Achieving high-performance with a sparse direct solver on Intel KNL