linked into libmpi
* add --enable-mca-direct to specify components that should be directly
called (instead of going through component structs and the like). The
components and component frameworks must explicitly support this.
Currently, only the TEG PML does so.
* Updated all the calls to the PML to use a macro so that they can either
be direct called or called through function pointer interfaces (aka
the component infrastructure)
This commit was SVN r5291.
Windows.
Added some functionality to replicate the waitpid and kill
under Windows.
Also created a ompi_environ.* file which contains the same
funcationality as environ.*, but we can't use that filename
since it is already defined under Windows.
This commit was SVN r5289.
Fix a bug (the fragment came from a cache so we should reinitialize the ptl_peer before calling the match or
otherwise the match does not set the peer).
Add more user friendly log messages.
This commit was SVN r5288.
compiler couldn't know that ompi_have_threads() would always be false,
so it was still running the tests. This saves about .04 microseconds
on the critical path, according to Tim.
Reviewed by George
This commit was SVN r5275.
- only show the basename of the argv[0] when displaying error messages
- if orterun is run with no arguments, show the help message
This commit was SVN r5269.
dimensions the f90 MPI bindings should support (they are strongly
typed, and the number of dimensions of choice arguments must be
specified -- it cannot be arbitrary). The default is four.
Note that even though increasing this value has essentially a linear
effect on the code, the multiplier constant is fairly large (only a
small number of functions have 2 choice buffers, so the exponential
factor is relatively small). Increasing this value increases the
amount of time f90 compilers will spend compiling src/mpi/f90/mpi.f90
(some compilers will crash if it is too big).
This commit was SVN r5268.
- Now we support GM 1.x
- We detect correctly the version of GM and activate some defines
- memory leaks fixes (dont release a fragment several times and release all fragments)
- protect the free_list while debug is activated
- change the place where we keep the informations about the local node as well as for the peers.
It should be a little bit cleaner as we have less copies.
- first step toward supporting OSes without memory registration.
- correctly select the communication protocol depending on the capabilities of the GM library.
There is still some work to be done as the detection is ... let's say user based at this moment (using
MCA parameters).
This commit was SVN r5248.
