Seminar Announcement
Rolf RabenseifnerUniversity of Stuttgart"Parallel Communication & MPI-I/O Benchmarking and Automatic MPI Profiling"
Date: Friday, May 18, 2001
Abstract:The talk will start with an overview on German HPC centers and on the High Performance Computing Center Stuttgart (HLRS) that has systems totaling more than 700 GFlops (Rmax), including new vector machines from NEC and Hitachi. The second part of the talk describes our automatic counter profiling of all MPI applications on our Cray T3E-900/512. A detailed summary of the hardware performance counters and the MPI calls of any MPI production program is gathered during execution and written on a special syslog file. User specific summaries are sent by e-mail to each user each week. Summary results of profiling our users for six months are presented. The third part of the talk describes two MPI benchmarks created to characterize the balanced system performance of high performance clusters and supercomputers: b_eff, the communication specific benchmark examines the parallel message passing performance of a system, and b_eff_io, which characterizes the effective I/O bandwidth. Both benchmarks have two goals: a) to get a detailed insight into the performance of different parallel communication and I/O patterns, and b) to obtain a single average bandwidth number. Results are given for IBM SPs, Cray T3E, NEC SX-5, and Hitachi SR 8000.
Email: rabenseifner@hlrs.de Speaker's web page: http://www.hlrs.de/people/rabenseifner
|