Talks and Poster Presentations (with Proceedings-Entry):
"High Performance Expectations for MPI";
Keynote Lecture: Austrian HPC Meeting 2017 - AHPC 2017,
Grundlsee, Austria (invited);
- 2017-03-03; in: "AHPC 2017, Austrian HPC Meeting 2017",
G. Baumgartner, J. Courian (ed.);
FSP Scientific Computing, University of Innsbruck,
A so-called high-quality implementation of MPI  first and foremost ensures best possible performance on the systems used by the application programmer. Second, this high-quality MPI implementation should also behave as expected. It is neither clear what exactly "best possible performance" means, nor what application programmers actually expect and could sensibly expect regarding MPI performance.
Performance guidelines  is an attempt to formalize in a verifiable way reasonable expectations on MPI performance without relying on abstract models of system and algorithm performance. Verified guidelines guard again unpleasant performance surprises in MPI applications, and ensure a degree of performance portability between systems and libraries. Violated guidelines alert the user to problematic features of her MPI library, and show the library implementer where improvements are needed. We give concrete examples
of performance guidelines for collective operations  and derived datatypes  for ensuring high quality of MPI library implementations.
Created from the Publication Database of the Vienna University of Technology.