Improving Performance of Sparse Numerical Linear Algebra Computations