High-accuracy Optimization by Parallel Iterative Discrete Approximation and Multi-GPU Computing
Center for Cognitive and Brain Science, The Ohio State University
The Ohio State University, 2014
@article{zhao2014highaccuracy,
title={High-accuracy Optimization by Parallel Iterative Discrete Approximation and Multi-GPU Computing},
author={Zhao, Di},
year={2014}
}
High-accuracy optimizer is the essential part of accuracy-sensitive applications such as computational finance and computational biology, and we developed single-GPU based Iterative Discrete Approximation Monte Carlo Search (IDA-MCS) in our previous research. However, single-GPU IDA-MCS is in low performance or even functionless for optimization problems with large number of peaks because of the capability constrains of single GPU. In this paper, with the novel idea of Iterative Discrete Approximation of the cost function by multi-GPU computing, we developed two multi-GPU versions of IDA-MCS: Domain Decomposition based IDA-MCS and Local Search based IDA-MCS, with the style of Single Instruction Multiple Data by CUDA 5.0, and we exhibit the performance of multi-GPU IDA-MCS by optimizing complicated cost functions. Computational results show that, by the same number of iterations, for the cost function with large number of peaks, the accuracy of multi-GPU IDA-MCS is more than thousands of times higher than that of single-GPU IDA-MCS; computational results also show that, the optimization accuracy from Domain Decomposition based multi-GPU IDA-MCS is much higher than that of Local Search based multi-GPU IDA-MCS.
November 13, 2014 by hgpu
