The International Conference for High Performance Computing, Networking, Storage and Analysis
New Parallelization Model of Sequential Monte Carlo Analysis with Prediction-Correction Computing.
Authors: Eiji Tomiyama (Research Organization for Information Science and Technology), Hiroshi Koyama (Research Organization for Information Science and Technology), Katsumi Hagita (National Defense Academy)
Abstract: Monte Carlo (MC) search with Metropolis judgment has been used widely to solve inverse problems; in particular, reverse Monte Carlo (RMC) analysis has been shown to make a possible configration of particles in terms of the structure factor S(q) obtained by X-ray scattering experiments. In general, MC search proceeds sequentially, because each judgement of the MC trial depends on its previous state. In the present study, we have developed a new “SimpleRMC” parallel code for RMC analysis for a system with millions of particles. In this code, two hotspots are identified and optimized for the histogram h(r) and difference of histogram Δh(r) calculations. For the histogram kernel, we achieved 339 Tflops performance (32.3% of the peak) on 8,192 nodes of the K computer using a 33,554,432-particle system. This newly developed “prediction–correction” method improves both the parallel performance of the Δh(r) calculation and its elapse time.