ZURUECK HOCH VOR INHALT SUCHEN

» Back to overview
Proposing Institution

Institut für Informatik, Lehrstuhl für Rechnertechnik und Rechnerorganisation ,TUM Garching
Project Manager

Dai Yang
Boltzmannstr. 3
85748 Garching
Abstract
On the emerging way toward exascale, the current MPI profiling interface (PMPI) may not be suitable anymore. It appears to be very rigid. The current PMPI Interface has a weak symbol intersection, which dependent on the OS. This means that the support is only existent for a single tool to be intercepted into the MPI execution. Also, it limits the tool to be monolithic. In order to add support for whole toolchain, changes have to be made on the current PMPI interface. In addition, with our new design, we will be able to do link time or runtime enablement, increasing flexibility for future MPI Profiling. In our proposal to the MPI Forum, we set up a requirement which preserves existing PMPI interface. Also, we plan to add support for pre and post call activation. With our new design, it can be ensured that future QMPI interface can be used for a arbitrary number of MPI Profiling tools running in a single MPI Process. It is planned that our prototype implementation and performance evaluation done on the SuperMUC is to be presented to the MPI Forum; In addition, it is very likely that our results is going to be embed into OpenMPI in the next 2 years.

Impressum, Conny Wendler