Underwood, Keith D
keith.d.underwood at intel.com
Sun Jan 24 17:19:31 CST 2010
> Still seems a little odd to have a solution for I/O functions and not
> for communication functions (the original ticket was for
> communication https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/117).
> If > 2GB can be written from a single process in a single
> call, they can also be communicated. If it is tedious to break up the
> write into smaller writes, the same applies to communication.
"can be" and "will be" are very different phrases. The feeling within the group that was there Thursday morning was that the "way 64 bit is done" (TM) is to replicate all of the function calls. The group felt like a compelling case had been made for I/O, because you would want to write all of your memory to file (e.g. checkpointing) and memory per node is growing beyond 2 GB. However, it was much harder to make the case that you really wanted to send that much data from a contiguous location to any other node. In fact, someone made the statement "all of the application users understand that you can't do that and still scale on any network in the foreseeable future".
Your interoperability comments may be a more motivating example, but as of Thursday morning, it was clear that there was not consensus among the forum members who were - especially given the pain that is going to be involved in fixing this for the general case.
More information about the mpi-forum