mpich.git
8 years ago[svn-r8400] intercomm Ibarrier modernization and bugfix
David Goodell [Mon, 11 Apr 2011 22:47:01 +0000]
[svn-r8400] intercomm Ibarrier modernization and bugfix

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.

8 years ago[svn-r8399] bugfix: intercomm bcast would hang forever
David Goodell [Mon, 11 Apr 2011 22:47:00 +0000]
[svn-r8399] bugfix: intercomm bcast would hang forever

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

No reviewer.

8 years ago[svn-r8398] add missing schedule barriers in intercomm allgather
David Goodell [Mon, 11 Apr 2011 22:46:59 +0000]
[svn-r8398] add missing schedule barriers in intercomm allgather

No reviewer.

8 years ago[svn-r8397] bugfix: MPIR_Reduce_scatter_block_rec_dbl internal blocklens calculations
David Goodell [Mon, 11 Apr 2011 22:46:57 +0000]
[svn-r8397] bugfix: MPIR_Reduce_scatter_block_rec_dbl internal blocklens calculations

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

No reviewer.

8 years ago[svn-r8396] add blocklens sanity check to MPID_Type_indexed
David Goodell [Mon, 11 Apr 2011 22:46:56 +0000]
[svn-r8396] add blocklens sanity check to MPID_Type_indexed

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.

8 years ago[svn-r8395] bugfix: remove leftover placeholder assertion in MPIR_Igather_inter
David Goodell [Mon, 11 Apr 2011 22:46:55 +0000]
[svn-r8395] bugfix: remove leftover placeholder assertion in MPIR_Igather_inter

No reviewer.

8 years ago[svn-r8394] bugfix: use the correct comm in MPIDU_Sched_start_entry
David Goodell [Mon, 11 Apr 2011 22:46:54 +0000]
[svn-r8394] bugfix: use the correct comm in MPIDU_Sched_start_entry

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@.

8 years ago[svn-r8393] correct MPI_C_BOOL size in "external32" (should be 1, was 4)
David Goodell [Mon, 11 Apr 2011 18:24:35 +0000]
[svn-r8393] correct MPI_C_BOOL size in "external32" (should be 1, was 4)

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.

8 years ago[svn-r8391] Fix missing changes to AR_FLAGS in simplement (a previous update failed...
William Gropp [Mon, 11 Apr 2011 12:36:17 +0000]
[svn-r8391] Fix missing changes to AR_FLAGS in simplement (a previous update failed to update all of them

8 years ago[svn-r8388] Don't create a link from mpicxx to mpic++ at build time. The install
Pavan Balaji [Sun, 10 Apr 2011 02:04:17 +0000]
[svn-r8388] Don't create a link from mpicxx to mpic++ at build time. The install
target in src/env already does this at install time.

Reviewed by goodell.

8 years ago[svn-r8387] Minor cleanup to the install-alt target for Hydra and improve the
Pavan Balaji [Sun, 10 Apr 2011 02:04:07 +0000]
[svn-r8387] Minor cleanup to the install-alt target for Hydra and improve the
comment describing it.

No reviewer.

8 years ago[svn-r8386] Improve comment which seemed to imply that this was fixed in
Pavan Balaji [Sun, 10 Apr 2011 02:03:58 +0000]
[svn-r8386] Improve comment which seemed to imply that this was fixed in
autoconf-2.62.

Reviewed by goodell.

8 years ago[svn-r8385] parkill is a developer tool. Missed out removing some instances of it
Pavan Balaji [Sun, 10 Apr 2011 02:03:46 +0000]
[svn-r8385] parkill is a developer tool. Missed out removing some instances of it
in r8378.

No reviewer.

8 years ago[svn-r8384] Removed a hack in configure to create a symbolic link from mpiexec to
Pavan Balaji [Sun, 10 Apr 2011 02:03:36 +0000]
[svn-r8384] Removed a hack in configure to create a symbolic link from mpiexec to
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.

8 years ago[svn-r8383] There is no need to install mpiexec and friends in the build directory
Pavan Balaji [Sun, 10 Apr 2011 02:03:30 +0000]
[svn-r8383] There is no need to install mpiexec and friends in the build directory
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.

8 years ago[svn-r8380] Fix duplicate comment.
Pavan Balaji [Fri, 8 Apr 2011 21:12:06 +0000]
[svn-r8380] Fix duplicate comment.

8 years ago[svn-r8379] PLPA is deprecated; yay
Pavan Balaji [Fri, 8 Apr 2011 21:12:04 +0000]
[svn-r8379] PLPA is deprecated; yay

8 years ago[svn-r8378] Parkill is a developer tool (not meant for users); the process manager
Pavan Balaji [Fri, 8 Apr 2011 21:12:02 +0000]
[svn-r8378] Parkill is a developer tool (not meant for users); the process manager
provides this functionality for the user. Don't install it.

Reviewed by buntinas.

8 years ago[svn-r8377] Use install-exec-hook instead of install-exec-local to ensure ordering
Pavan Balaji [Fri, 8 Apr 2011 21:12:00 +0000]
[svn-r8377] Use install-exec-hook instead of install-exec-local to ensure ordering
of operations during parallel builds.

8 years ago[svn-r8376] Remove information about MPD from the user and install guides. MPD is
Pavan Balaji [Fri, 8 Apr 2011 21:11:58 +0000]
[svn-r8376] Remove information about MPD from the user and install guides. MPD is
deprecated.

8 years ago[svn-r8374] bugfix: MPI_Get_elements with exotic types (tt#1467)
David Goodell [Fri, 8 Apr 2011 19:32:28 +0000]
[svn-r8374] bugfix: MPI_Get_elements with exotic types (tt#1467)

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@.

8 years ago[svn-r8370] fix intercomm MPI_Comm_create with an empty group on one side
David Goodell [Thu, 7 Apr 2011 20:19:11 +0000]
[svn-r8370] fix intercomm MPI_Comm_create with an empty group on one side

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

Reviewed by buntinas@.

8 years ago[svn-r8368] Fix #1465, problem building rlog (used to gather detailed performance...
William Gropp [Thu, 7 Apr 2011 19:40:28 +0000]
[svn-r8368] Fix #1465, problem building rlog (used to gather detailed performance data about the implmentation)

8 years ago[svn-r8366] Fix missing entries in many svn:ignore property lists
William Gropp [Thu, 7 Apr 2011 18:18:40 +0000]
[svn-r8366] Fix missing entries in many svn:ignore property lists

8 years ago[svn-r8358] Update the CHANGES file to include information on the hwloc upgrade.
Pavan Balaji [Thu, 7 Apr 2011 00:46:54 +0000]
[svn-r8358] Update the CHANGES file to include information on the hwloc upgrade.

8 years ago[svn-r8355] This is a hack-fix for ticket \#1434. We currently just disable all
Pavan Balaji [Thu, 7 Apr 2011 00:10:47 +0000]
[svn-r8355] This is a hack-fix for ticket \#1434. We currently just disable all
locality information in the process-mapping key, so MPICH2 does not
try to allocate shared memory between processes.

8 years ago[svn-r8354] PMI process mapping should directly use the global node list, rather
Pavan Balaji [Thu, 7 Apr 2011 00:10:41 +0000]
[svn-r8354] PMI process mapping should directly use the global node list, rather
than the proxy-specific node list. Also generalize a function that can
be used in other places in Hydra as well.

8 years ago[svn-r8353] Updated hwloc to 1.1.2.
Pavan Balaji [Thu, 7 Apr 2011 00:10:28 +0000]
[svn-r8353] Updated hwloc to 1.1.2.

8 years ago[svn-r8352] Merge the updated CHANGES file from the 1.4.x branch.
Pavan Balaji [Thu, 7 Apr 2011 00:09:39 +0000]
[svn-r8352] Merge the updated CHANGES file from the 1.4.x branch.

8 years ago[svn-r8343] update iccreate test to check that MPI_GROUP_EMPTY yields MPI_COMM_NULL
David Goodell [Wed, 6 Apr 2011 22:51:19 +0000]
[svn-r8343] update iccreate test to check that MPI_GROUP_EMPTY yields MPI_COMM_NULL

Unfortunately, MPICH2 currently fails this test.

No reviewer.

8 years ago[svn-r8342] delete incorrect comment in grouptest.c
David Goodell [Wed, 6 Apr 2011 22:51:18 +0000]
[svn-r8342] delete incorrect comment in grouptest.c

No reviewer.

8 years ago[svn-r8341] convert MPIR_Group_translate_ranks_impl to use mpi_errno errhandling
David Goodell [Wed, 6 Apr 2011 22:51:15 +0000]
[svn-r8341] convert MPIR_Group_translate_ranks_impl to use mpi_errno errhandling

No reviewer.

8 years ago[svn-r8340] convert MPIR_Group_compare_impl to use mpi_errno errhandling
David Goodell [Wed, 6 Apr 2011 22:51:13 +0000]
[svn-r8340] convert MPIR_Group_compare_impl to use mpi_errno errhandling

No reviewer.

8 years ago[svn-r8339] add a param for MPICH_ODD_EVEN_CLIQUES
David Goodell [Wed, 6 Apr 2011 22:51:11 +0000]
[svn-r8339] add a param for MPICH_ODD_EVEN_CLIQUES

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

No reviewer.

8 years ago[svn-r8337] enabled MPE2's configure to process --enable-g=* passed down from MPICH2.
Anthony Chan [Wed, 6 Apr 2011 19:03:59 +0000]
[svn-r8337] enabled MPE2's configure to process --enable-g=* passed down from MPICH2.

8 years ago[svn-r8333] Fall back to older but less memory-intensive version if memory is tight.
Robert Latham [Tue, 5 Apr 2011 16:43:53 +0000]
[svn-r8333] Fall back to older but less memory-intensive version if memory is tight.

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.

8 years ago[svn-r8332] environment variable to select "lockless" file system
Robert Latham [Tue, 5 Apr 2011 16:27:45 +0000]
[svn-r8332] environment variable to select "lockless" file system

from ibm: instead of a hard-coded "use bglockless for PVFS", default to PVFS
but let environment variables override selection.  Maybe someone installed a
new experimental file system?  Maybe a code is unable to prefix the file name
with 'bglockless:' ?  More flexibility without changing current behavior

8 years ago[svn-r8331] from ibm: pad to 16 bytes
Robert Latham [Tue, 5 Apr 2011 16:24:35 +0000]
[svn-r8331] from ibm: pad to 16 bytes

8 years ago[svn-r8330] from ibm: memory leak in alltoallv setup
Robert Latham [Tue, 5 Apr 2011 16:24:02 +0000]
[svn-r8330] from ibm: memory leak in alltoallv setup

8 years ago[svn-r8329] from IBM: better document the environment variables available for tuning
Robert Latham [Tue, 5 Apr 2011 16:07:29 +0000]
[svn-r8329] from IBM: better document the environment variables available for tuning

8 years ago[svn-r8326] fixed typo in printf reported by Fab Tiller
Rajeev Thakur [Mon, 4 Apr 2011 23:52:06 +0000]
[svn-r8326] fixed typo in printf reported by Fab Tiller

8 years ago[svn-r8323] added MPICH_ATTR_FAILED_PROCESSES key to mpi.h
Darius Buntinas [Mon, 4 Apr 2011 18:40:13 +0000]
[svn-r8323] added MPICH_ATTR_FAILED_PROCESSES key to mpi.h

8 years ago[svn-r8320] Use the maximum queue length for listen. "-1" apparently picks the
Pavan Balaji [Fri, 1 Apr 2011 14:43:57 +0000]
[svn-r8320] Use the maximum queue length for listen. "-1" apparently picks the
minimum length, not the maximum.

8 years ago[svn-r8319] Cleanup verbose output to be a bit more organized.
Pavan Balaji [Fri, 1 Apr 2011 14:43:50 +0000]
[svn-r8319] Cleanup verbose output to be a bit more organized.

8 years ago[svn-r8316] add nbc_pmpi_adapter.c for testing the NBC collectives
David Goodell [Thu, 31 Mar 2011 22:45:22 +0000]
[svn-r8316] add nbc_pmpi_adapter.c for testing the NBC collectives

Until we can convert the existing test suite to cover the new nonblocking
collectives, this is the best way that we have to test their implementation.
It simply intercepts all of the blocking collectives and calls the nonblocking
version followed by an MPI_Wait.

No reviewer.

8 years ago[svn-r8315] bugfix: relax/tighten error checking to support MPI_IN_PLACE in many...
David Goodell [Thu, 31 Mar 2011 22:45:21 +0000]
[svn-r8315] bugfix: relax/tighten error checking to support MPI_IN_PLACE in many NBCs

This change also covers cases where the sendbuf or recvbuf is undefined on one
or more processes, such as the recvbuf at non-root procs in MPIX_Ireduce.

No reviewer.

8 years ago[svn-r8314] bugfix: MPIX_Iallreduce support for MPI_IN_PLACE
David Goodell [Thu, 31 Mar 2011 22:45:19 +0000]
[svn-r8314] bugfix: MPIX_Iallreduce support for MPI_IN_PLACE

No reviewer.

8 years ago[svn-r8313] bugfix: comm_self was not hooked or committed like other comms
David Goodell [Thu, 31 Mar 2011 22:45:18 +0000]
[svn-r8313] bugfix: comm_self was not hooked or committed like other comms

No reviewer.

8 years ago[svn-r8312] bugfix: leak in NBC reduction ops
David Goodell [Thu, 31 Mar 2011 22:45:16 +0000]
[svn-r8312] bugfix: leak in NBC reduction ops

No reviewer.

8 years ago[svn-r8311] add maint/decode_handle script to help when debugging
David Goodell [Thu, 31 Mar 2011 22:45:14 +0000]
[svn-r8311] add maint/decode_handle script to help when debugging

No reviewer.

8 years ago[svn-r8310] add the "nonblocking2" test, enable the "nonblocking" test in testlist
David Goodell [Thu, 31 Mar 2011 15:53:51 +0000]
[svn-r8310] add the "nonblocking2" test, enable the "nonblocking" test in testlist

Also fix a small bug in the "nonblocking" test.

Reviewed by buntinas@.

8 years ago[svn-r8309] default implementation of MPIX_Ialltoallw
David Goodell [Thu, 31 Mar 2011 15:53:49 +0000]
[svn-r8309] default implementation of MPIX_Ialltoallw

Mostly straightforward port of MPI_Alltoallw, with an improvement to avoid
repeated memory allocation/deallocation that occurred inside of
MPIC_Sendrecv_replace.  The intercommunicator version is untested.

Reviewed by buntinas@.

8 years ago[svn-r8308] default implementation of MPIX_Iexscan
David Goodell [Thu, 31 Mar 2011 15:53:47 +0000]
[svn-r8308] default implementation of MPIX_Iexscan

This is a direct port of the recursive-doubling algorithm used for MPI_Exscan.

Reviewed by buntinas@.

8 years ago[svn-r8307] default implementation of MPIX_Iscan
David Goodell [Thu, 31 Mar 2011 15:53:46 +0000]
[svn-r8307] default implementation of MPIX_Iscan

This is a mostly mechanical transformation of the algorithms in MPI_Scan,
but the logic for the SMP scan was simplified and a bcast was removed in
that process.

Reviewed by buntinas@.

8 years ago[svn-r8306] commutil.c typo wibble
David Goodell [Thu, 31 Mar 2011 15:53:44 +0000]
[svn-r8306] commutil.c typo wibble

8 years ago[svn-r8304] Fix typo in coverage option in configure
William Gropp [Wed, 30 Mar 2011 03:27:48 +0000]
[svn-r8304] Fix typo in coverage option in configure

8 years ago[svn-r8296] Added checks in bcast to make sure the user passed the same buffer size...
Darius Buntinas [Mon, 28 Mar 2011 23:24:52 +0000]
[svn-r8296] Added checks in bcast to make sure the user passed the same buffer size at all processes.  Since bcast uses different algorithms depending on message size, this won't catch all errors.  Currently the mismatch is reported only at the processes directly communicating with the root in the bcast algorithm.  Reviewed by goodell@.

8 years ago[svn-r8289] ibcast whitespace wibble
David Goodell [Sat, 26 Mar 2011 03:06:38 +0000]
[svn-r8289] ibcast whitespace wibble

Reviewed by buntinas@.

8 years ago[svn-r8288] performance fix for bcast via recursive-doubling scatter/allgather
David Goodell [Sat, 26 Mar 2011 03:06:35 +0000]
[svn-r8288] performance fix for bcast via recursive-doubling scatter/allgather

Changeset r3686 refactored the bcast algorithms into separate functions,
but did so incorrectly.  After r3686, the bcast via
scatter-then-allgather that used the recursive-doubling allgather did
not accurately calculate the amount of data present at each process
after the scatter step.  Instead, it performed the equivalent of an
allgatherv with recvcounts[root]=count and recvcounts[non-root]=0.

This error turns an algorithm with approximate cost:

  lgp.alpha + n.((p-1)/p).beta

into an algorithm with approximate cost:

  lgp.alpha + n.lgp.beta

Reviewed by buntinas@.

8 years ago[svn-r8287] eliminate O(p) memory allocation in MPIR_Bcast_scatter_ring_allgather
David Goodell [Sat, 26 Mar 2011 03:06:32 +0000]
[svn-r8287] eliminate O(p) memory allocation in MPIR_Bcast_scatter_ring_allgather

The new version calculates the recv counts and displacements on the fly,
entirely avoiding non-tmp_buf memory allocation.

Reviewed by buntinas@.

8 years ago[svn-r8286] simplify MPIR_Ibcast_SMP
David Goodell [Sat, 26 Mar 2011 03:06:29 +0000]
[svn-r8286] simplify MPIR_Ibcast_SMP

Unlike MPIR_Bcast_SMP, it makes no attempt to choose the algorithm that will
be used by a subcommunicator.  It simply delegates that to the coll_fns
table that is associated with the subcommunicator.

Reviewed by buntinas@.

8 years ago[svn-r8285] port remaining algorithms to MPIX_Ibcast impl
David Goodell [Sat, 26 Mar 2011 03:06:25 +0000]
[svn-r8285] port remaining algorithms to MPIX_Ibcast impl

The two scatter-then-allgather algorithms from MPI_Bcast have been added.
MPIR_Ibcast_intra has also been updated to select among the three total
algorithms appropriately.  The translation to NBC code was not
purely mechanical, as it required computing recv counts and offsets for
algorithms that previously relied on MPIR_Get_count_impl.

Reviewed by buntinas@.

8 years ago[svn-r8284] convert MPIR_Ibcast_binomial to be non-static
David Goodell [Sat, 26 Mar 2011 03:06:22 +0000]
[svn-r8284] convert MPIR_Ibcast_binomial to be non-static

Also declutter old-style barriers by using MPID_SCHED_BARRIER instead.

Reviewed by buntinas@.

8 years ago[svn-r8283] MPIR_Iallgather_ring comment wibble
David Goodell [Sat, 26 Mar 2011 03:06:19 +0000]
[svn-r8283] MPIR_Iallgather_ring comment wibble

Reviewed by buntinas@.

8 years ago[svn-r8282] default implementation of MPIX_Iallgatherv
David Goodell [Sat, 26 Mar 2011 03:06:15 +0000]
[svn-r8282] default implementation of MPIX_Iallgatherv

This is an adaptation of the algorithms in MPI_Allgatherv.  The _bruck
and _rec_dbl algorithms were altered slightly to statically calculate
the recv counts instead of using MPIR_Get_count_impl, thus avoiding
jumping through callback hoops.  The intercomm version has not been
tested.

Reviewed by buntinas@.

8 years ago[svn-r8280] Added copyright notice
Guillaume Mercier [Wed, 23 Mar 2011 15:18:48 +0000]
[svn-r8280] Added copyright notice

8 years ago[svn-r8279] To check whether a compiler will link an object compiled with -O2 and...
Darius Buntinas [Tue, 22 Mar 2011 21:35:17 +0000]
[svn-r8279] To check whether a compiler will link an object compiled with -O2 and an object compiled without -O2, we were comparing the stdout+err of a compile+link of a singe file with a compile+link of a .c and a .o.  PGCC gives different output when linking one file vs two files, so we got false negatives.  This fixes that by making sure we do a compile+link of a .c and .o in both cases.

8 years ago[svn-r8277] Added a test for intercomm reduce scatter, and corrected a few bugs in...
William Gropp [Tue, 22 Mar 2011 21:22:25 +0000]
[svn-r8277] Added a test for intercomm reduce scatter, and corrected a few bugs in error reporting in reduce-scatter tests

8 years ago[svn-r8275] Comment out unreachable code. Compilers that compile unreachable code...
Jayesh Krishna [Mon, 21 Mar 2011 20:34:09 +0000]
[svn-r8275] Comment out unreachable code. Compilers that compile unreachable code, eg: PGI compilers, failed - function defined only on Windows was present in this part of the code - in the unreachable part of the code.

8 years ago[svn-r8272] Adding missing fn_label - missed in r8234
Jayesh Krishna [Mon, 21 Mar 2011 20:02:27 +0000]
[svn-r8272] Adding missing fn_label - missed in r8234

8 years ago[svn-r8271] Adding missing HAVE_SIGNAL for code added in r8113
Jayesh Krishna [Mon, 21 Mar 2011 19:58:29 +0000]
[svn-r8271] Adding missing HAVE_SIGNAL for code added in r8113

8 years ago[svn-r8270] Bug-fix: We need to be careful on when we want to close the socket. We
Pavan Balaji [Mon, 21 Mar 2011 19:49:30 +0000]
[svn-r8270] Bug-fix: We need to be careful on when we want to close the socket. We
were closing the socket when the stdout socket was full, instead of
retrying.

8 years ago[svn-r8268] Fix error in finding routines for which F90 interfaces should be created...
William Gropp [Sun, 20 Mar 2011 19:37:15 +0000]
[svn-r8268] Fix error in finding routines for which F90 interfaces should be created, also fix some missing LOGICAL args

8 years ago[svn-r8266] Cleanup testlist for some C++ tests
William Gropp [Sun, 20 Mar 2011 19:33:40 +0000]
[svn-r8266] Cleanup testlist for some C++ tests

8 years ago[svn-r8264] Add a test on MPI::COMM_NULL and re-enable a test disabled before the...
William Gropp [Sun, 20 Mar 2011 19:06:04 +0000]
[svn-r8264] Add a test on MPI::COMM_NULL and re-enable a test disabled before the 1.0.5 release and mistakenly never re-enabled

8 years ago[svn-r8262] Bug-fix: signals that are not handled by Hydra should be passed on to
Pavan Balaji [Sun, 20 Mar 2011 04:29:54 +0000]
[svn-r8262] Bug-fix: signals that are not handled by Hydra should be passed on to
the application. Fixes ticket #1460.

8 years ago[svn-r8260] If error handler for files is errors return, there will be no function...
William Gropp [Sat, 19 Mar 2011 15:30:51 +0000]
[svn-r8260] If error handler for files is errors return, there will be no function to call, so we need to return immediately

8 years ago[svn-r8259] Fix the invocation of the error handler - dont just throw an exception.
William Gropp [Fri, 18 Mar 2011 15:38:28 +0000]
[svn-r8259] Fix the invocation of the error handler - dont just throw an exception.

8 years ago[svn-r8258] Enhance output on error for namepub test in C++
William Gropp [Wed, 16 Mar 2011 22:48:49 +0000]
[svn-r8258] Enhance output on error for namepub test in C++

8 years ago[svn-r8257] Move IO parts of C++ error handling tests to the io directory. Also...
William Gropp [Wed, 16 Mar 2011 22:39:58 +0000]
[svn-r8257] Move IO parts of C++ error handling tests to the io directory.  Also a few minor fixes to ensure the expected error handler (throws_exceptions) is used

8 years ago[svn-r8255] default implementation of MPIX_Iallgather
David Goodell [Wed, 16 Mar 2011 20:54:47 +0000]
[svn-r8255] default implementation of MPIX_Iallgather

This is a straightforward port of the algorithms currently used in
MPI_Allgather.  The intercomm implementation has not been tested.

Reviewed by buntinas@.

8 years ago[svn-r8254] fix typo in MPI_Wait error message
David Goodell [Wed, 16 Mar 2011 20:54:43 +0000]
[svn-r8254] fix typo in MPI_Wait error message

No reviewer.

8 years ago[svn-r8252] fixed the coverage test to use AC_LINK_IFELSE.
Anthony Chan [Wed, 16 Mar 2011 20:25:53 +0000]
[svn-r8252] fixed the coverage test to use AC_LINK_IFELSE.

8 years ago[svn-r8249] Adding RMA datatype tests that exercise a bug on BG/P.
James Dinan [Wed, 16 Mar 2011 17:38:28 +0000]
[svn-r8249] Adding RMA datatype tests that exercise a bug on BG/P.

8 years ago[svn-r8247] removed -D flags from FFLAGS and FCFLAGS to resolve req#1450.
Anthony Chan [Tue, 15 Mar 2011 23:54:06 +0000]
[svn-r8247] removed -D flags from FFLAGS and FCFLAGS to resolve req#1450.

8 years ago[svn-r8235] added svn:ignore for README.envvar
Darius Buntinas [Mon, 14 Mar 2011 20:40:15 +0000]
[svn-r8235] added svn:ignore for README.envvar

8 years ago[svn-r8234] Added MPICH_PORT_RANGE to the param utility. Reviewed by Goodell
Darius Buntinas [Mon, 14 Mar 2011 20:29:02 +0000]
[svn-r8234] Added MPICH_PORT_RANGE to the param utility.  Reviewed by Goodell

8 years ago[svn-r8233] Added support for ranges to the param utility. Reviewed by Goodell
Darius Buntinas [Mon, 14 Mar 2011 20:19:45 +0000]
[svn-r8233] Added support for ranges to the param utility.  Reviewed by Goodell

8 years ago[svn-r8230] Revert r8222. When a process terminates, kill only the processes we
Pavan Balaji [Mon, 14 Mar 2011 03:05:58 +0000]
[svn-r8230] Revert r8222. When a process terminates, kill only the processes we
spawned instead of the entire process group as the "mpiexec" process
might be in the same process group as well.

8 years ago[svn-r8227] Provide a link to the wiki page link with the mpiexec -help
Pavan Balaji [Sat, 12 Mar 2011 20:10:51 +0000]
[svn-r8227] Provide a link to the wiki page link with the mpiexec -help
output. Fixes ticket \#1454.

No reviewer.

8 years ago[svn-r8226] Fix a bug in the error handling part when the user provides a bad
Pavan Balaji [Sat, 12 Mar 2011 20:10:39 +0000]
[svn-r8226] Fix a bug in the error handling part when the user provides a bad
binding option. Fixes ticket \#1456.

No reviewer.

8 years ago[svn-r8223] Cleanup of signals. SIGINT and SIGALRM are special signals that have
Pavan Balaji [Sat, 12 Mar 2011 06:23:51 +0000]
[svn-r8223] Cleanup of signals. SIGINT and SIGALRM are special signals that have
specific meanings in Hydra (abort and checkpoint
respectively). Everything else is just passed to the applications.

8 years ago[svn-r8222] Kill all processes in the process group instead of just the processes
Pavan Balaji [Sat, 12 Mar 2011 06:23:47 +0000]
[svn-r8222] Kill all processes in the process group instead of just the processes
spawned by Hydra.

8 years ago[svn-r8220] Generalize the signaling mechanism in Hydra and allow signals such as
Pavan Balaji [Sat, 12 Mar 2011 05:00:54 +0000]
[svn-r8220] Generalize the signaling mechanism in Hydra and allow signals such as
SIGKILL to be forwarded to the application processes.

No reviewer.

8 years ago[svn-r8219] When a process terminates badly kill all processes in all process
Pavan Balaji [Sat, 12 Mar 2011 05:00:45 +0000]
[svn-r8219] When a process terminates badly kill all processes in all process
groups. We had changed this earlier to kill only processes in the dead
process' group to match the MPI standard, but on careful reading that
should only occur when the child has disconnected from the parent
group. Since PMI does not give us this information, Hydra cannot
detect this case and has to cleanup all processes every time.

No reviewer.

8 years ago[svn-r8216] Checked in a temporary fix that replaced all .o/"" by EXEEXT and OBJEXT...
Anthony Chan [Fri, 11 Mar 2011 22:38:33 +0000]
[svn-r8216] Checked in a temporary fix that replaced all .o/"" by EXEEXT and OBJEXT.  This is really a hack as romio/configure.in and aclocal_romio.m4 are some of most outdated configure code within MPICH2.  The issue is that it does not use AC_*_IFESLE or AC_TRY_*, instead it explicitly defines compile and link commands without correctly using other important compiler flags, e.g. CPPFLAGS, LDFLAGS and LIBS.
This is inconsistent with GNU coding standard.  The consequence could be important user or mpich2-modified flags can not be passed down to ROMIO through standard practive.  Also, romio configure runs a weak symbol test which has been done in upper level configure already, this info could have been passed down from upper level configure...

8 years ago[svn-r8214] Added code to generate README.envvar file describing MPICH environment...
Darius Buntinas [Fri, 11 Mar 2011 21:15:53 +0000]
[svn-r8214] Added code to generate README.envvar file describing MPICH environment variables.  Reviewed by goodell@

8 years ago[svn-r8212] remove inaccurate comments in alltoall.c
David Goodell [Fri, 11 Mar 2011 19:51:01 +0000]
[svn-r8212] remove inaccurate comments in alltoall.c

No reviewer.

8 years ago[svn-r8211] default implementation of MPIX_Ialltoall
David Goodell [Fri, 11 Mar 2011 19:51:00 +0000]
[svn-r8211] default implementation of MPIX_Ialltoall

This is a straightforward port of the algorithms in MPI_Alltoall.  The
intercommunicator and MPI_IN_PLACE cases have not been tested.

Reviewed by buntinas@.

8 years ago[svn-r8210] add some debugging sanity checking to MPID_Sched_copy
David Goodell [Fri, 11 Mar 2011 19:50:59 +0000]
[svn-r8210] add some debugging sanity checking to MPID_Sched_copy

This makes sorting out truncation errors easier because it becomes clear
exactly where and how the truncation is occurring.  This code is disabled
in fast builds.

Reviewed by buntinas@.