A Hierarchical Approach for Parallelization of a Global Optimization Method for Protein Structure Prediction

We discuss the parallelization of our protein structure prediction algorithm on distributed-memory computers. Because the computation can be represented as a search through a vast tree of possible solutions, a hierarchical approach that assigns subtrees to different groups of processors allows us to partition the work efficiently and maintain information updated without incurring significant communication overhead. Our results show that a dynamic strategy for load balancing outperforms the static one.