Affiliations: [a] Radiation Medicine Program, Princess Margaret Hospital and Department of Radiation Oncology, University of Toronto, Toronto, ON, M5G 2M9, Canada | [b] Department of Physics, Ryerson University, Toronto, ON, M5B 2K3, Canada. E-mail: james.chow@rmp.uhn.on.ca
Abstract: This study investigated the performance of Cell processor in Monte Carlo simulation using the EGSnrc code. The execution speeds of two processor-intensive functions namely, HOWNEAR and RANMAR_GET in the EGSnrc code based on the informal 20–80 rule were measured, using the profiler gprof, which specified the number of executions and total time spent in the function. The functions were re-implemented in a testing architecture designed for Cell processor using a Playstation3 (PS3) system. This specific framework was designed to divide the program functionality in two-isolating branching code for execution on the powerPC processing element (PPE) and synergistic processing elements (SPEs) in the system. Depth dose profiles in heterogeneous phantoms using photon and electron beams calculated by a CPU and Cell processor were compared and there was no difference in the results. It was found that our examined algorithms could be parallelizable on the Cell processor, provided that an architectural change of the EGSnrc was made so that the code was initialized on the PPE, and execution of the simulation loop within the program be split between the PPE and SPE. Since it was found that the EGSnrc performance was currently limited by the PPE using the Cell processor, a computing platform of PC coupled with a general purpose programming on graphics processing units or GPGPU may provide a more viable avenue for acceleration.
Keywords: Cell processor, Monte Carlo simulation, performance, parallel computing, GPU, PS3, EGSnrc
