Proper parallel Monte Carlo for computed tomography of volcanoes

Abstract

Most Monte Carlo simulations can be parallelized, or at least easily distributed. However, the parallelization of such programs is not mainstream. In this paper, we propose good practices to distribute a Monte Carlo simulation for computed tomography. We apply this to large edifices (such as volcanoes or pyramids) and give a particular focus on performances. The work included first the optimization of an existing sequential prototype and second the use of reliable parallel random streams. The optimized parallel version runs on a Symmetric Multi-Processor (SMP) and shows a fine and scalable speedup. The cumulated optimizations (sequential and parallel) enabled a speedup of 400X on a 32 cores SMP.

Parallel Random Numbers, Simulation, and Reproducible Research

“…Researchers have highlighted many problems in various application frameworks, particularly in nuclear medicine simulations, which often require more than 10 20 pseudorandom numbers deployed on thousands of processors. [2][3][4][5][6][7] When considering stochastic simulations, pseudorandom numbers must be generated in parallel, so that each computing element can autonomously obtain its own stream of random numbers independently of the other computing elements. If such independence isn't guaranteed, parallelism is affected and the independence of parallel streams is questioned, with potentially serious consequences because the quality of the final simulation results could be heavily damaged.…”

“…Researchers have highlighted many problems in various application frameworks, particularly in nuclear medicine simulations, which often require more than 10 20 pseudorandom numbers deployed on thousands of processors. [2][3][4][5][6][7] When considering stochastic simulations, pseudorandom numbers must be generated in parallel, so that each …”