Are nonblocking networks really needed for high-end-computing workloads?