A Divide-and-Conquer Procedure for Sparse Inverse Covariance Estimation