We report on a collaborative project to investigate the input/output (I/O) performance of the AIPS++ system with particular emphasis on the the I/O characteristics of our parallel, MPI-based application pimager. The AIPS++ parallel group has been working extensively with the Pablo group of the University of Illinois Computer Science department. The Pablo group has developed an I/O and event tracing library that records the counts and timing of I/O calls in an application program, as well as tools to analyze the output trace records. A graduate student from the Pablo project, working closely with the AIPS++ parallel group, has included the I/O instrumentation calls into the AIPS++ code to allow I/O and event tracing during the execution of any AIPS++ application. The AIPS++ parallel and Pablo groups have carried out several joint experiments using the parallel version of the imager application ( pimager). Pimager does parallel deconvolution of a spectral-line data set using the Clark CLEAN algorithm; parallelism is implemented with the Message Passing Interface (MPI). The experiments were run on a 128-processor SGI Origin 2000 at NCSA. The I/O characteristics of the pimager program from these experiments are presented.
|Title of host publication
|Astronomical Data Analysis Software and Systems IX
|N. Manset, C. Veillet, D. Crabtree
|Astronomical Society of the Pacific
|Published - 2000
|Astronomical Society of the Pacific Conference Series