Performance of the MP Lite Message-Passing Library on Linux Clusters

摘要

MP Lite is a light-weight message-passing library designed to deliver the maximum performance to applications in a portable and user-friendly manner. It supports a subset of the most commonly used MPI commands, or it can be run using its own simpler syntax. The high level of performance is achieved by implementing a core set of communication primitives in a very clean fashion. MP-Lite resorts to extra buffering only when necessary. Its small size also makes it an ideal tool for performing message-passing research. On Unix clusters, MP-lite can run as a thin layer on top of TCP, providing all the performance that TCP offers for Applications that can take advantage of the restricted mode. The Default signal based interrupt dirven mode is more robust and still provides around 90% of the TCP performance on Gigabit Ethernet networks. Current research involves the Development of an MP-Lite module for the M-VIA OS-bypass library. Bypassing the OS reduces the message latency and streamlines the flow of data to the network hardware. Reducing the memory-to-memory copies puts much less strain on te memory bus, allowing more of the full bandwidth of the PCI bus to be realized when coupled with Gigabit-level networking. MP lite also has the capacility of striping data across multiple network interfaces per computer. If done at the TCP level, dual Fast Ethernet interfaces double the performance for PC Clusters at a Minimal increase to the overall cost. Current work with the M-VIA module allows effective channel bounding of at least three Fast Ethernet interfaces per maching for PC cluster, providing a scalable networking solution for low cost clusters.