluster (HPCC).