Multilevel Parallelized Forward Algorithm for Hidden Markov Models on IBM Roadrunner Clusters

In this work we propose an efficient parallel algorithm to evaluate an observation sequence on Hidden Markov Model starting from the sequential Forward Algorithm (FA). The Cell Broadband Engine (Cell/B.E.) hybrid architecture, allows us to approach two levels of parallelization in developing our algorithms. Two strategies were implemented and tested in order to obtain a parallel version of the FA, using solely the Message Passing Interface (MPI) and, in the second case, using the MPI along with the Synergistic Processing Elements (SPEs) on the Cell/B.E. Processors. Performance tests were run for the two approached techniques on an IBM Roadrunner cluster with Power XCell8i processors. We found that our parallel versions of FA performed approximately 37 times faster on Cell/B.E. Processors without SPE cores, and 75 times faster with SPE cores the, compared to the serial algorithm.