Shared-memory alternatives for parallel image reconstruction

Shared-memory multiprocessor computers are nowadays ubiquitous and can deliver substantial computing power. However, users are not always able to exploit this computing power because they either lack an in-depth knowledge of parallel computation or the tools they have In this paper, we investigate two new tools for shared-memory multi-threading programming , namely OpenMP and Partitioned Global Address Space, both of which promise to ease the burden of designing and implementing parallel image reconstruction algorithms. With these tools we implemented EM-ML and Separable Surrogate MAP reconstruction programs for PET and compare their performance against the traditional Pthreads multi-threading approach. We show that these new tools can parallelize complicated reconstruction code with the simply addition of a few new language directives, and more importantly, that this can be done incrementally starting from existing sequential code. We measure reconstruction execution time in a 24 core computer for two small animal PET imaging data sets and compute speedup gains.