[svn-r7720] Fix collectives to not hang if the communicator contains a failed process.  The collectives will not return an error immediately upon detecting a failure, rather they'll return the error at the end of the function and continue the communication pattern so that other processes waiting to receive messages will not hang.  This means that, although the collective should complete at all processes, some processes will receive an error, and some processes may not get a valid result.  Since some processes may not receive an error and still receive an invalid result, a separate mechanism is needed to confirm that the collective has completed correctly, such as MPI_Comm_validate of the MPI3 FT proposal.

at all. We just change to the base directory and run the executable.

processes. If we forked out the MPI processes (PMI_FD format), we
preload the PMI FDs as well. If we are using a PMI PORT, we wait for
the application to send a PMI initialization message before loading
the PMI FD.

[svn-r7705] Updated the test on the nameserver choice to be compatible with the new default; remove reference to long-deleted ldap nameserver

[svn-r7702] Make use of the atomic increment of the completion_count to bump the progress engine when a checkpoint is initiated of a failed process is detected.  Also jump out of blocking_recv if the completion_count is bumped.

yet, we still may have posted recvs; we need to look for those in a
failure case.  Similarly, we may have sends in the send queue even if
the connection hasn't been fully established.

library.  if for some reason we need the MPI implementation, use
--with-mpi-impl (but it looks like that historical flag doesn't actually do
anything any longer

used is to chdir to the directory where the executable is located and
launch the program from there. Fixes ticket #1160.

we are still not handling the case where a process running in the
background might look for STDIN (in which case it should get an
EOF). That's a random corner case, which we don't support right now.

socket. This makes STDIN management independent of which bootstrap
server we use. This provides the necessary setup to enable ticket

proxies. But even if that's not successful, mpiexec should still exit.

-outfile-pattern and -errfile-pattern.

This fixes ticket #1122.  This fix hasn't been comprehensively tested,
but it has worked fine everywhere that I've tried it.

No reviewer.

[svn-r7687] Changed global completion count to use atomics rather than locks.  This will allow us to update the completion count asynchronously from other threads or interrupt contexts

Prevents duplicate storage on platforms like darwin that don't support
weak symbols and therefore have two libraries in order to support the
PMPI profiling interface.  Previously, there was also a copy in
libpmpich.a, which contains the actual "MPI_" versions of the routines.

No reviewer.

In the old code it was possible that ROMIO freed the keyval multiple
times, rather than just once.  This was fine in MPICH2, which is robust
in the face of such behavior, but caused problems for ROMIO over Open
MPI.  We now utilize the MPI-2.2 LIFO ordering for attribute destruction
in order to free the keyval exactly once.

Reported by Pascal Deveze and related to ticket #222.

Reviewed by robl@.

This is more natural to users because it behaves just like printf escape
specifiers.

Reviewed by balaji@.

This was a feature request from Quincey in order to improve the
effectiveness of the HDF5 test suite.

No reviewer.

This builds on r7653 and r7654.

No reviewer.

[svn-r7671] merged in error-return branch.  This branch completes with an error recvs on failed VCs or anysource recvs on comms with failed VCs.  See log for r7669 for details.

No reviewer.

Bill was sharp and caught my mistake, which didn't entirely replace
sizeof(void*) with align_sz.

No reviewer.

Reviewed by buntinas@.

Thanks to Cray for reporting these issues.

No reviewer.

 #1131.

expression to be prepended to the output. -prepend-rank is a subset of
this generalized capability.

1. Move UI specific parameters to a different structure, so other
components don't see them.

2. Rename HYD_handle to HYD_server_info to make the data in it
clearer.

3. Rename hydra.h to hydra_server.h (this is only used by the server
side). Also make sure the proxy side code does not include it.

4. Merge hydra_base.h and hydra_utils.h and make it hydra.h, as these
are the utility functions used for both the servers and the proxies.

1. Point out that C++, F77 and F90 support is optional and the users
can disable them.

2. Remove the recommendation for VPATH builds. While this is a
convenient option for system administrators or developers who maintain
multiple copies of MPICH2 builds, it is confusing for users.

3. Remove redundant description of configure options.

4. Remove the MPICH_INTERFACE_HOSTNAME environment variable from the
README, as in the default PM, we automatically pass it based on the
host list provided by the user.

5. Cleaned up several descriptions.

ticket #1141.

used at several places now.