GitLab

The old one was stale and yielded a 404.

No reviewer.

This was the only remaining non-SMP algorithm to port from MPI_Reduce.

Reviewed by buntinas@.

This fixes MPIX_Ialltoallw as used in the coll/icalltoallw test.

No reviewer.

No reviewer.

The bug fix was a devilish stack corruption bug that was caused by using
a stack variable as the broadcast buffer.  After MPIR_Ibarrier_inter
returned, the memory for the variable was reused for other purposes,
resulting in erratic corruption and segfaults in the progress engine.
Mallocing the buffer instead resolves this problem.

The modernization includes switching to the MPID_SCHED_BARRIER macro and
dispatching sub-collectives via the coll_fns table instead of statically.

No reviewer.

Also, dispatch to the localcomm's coll_fns table instead of statically
selecting an intracomm bcast routine.

No reviewer.

No reviewer.

The code was incorrectly calculating the blklens array entries, leading
to negative values that caused strange assertions in Nemesis to trigger.

No reviewer.

Users will never encounter this directly because MPI_Type_indexed
and friends perform checking, but MPIR_Type_indexed_impl and friends
do not.  Without this check, much more confusing assertions will trigger
later deep in nemesis fastbox code when negative blocklens are provided.

No reviewer.

No reviewer.

The code did not originally track per-entry communicators, so it was
correct in the past to use a single comm stored in the schedule.

This fixes most intercommunicator collectives, which were broken because
intercomm collectives commonly use multiple comms (the real comm and the
local comm).  SMP collectives similarly were broken.

Reviewed by buntinas@.

I don't think that this has any ABI implications, since we don't
actually have "external32" support and it doesn't change any sizes for
any of the actual routines or symbols that the library exposes to
applications.

No reviewer.

[svn-r8391] Fix missing changes to AR_FLAGS in simplement (a previous update failed to update all of them

target in src/env already does this at install time.

Reviewed by goodell.

comment describing it.

No reviewer.

autoconf-2.62.

Reviewed by goodell.

in r8378.

No reviewer.

mpirun. Each process manager should automatically take care of what
all executables it exposes to the user. For example, some process
managers do not even call their executable "mpiexec" (MVAPICH's
process manager).

Reviewed by goodell.

at build time, and then separately install them into the install
directory at install time. Also get rid of the mpich2-build-install
targets for all process managers. mpiexec and friends are no longer
installed at build time.

Reviewed by goodell.

provides this functionality for the user. Don't install it.

Reviewed by buntinas.

of operations during parallel builds.

deprecated.

The "type contents" is used to calculate the number of basic elements,
and for some reason the pointers to the type contents were recalculated
at the upper level instead of calling MPID_Type_access_contents.  This
duplicated logic didn't account for padding between the arrays in some
cases, which led to calculations based off of undefined data.  However
only fairly complicated types ever triggered this bug because it requires
both more than one basic element type and a type constructor with an odd
number of ints in the contents.

This commit includes a new "datatype/get-elements" regression test.

Thanks to Fab Tillier @ Microsoft for reporting the bug.

Reviewed by buntinas@.

This fixes the improved version of the comm/iccreate test that was added
to the trunk in r8343.

Reviewed by buntinas@.

[svn-r8368] Fix #1465, problem building rlog (used to gather detailed performance data about the implmentation)

locality information in the process-mapping key, so MPICH2 does not
try to allocate shared memory between processes.

than the proxy-specific node list. Also generalize a function that can
be used in other places in Hydra as well.

Unfortunately, MPICH2 currently fails this test.

No reviewer.

No reviewer.

No reviewer.

No reviewer.

MPICH_NO_LOCAL was already handled by the MPIR_Param system, but this
parameter was not for some reason.

No reviewer.

This one is all me:  under certain workloads, ADIOI_R_Exchange_data_alltoallv
will allocate too-big temporary buffers (on some nodes), causing an out-of
memory assertion and general applicaiton sadness.  The Kernel_GetMemorySize
funciton call is indeed BlueGene/P specific, but the ALCF folks tell us that
when a malloc fails it leaves some lint in the OS state.