Optimizing the Scalability of Parallelized GATE Simulations

GATE is a GEANT4 application toolkit for accurate simulation of positron emission tomography (PET) and single photon emission computed tomography (SPECT) systems. As Monte Carlo simulations are CPU-intensive, simulations often take up to several days to complete with state-of-the-art single-CPU computers. However, Monte Carlo simulations are also excellently suited for parallelization, theoretically showing a linear speed-up as a function of the number of processing nodes. Previously a parallel computing platform was developed in order to reduce the overall computing time of GATE experiments. It was comprised of a job splitter, to subdivide the simulation using time-domain decomposition and a file merger to merge the data output. However, at that time three factors limited the scalability of this platform. Firstly, isotopes with a short half life led to an inefficient load balancing, as an unequal amount of events was processed on each node. Secondly, a long setup time was required for SPECT collimator geometries. Thirdly, the merge overhead of the single file merger greatly limited the scalability. This paper provides a solution for each of these problems, thereby improving the overall scalability of the aforementioned cluster platform significantly.