Decision-theoretic exploration of multiProcessor platforms

In this paper, we present an efficient technique to perform design space exploration of a multi-processor platform that minimizes the number of simulations needed to identify the power-performance approximate Pareto curve. Instead of using semi-random search algorithms (like simulated annealing, tabu search, genetic algorithms, etc.), we use domain knowledge derived from the platform architecture to set-up exploration as a decision problem. Each action in the decision-theoretic framework corresponds to a change in the platform parameters. Simulation is performed only when information about the probability of action outcomes becomes insufficient for a decision. The algorithm has been tested with two multi-media industrial applications, namely an MPEG4 encoder and an Ogg-Vorbis decoder. Results show that the exploration of the number of processors and two-level cache size and policy, can be performed with less than 15 simulations with 95% accuracy, increasing the exploration speed by one order of magnitude when compared to traditional operation research techniques.