For manufacturing of tera-scale cluster systems with thousands of processors, the performance of collective communication and fine-grained message passing needs further improvement.