An MPI messaging layer for network processors

• Main Author: Wijeyeratnam, Ashley
• Language: English
• Published: 2009
• Online Access: http://hdl.handle.net/2429/9809

We describe the design and implementation of MPI-NP, a Myrinet communication system tailored to support LAM, a public domain version of MPI. The goals of MPI-NP are to reduce the time spent on the host for message processing, and to promote the overlap of computation and communication. MPI-NP achieves these goals by off-loading as much of the computation as possible to the network processor. MPI-NP relieves the host of several tasks, making more CPU cycles available to the application, but pays a price for heavy utilization of the slower network processor, by adding a significant overhead to message latency. Although part of the overhead can be attributed to the hardware of our testbed, the philosophy of MPI-NP characteristically does not provide the best latency possible because of performance disparities between host and network processors. Users are left with the choice of deciding on the trade-off of latency versus host overhead. Applications which are not latency bound can expect to perform well on MPI-NP.