Parallel Monte Carlo simulation and model reduction

Abstract

The high portability and robustness of Monte Carlo algorithm come with its high computational cost, as every dimension in the problem space needs to be sampled randomly. Therefore, parallelizing the Monte Carlo algorithm is necessary for improving performance.

We investigated the simulation problem for predicting columnar clustering. A reduced model was built using logistic regression to improve the computing performance. A large number of computer experiments were performed on Blue Gene Q and the results were used to train the logistic regression model using the machine learning library in Spark. The logistic regression model was evaluated by area under receiver operating characteristics curve, precision and recall.

In this work, a C++ software package is developed for running large scale Potts Monte Carlo simulation in parallel based on the MMSP framework[1]. The package is applied to model material microstructure evolution. Specifically, a new feature of parallel biased Monte Carlo sampling is developed and implemented. The biased Monte Carlo sampling can be used to simulate field gradients, e.g. temperature gradients. This enables solving more general problems, since gradients often present in physical processes. Also, we validated the parallel algorithm by two test cases. Besides, the scalability of the parallel algorithm was investigated on Blue Gene Q at RPI with different computing configurations.