In the last decade, Compressive Sensing (CS) has emerged as the most promising, model-driven approach to accelerate MRI scans. CS relies on the key sparsity assumption and proposes random sampling for data acquisition. The practical CS approaches in MRI employ variable-density (VD) sampling, where samples are drawn at random based on a parametric probability model which focuses on the center of the Fourier domain. In stark contrast to this model-driven sampling approaches, we propose a data-driven framework for optimizing sampling in parallel (multi-coil) MRI. Our approach does not assume any structure in the data, and instead optimizes a performance metric (e.g. PSNR) for any given reconstruction algorithm, based on our earlier learning-based sampling framework previously applied to 2D MRI which we also extend to 3D MRI setting in this work by employing lazy evaluations in the greedy algorithm. We show boosted performance for the parallel MRI based on this sampling approach and highlight the inefficiency of variable density approaches. This suggests that data-driven sampling methods could be the key to unlocking the full power of CS applied to MRI.