As currently stated in the proposal, the info argument is opaque to MPI
and is directly passed to the runtime environment, which uses it to
determine where and how to spawn processes. If the info argument is
empty, the runtime environment decides where to spawn processes. Portable
programs not requiring detailed control over process locations should use
an empty argument.
MPI does not specify the format of the info argument, which is left
up to the MPI implementation. The format is expected to be tailored
to the particular runtime system being used.
Because the format can be tailored to a particular system, and
is left up to the implementation, the info argument should be
general enough to allow programs to pass information to runtime
environments with different levels of sophistication.
I would like to propose that the info argument to MPI_SPAWN is a handle.
The following functions would be added to create/free the handle:
MPI_INFO_CREATE(info, handle)
IN info a string providing information to
the runtime system on how/where to
start processes
OUT handle MPI_INFO_HANDLE
MPI_INFO_ADD(old_handle, info, new_handle)
IN old_handle MPI_INFO_HANDLE
IN info string providing additional information on
how/where to start processes
OUT new_handle MPI_INFO_HANDLE
MPI_INFO_FREE(handle)
IN handle MPI_INFO_HANDLE
The following is a variable defined for both C and Fortran to be used
when a program passes info as an empty argument:
MPI_INFO_NULL
Discussion
----------
The original proposal defines info as a string. This proposal has
fewer functions, but is less flexible. Another drawback with having
info as a string, is that info is limited on some systems to 256 characters
in Fortran, and the possibility of users specifying a string with more
that 256 characters is quite feasible.
The new proposal allows more flexibility. Users can still pass a string,
and can add to it, solving the 256 character string length problem.
It also allows more complex information in the form of structures
for example, to be passed to the runtime environment. The new proposal
also defines MPI_INFO_NULL so that users can easily pass an empty
argument from either Fortran or C.
Rolf suggested to me that the MPI_INFO_ADD() and MPI_SPAWN functions free
the handle instead of having a separate function. This would simplify
the interface for a single spawn call, but would complicate it if a user
wanted to set up the info argument, then do multiple spawns, reusing
the info argument. Since MPI_SPAWN would free the handle, the user would
have to make another MPI_INFO_CREATE call ahead of time to duplicate the
handle.
Comments? %:)
Heidi Poxon E-Mail: hlg@cray.com
Cray Research, Inc. Phone: 612/683-5677
655F Lone Oak Drive
Eagan, MN 55121