High Performance Computing of Fast Independent Component Analysis for Hyperspectral Image Dimensionality Reduction on MIC-Based Clusters

Fast independent component analysis (Fast ICA) for hyper spectral image dimensionality reduction is computationally complex and time-consuming due to the high dimensionality of hyper spectral images. By analyzing the Fast ICA algorithm, we design parallel schemes for covariance matrix calculating, white processing and ICA iteration at three parallel levels: multicores, many integrated cores (MIC), and clusters. Then we present a series of optimization methods for different hotspots, and measure their performance effects. All the work has been implemented in a framework called Ms-Fast ICA. Our experiments on the Tianhe-2 Supercomputer show that the Ms-Fast ICA algorithm has a good scalability, and it can reach a maximum speed-up of 410 times on 64 nodes with 192 Intel Xeon Phis.