Program Information
Coupled Photon-Electron Monte Carlo Simulation for Heterogeneous Computing Systems
E Doerner*, C Rebolledo , V Gomez , Institute of Physics, Pontificia Universidad Catolica de Chile, Santiago, Chile
Presentations
SU-I-GPD-T-352 (Sunday, July 30, 2017) 3:00 PM - 6:00 PM Room: Exhibit Hall
Purpose: To develop a Monte Carlo platform that enables the coupled photon-electron transport simulation for heterogeneous computing systems through the OpenCL framework.
Methods: Our platform, named OCL_MC, relies on the physical model implemented in EGSnrc and adapts it for execution across heterogeneous platforms consisting of central processing units (CPUs) and graphics processing units (GPUs) using OpenCL. The simulation process, memory-access scheme and data transfer mechanism across computing devices have been optimized in order to exploit the potential of the underlying hardware, without sacrificing the reliability of the original implementation. Runtime and accuracy of OCL_MC was tested against a selection of phantoms and two clinical beams of 6MV and 15 MeV, read from phase space files. All phantoms consist of 61x61x150 voxels with a size of 0.5x0.5x0.2 cm3. We used a homogenous water phantom and two water phantoms with a layer of bone and lung ranging from z = 5 to 10 cm. Performance tests were carried out two systems consisting on one Intel Core i7-4820K processor with two AMD Radeon HD7970 and two Intel Xeon E5-2620 16 Core with three NVIDIA Tesla K40 cards, respectively.
Results: We could demonstrate that the dose distributions calculated with OCL_MC are in excellent agreement with EGSnrc. Our implementation scales very well across the different processors available in the studied heterogeneous platforms achieving a speed-up of up to 13x compared to the original EGSnrc parallel implementation.
Conclusion: In this work we presented an OpenCL based Monte Carlo platform that is capable of leverage the potential of modern heterogeneous computing systems consisting of different kinds of processors. Thanks to this capability, OCL_MC is capable of obtaining a significant speed-up when compared with the original implementation without introducing any approximation to the underlying physics that would compromise the accuracy and reliability of the dose calculations.
Funding Support, Disclosures, and Conflict of Interest: The authors would like to acknowledge project FONDECYT Iniciacion No. 11150094 for financial support.
Contact Email: