SPMD cluster-based parallel 3D OSEM

Abstract

This study empirically compares two approaches to parallel 3D OSEM that differ as to whether calculations are assigned to nodes by projection number or by transaxial plane number. For projection space decomposition (PSD), the forward projection is completely parallel, but backprojection requires a slow image synchronization. For image space decomposition (ISD), the communication associated with forward projection can be overlapped with calculation, and the communication associated with backprojection is more efficient. To compare these methods, an implementation of 3D OSEM for three PET scanners is developed that runs on an experimental 9-node, 18-processor cluster computer. For selected benchmarks, both methods exhibit speedups in excess of eight or nine nodes, and comparable performance for the tested range of cluster sizes.