Parallel programming for OSEM reconstruction with MPI, OpenMP, and hybrid MPI-OpenMP

To improve the parallel efficiency (PE) of the Ordered-Subsets Expectation-Maximization (OSEM) algorithm for 3D PET image reconstruction, we implemented the algorithm with 1) an OpenMP and 2) a hybrid Message Passing Interface (MPI)-OpenMP model on the basis of a standard MPI implementation. The motivation was to reduce the interprocessor data exchange time which was the dominant PE limiting factor of the MPI model when large number of processors was used. The OpenMP model used a fine-grained approach and showed significant speedup only up to 2-3 processors for both the true shared memory and the single system image (SSI) distributed shared memory architectures. The hybrid MPI-OpenMP model achieved a consistent improvement of similar to 10% in terms of speedup factor on a large number of parallel processors compared to the pure MPI approach. As clusters of larger symmetric multiprocessor (SMP) machines continue to become more cost effective, we expect this hybrid MPI-OpenMP approach to be increasingly valuable to accelerate 3D PET reconstructions, and other applications with similar computational characteristics.