MPI-StarT: Delivering Network Performance to Numerical Applications

We describe an MPI implementation for a cluster of SMPs interconnected by a high- performance interconnect. This work is a collaboration between a numerical applications programmer and a cluster interconnect architect. The collaboration started with the modest goal of satisfying the communication needs of a specific numerical application, MITMatlab. However, by supporting the MPI standard MPI-StarT readily extends support to a host of applications. MPI-StarT is derived from MPICH by developing a custom implementation of the Channel Interface. Some changes in MPICH´s ADI and Protocol Layers are also necessary for correct and optimal operation. MPI-StarT relies on the host SMPs´ shared memory mechanism for intra-SMP communication. Inter-SMP communication is supported through StarT-X. The StarT-X NIU allows a cluster of PCI-equipped host platforms to communicate over the Arctic Switch Fabric. Currently, StarT-X is utilized by a cluster of SUN E5000 SMPs as well as a cluster of Intel Pentium-II workstations. On a SUN E5000 with StarT-X, a processor can send and receive a 64-byte message in less than 0.4 and 3.5 usec respectively and incur less than 5.6 usec user-to-user one-way latency. StarT-X´s remote memory-to-memory DMA mechanism can transfer large data blocks at 60 MByte/sec between SUN E5000s. This paper outlines our effort to preserve and deliver this level of communication performance through MPI-StarT to user applications. We have studied the requirements of MITMatlab and the capabilities of StarT-X and have formulated an implementation strategy for the Channel Interface. In this paper, we discuss some performance and correctness issues and their resolutions in MPI-StarT. The correctness issues range from the handling of arbitrarily large message sizes to deadlock-free support of nonblocking MPI operations. Performance optimizations include a shared-memory-based transport mechanism for intra- SMP communication and a broadcast mechanism that is aware of the performance difference between intra-SMP and the slower inter-SMP communication. We characterize the performance of MPI-StarT on a cluster of SUN E5000s. On SUN E5000s, MPI processes within the same SMP can communicate at over 150 MByte/sec using shared memory. When communicating between SMPs over S- tarT-X, MPI-StarT has a peak bandwidth of 56 MByte/sec. While fine-tuning of MPI-StarT is ongoing, we demonstrate that MPI-StarT is effective in enabling the speedup of MITMatlab on a cluster of SMPs by reporting on the performance of some representative numerical operations.