Monte Carlo Dose Calculation for TomoTherapy Using Cloud Computing
Q Chen1*, M Humphrey1, K Ding1,2, E Sterpin3, P Read1, J Larner1, (1) University of Virginia, Charlottesville, VA, (2) Johns Hopkins University , Baltimore, MD, (4) Universite Catholique de Louvain, Brussels, BelgiumSU-E-T-521 Sunday 3:00PM - 6:00PM Room: Exhibit Hall
Purpose: Monte Carlo (MC) dose calculation has been regarded as the gold standard for accuracy, especially in the presence of patient in-homogeneities. However, the calculation can be very time consuming as a large number of particle histories are needed to reduce the statistical noise to a level acceptable for clinical use. The purpose of this work is to explore the option of cloud computing to achieve ultra-fast MC calculation.
Methods: A MC package developed for TomoTherapy, TomoPen, was ported to run on Amazon Elastic Cloud (EC2). The executable as well as the phase space file for the MC program was packaged into an Amazon machine image (AMI), which was loaded from Amazon Elastic Block Store (EBS) to each cloud node during initialization process. Message-Passing-Interface (MPI) was used to distribute computing tasks to different nodes. Statistics independence of each node was achieved by reading from different location of the phase-space files. Fifty clinical plans were used to benchmark the cloud computing performance.
Results: Benchmarking cases were tested on EC2 with different type and number of computing instances. It was found that the program run time scaled inversely with the number of computing instances. The most cost-effective EC2 instance was found to be the extra-large high-CPU instance. An EC2 cloud with 32 extra-large high-CPU computing instance was used to compute MC dose for 50 patient plans. On average, it took 22 seconds total to transfer CT to/dose from cloud and 117 seconds calculation time to get a MC dose with statistical uncertainty of less than 2%. The total cost was less than $3 for one hour when spot instance is used.
Conclusion: Cloud computing provides a promising platform to regroup computing resources on-demand to improve the speed of MC calculation at very little cost.
Funding Support, Disclosures, and Conflict of Interest: This study is supported in part by UVa George Amorino Pilot Grant.