- move the attribute init section in ompi_mpi_init() down below where
communicators are setup (we need MPI_COMM_WORLD to be setup before
attributes and keyvals are setup)
- removed confusing extra wrapper class around ompi_hash_table_t;
looks like it was a victim of slow eroding of members so I put took
it out back and put it out of its misery
- added preliminary definitions for all the pre-defined keyvals.
Still need more work here to assign their initial values, but I
think Edgar was running into an atrribute problem and it may have
been that the pre-defined attrs didn't yet exist.
- removed some LAM-specific predefined keyvals from mpi.h
This commit was SVN r2695.
invoking an errhandler in MPI_INIT when we don't yet have a
communicator to invoke them on. Instead, remove some of the logic
from the ERRHANDLER macros and just invoke the back-end function
directly in MPI_INIT.
This commit was SVN r2692.
configure so that we can be smarter about selection logic (note: we are
not yet smarter about selection logic - coming soon). To keep things
sane, moved a bunch of files into rms/src/ from rms/.
* Since we were already forcing an autogen, move all the files from
rsh/src/ to rsh/ for the rsh pcm component. Makes things a bit quicker
on the compile, with no real downside
NOTE: You will have to run autogen / configure after this update.
This commit was SVN r2683.
connection management, the first *REAL* infiniband send has
to be triggered from the progress engine. Taking care of this
case with this code.
This commit was SVN r2674.
such as MPI_DUP_FN: MPI says that they have to be available via that
name in both C and Fortran. However, for implementation reasons, we
have to have them as separate functions. But if a Fortran compiler's
naming convention is ALL_CAPS, this is not possible. Hence, we need
to put in #define's in mpi.h to change the C names something like
this: MPI_DUP_FN -> OMPI_C_MPI_DUP_FN. This was actually done a long
time ago.
However, the source code where those C functions actually live
(src/mpi/c/attr_fn.c) still reflected the old names (e.g.,
MPI_DUP_FN). This is fine, actually -- mpi.h would come in and
#define them to their real names (OMPI_C_MPI_DUP_FN). So it was
functionally correct, but confusing at hell (and it just bit Edgar and
me today). So I'm changing src/mpi/c/attr_fn.c to use the real names
(OMPI_C_MPI_DUP_FN) so that tools like ctags and grep can find them in
the source code when you go looking. Plus, the code is just more
clear that way. I also put in massive comments about this in
src/mpi/c/attr_fn.c and src/mpi/f77/attr_fn_f.c so that we remember
why the heck we did this. :-)
This commit was SVN r2670.
- stop mca_oob_parse_contact_info from trashing the input string
- get gpr_proxy to reissue the non-blocking receive for notify messages
- add a diagnostic message to the gpr_proxy
This commit was SVN r2669.
- The official OSX workaround for ROMIO is to configure it with the
-arch=freebsd switch (that's what MPICH does!). This switch no
longer works because I updated their configure script to a modern
version of Autoconf, so I put a[nother] test in ROMIO's configure to
detect this situation and do the workaround.
- Realize that I forgot that ROMIO uses the MPI profiling layer, so if
the profiling layer is not enabled, ROMIO will fail to link in
dynamic environments (or otherwise Not Work). So put a thingy in
its configure script testing to see if the MPI profiling layer is
enabled or not. If it's not enabled, don't build ROMIO.
- Make the lack of any available io components *NOT* be an error
during MPI_INIT (e.g., what if you didn't enable profiling, but you
don't use any MPI_FILE* functions in your program? This is quite a
common case). If no io components are available, then an
MPI exception will be thrown during MPI_FILE_OPEN and
MPI_FILE_DELETE (which, unfortunately, according to the standard
uses the default handler on MPI_FILE_NULL, which, unfortunately
again, is MPI_ERRORS_RETURN). So users really have to check for
this situation. :-\
This commit was SVN r2661.
Fixed ompid so it reissues the non-blocking receive - should now be close to ready for primetime. Fixed some logic in the svc framework that wasn't checking properly for action flags.
This commit was SVN r2660.
