On the memory footprint issue, I'd like to add that it is possible to structure an MPI implementation to only require the code associated with the routines that are used in the application, even when the application must be statically linked (as is the case for Blue Gene/ L, for example). This can be done even if there are actions that the application needs or wants to perform at the time of MPI_Finalize. There are some things that can be done to make this easier for the implementation, but (like Dick) I'm not convinced that a subset is required as a way to limit the memory footprint. Bill William Gropp Paul and Cynthia Saylor Professor of Computer Science University of Illinois Urbana-Champaign *