Description of problem fixed
-----------------------------

This efix targets the following issues:

1 - Intermittent data corruption using mpi_get with derived mpi types whereby an optimization in pami
    which split messages which were on a line in E dimension and over 2k in size was disabled and the 
    data corruption disappeared.  After consultation with IBM research it was deemed this optmization
    has very marginal impact on performance which was further verified by the user.

2 - Allgather optimization using allreduce had in issue if number of bytes were not div by 4. The fix 
    avoids the optimized allgather over allreduce for messages that are not div by 4 (MPICH fix).

3 - An optimized allreduce was using in non-comm-world communicators in multithread MPI mode. This
    optimized collective tree allreduce doesn't support multithreaded mode for non-comm-world. The
    fix is to use another algorithm for allreduce when this is the case.

4 - A wrong injChannel was advanced for some collectives. This causes a hang in certain situations
    when MU doesn't complete immediately causing all pt2pt and collectives communication to block.


Steps to take before installing the RPMs for this fix
-----------------------------------------------------

NONE

Steps to take after installing the RPMs for this fix
----------------------------------------------------

NONE

Getting the list of installed fixes
-----------------------------------

To find out which efixes are currently installed and not superceded, run the following command:

rpm -qf `/bin/ls -d /bgsys/drivers/V?R?M? | tail -1` /opt/teal/ibm --queryformat "%{release}\n" | sort -nu