| In this thesis, we examine the MPI characteristics of the three applications in the NPB-MZ suite as well as two applications ( SPECseis and SPECenv) in the SPEChpc2002 suite in terms of their point-to-point and collective communications. We also evaluate the impact of different number of processors as well as different problem sizes on the communication characteristics of these applications. Overall, our experiments reveal that the applications studied have diverse communication patterns, and that they are sensitive to the changes in the system size and the problem size.; This thesis presents an in-depth evaluation of the new Myrinet two-port networks at the user-level (GM), MPI-level, and at the Aggregate Remote Memory Copy Interface (ARMCI) level. High-performance interconnects such as Myrinet provide a one-sided communication model, referred to as Remote Direct Memory Access (RDMA), which is not utilized in many parallel applications, such as NPB-MZ. We realized that non-blocking operations perform better than blocking, and two-port communication outperforms one-port communication. We noticed that for messages larger than 8KB, ARMCI non-blocking Put has a better performance than MPI Send/Receive operations. (Abstract shortened by UMI.)... |
