High performance GPU implementation of k-NN based on Mahalanobis distance

The k-nearest neighbor (k-NN) is a widely used classification technique and has significant applications in various domains. The most challenging issues in the k-nearest neighbor algorithm are high dimensional data, the reasonable accuracy of results and suitable computation time. Nowadays, using parallel processing and deploying many-core platforms like GPUs is considered as one of the popular approaches to improving these issues. In this paper, we present a novel and accurate parallel implementation of k-NN based on Mahalanobis distance metric in GPU platform. We design and implement k-NN for GPU architecture and utilize mathematic and algorithmic techniques to eliminate repetitive computations. Moreover, in addition, to taking advantage of different parallelism techniques, we improve warp management to gain maximum speed up in this implementation. Via Compute Unified Device Architecture (CUDA)-enabled GPUs, the acceleration is considerable as experimental results show the 110X speedup with respect to the single core CPU implementation. Furthermore, we measure the energy and power consumption of this algorithm for both CPU and GPU platforms, where GPU is more energy efficient regarding this application.