A multi-objective decision-theoretic exploration algorithm for platform-based design

Abstract

This paper presents an efficient technique to perform multi-objective design space exploration of a multiprocessor platform. Instead of using semi-random search algorithms (like simulated annealing, tabu search, genetic algorithms, etc.), we use the domain knowledge derived from the platform architecture to set-up the exploration as a discrete-space multi-objective Markov Decision Process (MDP). The system walks the design space changing its parameters, performing simulations only when probabilistic information becomes insufficient for a decision. The algorithm employs a novel multi-objective value function and exploration strategy, which guarantees high accuracy and minimizes the number of necessary simulations. The proposed technique has been tested with a small benchmark (to compare the results against exhaustive exploration) and two large applications (to prove effectiveness in a real case), namely the ffinpeg transcoder and pigz parallel compressor. Results show that the exploration can be performed with 10% of the simulations necessary for state-of-the-art exploration algorithms and with unrivaled accuracy (0.6 ± 0.05% error).