Program Information
Benchmarking and Optimization of a Monte Carlo Based Robust Optimization System for Intensity Modulated Proton Therapy (IMPT) On Multi-GPU Cluster
A Abdel-Rehim*, H Wan Chan Tseung , J Ma , H Kamal Sayed , M Herman , C Beltran , Department of Radiation Oncology, The Mayo Clinic, Rochester, MN 55905
Presentations
SU-I-GPD-T-161 (Sunday, July 30, 2017) 3:00 PM - 6:00 PM Room: Exhibit Hall
Purpose: To benchmark and optimize the performance of a Monte Carlo based robust optimization system for Intensity Modulated Proton Therapy (IMPT) on multi-GPU cluster.
Methods: A set of cases with varying complexity was used to benchmark a locally developed multi-GPU Monte Carlo and robust optimization code which also includes an LET dependent biological model for IMPT. The computer cluster used has multiple nodes each equipped with four Tesla K80 GPUs connected through an InfiniBand network. Cases considered include Head and Neck, prostate and pediatric brain. Both physical (RBE=1.1) and biological dose constraints were considered. For the robust optimization, 9 scenarios with positional shifts of +/-3mm and 3% range uncertainty were considered. Each part of the code is timed for varying number of GPUs. The main bottlenecks and how to resolve them were identified in order to make the usage of the optimizer clinically viable.
Results: The two major components of the total compute time were the production and storage of the dose influence map with Monte Carlo and the numerical optimization procedure. Two important bottlenecks were found to be the Monte Carlo computational time and storage of the dose influence map. To reduce the Monte Carlo computing time, we down-sample the CT by a factor of 2 in each direction, giving a 2.5x2.5x2.0 mm3 grid, before calculating the dose influence map. A factor of 3 reduction in computational time was measured. In order to reduce data movement between the GPU and the CPU as well as avoiding storing data to disk during the optimization, the Monte Carlo and optimizer were fully integrated and efficient usage of memory on the GPU was implemented.
Conclusion: With a properly optimized Monte Carlo and optimization software, robust optimization of IMPT planning can be made clinically viable using a relatively small GPU cluster.
Contact Email: