Parallel Gaussian Process Regression for Big Data: Low-Rank Representation Meets Markov Approximation