D. Turner, S. Selvarajan, X. Chen, and W. Chen (USA)
MP_Lite, MPICH, MPI, VIA
The inter-processor communication rate is the major limiting factor to efficiently using all types of parallel computers. The message-passing system is often part of the problem, restricting the bandwidth that the underlying network offers. MP_Lite is a lightweight message passing library designed to deliver the maximum performance to applications. It supports the most common MPI functions, including blocking and asynchronous communications and the basic global operations. It is being used to investigate new methods for improving message-passing performance. A SIGIO interrupt driven approach allows MP_Lite to deliver nearly all the performance that TCP offers while maintaining message progress at all times. The VIA module bypasses the OS to provide a more streamlined path from applications to the network. A shared-memory segment module provides faster communications within SMP nodes. An MPICH-MP_Lite implementation has been developed to pass these benefits to a full MPI library. These new approaches provide better performance to message-passing applications in a broad range of high-performance computing environments.
Important Links:
Go Back
