Masamichi Takagi authored Jan 20, 2014 and Pavan Balaji committed Jan 19, 2014

Fix warnings and errors in the IB netmod when configured with
--enable-strict and --enable-g=all.

Signed-off-by: Pavan Balaji <balaji@mcs.anl.gov>

Masamichi Takagi authored Jan 20, 2014 and Pavan Balaji committed Jan 19, 2014

Signed-off-by: Pavan Balaji <balaji@mcs.anl.gov>

No interfaces were added or removed.  There were just source code
changes.

Signed-off-by: Ken Raffenetti <raffenet@mcs.anl.gov>

No Reviewer.

On some machines the iterations take unusually long.  If they are
getting to be larger than a predefined amount, break out of that loop.

Fixes #1669.

Signed-off-by: Ken Raffenetti <raffenet@mcs.anl.gov>

The code was unparseable to make any changes.

Kenneth Raffenetti authored Dec 20, 2013 and Pavan Balaji committed Jan 18, 2014

Simplify logic in compile wrapper scripts. Use configure substitutions
where possible to better match pkg-config style.

Signed-off-by: Pavan Balaji <balaji@mcs.anl.gov>

Includes the following modifications by Pavan Balaji:

Remove the PAC_COMPILER_SHLIB_FLAGS usage, instead of modifying the
macro in confdb.

The ordering of flags in mpicc and friends does not match that of
pkg-config.  This is because of two reasons.

1. pkg-config reorders flags when it outputs them.  This requires us
to manually adjust the flags in mpicc to match up, and is error prone.

2. mpicc and friends provide LDFLAGS before the user-specified flags,
followed by the include and library directories.  This is to make sure
that the LDFLAGS are listed before the application source file.
Reordering them to match pkg-config loses this flexibility.

Signed-off-by: Ken Raffenetti <raffenet@mcs.anl.gov>

Kenneth Raffenetti authored Dec 09, 2013 and Pavan Balaji committed Jan 18, 2014

Add rpath flags to pkg-config to match compiler wrappers. Fixes #1044

Signed-off-by: Pavan Balaji <balaji@mcs.anl.gov>

Some datatype performance tests in the MPICH test suite fail:
(perf/twovec,  perf/nestvec, perf/nestvec2, perf/indexperf,
perf/transp-datatype).

This changeset introduces a few optimizations that operate on the
dataloop representation to make it more performant.  perf/indexperf
should still fail under these changes.

Original-author: Bill Gropp <wgropp@illinois.edu>

See #1788, for which this resolves some but not all performance issues.

Signed-off-by: Rob Latham <robl@mcs.anl.gov>

The test in test/mpi/perf/twovec made invalid assumptions about the
performance of two MPI datatype creation routines.  This is a hard test to
get right, but this version is more likely to avoid falsely signalling
an error.

This was meant to test out the case when MPI_Test is not nonblocking.
However, we ended up assuming that MPI_Win_lock will be nonblocking.
That is not specified by the standard and might not be true.
Commenting this out till be find a better way to test the original
problem with MPI_Test.

Fixes #1910.

Signed-off-by: Rajeev Thakur <thakur@mcs.anl.gov>

Fix a transposition in use of fprintf and remove benign trailing blanks

Added mpivars to the programs known to Automake so that it will
build and install it.

If we need to build a separate profiling library, make sure we add that to
the link link.

This adds a program that should work with any MPI 3 implementation to
show the control and performance variables defined at MPI_Init time.

The original code here was a bit subtle, so this fix adds a comment
explaining the particular test and provides the fallback message
when a datatype has no name associated with it.

When mpich is configured with ./configure --enable-coverage ..., we put
-fprofile-arcs -ftest-coverage to CXXFLAGS. According to gcc manual,
-fprofile-arcs -ftest-coverage can be replaced by --coverage. But unfortunately,
I found though both add -lgcov when linking libmpichcxx.so, -lgcov is put to the
end of the link command, which is too late.

Adding -lgcov directly to LIBS will put -lgcov before -lc, which makes
the linker correctly resolve symbol 'atexit' in linking.

Fixes #2000

Signed-off-by: Huiwei Lu <huiweilu@mcs.anl.gov>

Signed-off-by: Huiwei Lu <huiweilu@mcs.anl.gov>

Rob Latham authored Jan 11, 2014 and Pavan Balaji committed Jan 11, 2014

Clean up memory allocated in write (for read-modify-write) when
encountering error.  Still todo: blue gene, lustre, and any other driver
that acts this way.

closes #1994

Signed-off-by: Pavan Balaji <balaji@mcs.anl.gov>

Michael Blocksome authored Jan 08, 2014 and Pavan Balaji committed Jan 10, 2014

Signed-off-by: Su Huang <suhuang@us.ibm.com>

Michael Blocksome authored Jan 08, 2014 and Pavan Balaji committed Jan 10, 2014

Signed-off-by: Su Huang <suhuang@us.ibm.com>
Signed-off-by: Pavan Balaji <balaji@mcs.anl.gov>

Give a "likely" attribute to the branch that checks if an MPI process
is initialized in THREAD_MULTIPLE or not.  This should reduce the cost
of the branch for applications that require THREAD_MULTIPLE.

The per-object mode in the pamid branch is intended for the case where
comm-threads are used.  In this case, the application is mostly in
THREAD_MULTIPLE mode, even if it didn't explicitly initialize it as
such.  The only time it'll not be in that mode is during
initialization before the appropriate mutexes have been setup.

See #1900.

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>
Signed-off-by: Rob Latham <robl@mcs.anl.gov>

Disabling runtime check for thread-safety is not a correct option.
This would cause memory allocation and string manipulation routines to
be unusable before the thread-safety level is set.  Fixes #1900.

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>
Signed-off-by: Rob Latham <robl@mcs.anl.gov>

This patch has two related parts.

1. We initialize the isThreaded runtime check to FALSE before doing
the cvar checks.  This is because the cvar initialization uses memory
allocation and string duping, which internally use mutexes.  But the
mutexes are not initialized yet, so disabling isThreaded would make
sure we don't use them.

2. Updates to the pamid device to respect the isThreaded variable.

This is needed since we were initializing cvars before setting the
thread-level, so the thread-safety macros are not initialized at that
point.  See #1900.

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>
Signed-off-by: Rob Latham <robl@mcs.anl.gov>

This test still needs to be corrected, but for now it will be xfail. It is
being tracked with ticket #1996.

No reviewer

Wesley Bland authored Jan 09, 2014 and Pavan Balaji committed Jan 09, 2014

This test needs to be debugged, but it can wait for v3.1.1 since this is an
experimental failure.

Signed-off-by: Pavan Balaji <balaji@mcs.anl.gov>

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>

Sameer Kumar authored Jan 07, 2014 and Pavan Balaji committed Jan 08, 2014

Full fix

Fix for get accumulate that sends contig ack back and then scatters result buffer on the src node.

Remove unused params.

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>

Su Huang authored Jan 06, 2014 and Pavan Balaji committed Jan 08, 2014

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>

Su Huang authored Jan 06, 2014 and Pavan Balaji committed Jan 08, 2014

Also clean up warnings.

  src/mpid/pamid/src/dyntask/mpidi_port.c:53: warning: useless storage class
  specifier in empty declaration

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>

Su Huang authored Dec 31, 2013 and Pavan Balaji committed Jan 08, 2014

The segfault occurred while trying to free an already freed request
handle in MPIDI_Win_DoneCB. To fix the problem, the second MPIU_Free(req)
is removed from the routine.

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>

Su Huang authored Oct 30, 2013 and Pavan Balaji committed Jan 08, 2014

PAMID: fixes for RMA shared related, lock_all and unlock_all functions

The following has been updated:
- MPID_Win_allocated_shared
- MPID_Win_shared_query
- several functions in mpid_win_lock_all.c, mpid_win_lock.c, mpid_win_free.c and
  structures in mpidi_onesided.h, mpidi_datatypes.h have been modified for
  scalability support for MPID_Win_lock_all() and MPID_Win_unlock_all()

PAMID: fix MPID_Win_allocate_shared, MPID_Win_lock_all/unlock_all and MPID_Win_shared_query

Add changes based on code review feedback from the team:
1) update getPageSize to ensure that the pagesize is obtained from the range of
   passed in address.
2) don't call dispatcher if lockQ[index].done == 1
   if (!lockQ[index].done)
     MPID_PROGRESS_WAIT_WHILE(lockQ[index].done == 0);
3) in mpid_win_shared_query()
   replace
     MPID_assert(win->create_flavor == MPI_WIN_FLAVOR_SHARED);
   by
     MPIU_ERR_CHKANDSTMT((win->create_flavor != MPI_WIN_FLAVOR_SHARED), mpi_errno,
                         MPI_ERR_RMA_FLAVOR, return mpi_errno, "**rmaflavor");
3) some minor fixes.

PAMID: modify MPI_Win_flush_local, MPI_Win_lock etc for better performance

The current implementation for the subject mentioned function is to allocate
two counters for each rank in the window group. The design could cause a scaling
issue.  The fix is to update MPI_Win_flush_local etc with two counters per
window approach.

The changes also include the follwoing:
- provides mutex_lock/mutex_unlock for atomic operations in shared window
- fixes some bugs in handling shared window.
- removes the shared segment with IPC_RMID in MPI_Win_free.

PAMID: fixed base address for each rank in a window group

For shared window, the base address for each rank should not be
exchanged among ranks in a window group. Without the fix, the job
will be terminated with segfault.

(ibm) F189033
(ibm) D194640

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>

Su Huang authored Oct 16, 2013 and Pavan Balaji committed Jan 08, 2014

added MPID_Win_allocate_shared and MPID_Win_shared_query support and
fixes on passing incorrect rank to MPIDI_WinCtrlSend in MPIDI_WinPost_post
and MPIDI_WinComplete_post

minor fixes for 'win allocate shared'

- compiles for bgq
- UNRESOLVED: undefined shm key on bgq, or any other time the required
  environment variables are not set
- UNRESOLVED: multiple calls to 'win allocate shared' will use the same
  shm key

(ibm) F189033

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>

Renamed WIN_SEPARATE to WIN_UNIFIED.  pamid's WIN_SYNC already does a
full memory barrier, so we should be all set for UNIFIED support.

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>

the code in glue_romio replicates a bit of pointer-size assertion logic.
it's only used by ROMIO, though.  if mpitypedefs.h is not included (see
recent standalone-romio cleanups) we will get an undefined symbol at
link time (MPI_AINT_CAST_TO_VOID_PTR).  Use equivalent
ADIOI_AINT_CAST_TO_VOID_PTR instead for this ROMIO-related code.

Signed-off-by: Jed Brown <jedbrown@mcs.anl.gov>

The current implementation of SMP-awareness in MPI_Barrier was in
MPIR_Barrier_impl.  This makes the default implementation of barrier
SMP-aware.  However, if a device overrides barrier and then calls back
the default implementation through MPIR_Barrier, it can no longer take
advantage of SMP-awareness.  This patch moves the SMP-aware
implementation to MPIR_Barrier_intra, which is called by
MPIR_Barrier_impl through MPIR_Barrier, in the default implementation.
See #1957.

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>

Use MPIR_Bcast as an inline function, and let the compiler do its job,
instead of duplicating the code.

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>

If the user directly provided the host list or if the user provided a
modifier to the list of hosts we obtained from the resource manager
(e.g., changed the number of processes per node, or sorted them), we
consider the host list as user-provided.  Fixes #1575.

Signed-off-by: Junchao Zhang <jczhang@mcs.anl.gov>

Signed-off-by: Junchao Zhang <jczhang@mcs.anl.gov>