A parallel accelerated approach of HMM Forward Algorithm for IBM Roadrunner clusters

In this work we propose an efficient parallel approach of Hidden Markov Model classification based on Cell Broadband Engine (Cell/B.E.) architecture. The Cell/B.E. technology offers us the possibility of developing two level parallelization algorithms due to the hybrid architecture feature of this kind of processors. Based on this concept, we developed and analyzed two different techniques of parallelization of Forward Algorithm using the Message Passing Interface (MPI) model and the MPI combined with the Synergistic Processing Elements (SPEs) benefits on the Cell Broadband Engine processors. We evaluate the performance of our parallel approach for Markov models with long sequence length and large state spaces on an IBM Roadrunner cluster equipped with PowerXCell8i processors. We obtained encouraging results in terms of parallel efficiency gained on Cell/B.E. processors.