Designing an Efficient Kernel-Level and User-Level Hybrid Approach for MPI Intra-Node Communication on Multi-Core Systems.