Optimizing a conjugate gradient solver with non-blocking collective operations