Over the last few decades, feature selection has been a hot research area in pattern recognition and machine learning, and many famous feature selection algorithms have been proposed. Among them, feature selection using positive approximation(FSPA) is an accelerator for traditional rough set based feature selection algorithms, which can significantly reduce the running time. However, FSPA still cannot handle large scale and high dimension dataset due to the memory constraints. In this paper, we propose a parallel implementation of FSPA using MapReduce framework, which is a programming model for processing large scale datasets. The experimental results demonstrate that the proposed algorithm can process large scale and high dimension dataset efficiently on commodity computers.