For Active Target synchronization, the original implementation
does not guarantee the completion of all ops on target side
when Win_wait / Win_fence returns. It is implemented using a
counter, which is decremented when the last operation from that
origin finishes. Win_wait / Win_fence waits until that counter
reaches zero. Problem is that, when the last operation finishes,
the previous GET-like operation (for example with a large data
volume) may have not finished yet. This breaks the semantic of
Win_wait / Win_fence.

Here we fix this by increment the counter whenever we meet a
GET-like operation, and decrement it when that operation finishes
on target side. This will guarantee that when counter reaches
zero and Win_wait / Win_fence returns, all operations are completed
on the target.

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

This reverts commit 74189446.

Libraries needed to build MPICH (and ROMIO) were missing from EXTERNAL_LIBS.
This posed an issue when interlibrary dependencies were not supported, as in
static builds, causing user programs to fail to build. Fixes #2172

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

Ignore tag matching when MPI_ANY_TAG is passed to a recv operation.

Signed-off-by: Sangmin Seo <sseo@anl.gov>

Since flush call guarantees the completion of operations on target side,
a memory barrier on target side is necessary in order to ensure all store
instructions are exactly performed before that flush call finished on origin
side.

Signed-off-by: Xin Zhao <xinzhao3@illinois.edu>

  Added an empty implementation for MPID_Comm_get_all_failed_procs
  and MPID_Comm_revoke in pamid.

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

Signed-off-by: Sameh Sharkawi <sssharka@us.ibm.com>

Signed-off-by: Sameh Sharkawi <sssharka@us.ibm.com>

Pass tail pointer to single direction linked-list API so that
we optimize CONCAT and APPEND operations from O(N) to O(1).

Note that we do not need MPL_LL_APPEND_VS2008 anymore because
we no longer need to work around temporary variable. Similiarly,
we can support MPL_LL_CONCAT for VS2008 now.

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

The original implementation of FENCE and PSCW does not
guarantee the remote completion of issued-out RMA operations
when MPI_Win_complete and MPI_Win_fence returns. They only
guarantee the local completion of issued-out operations and
the completion of coming-in operations. This is not correct
if we try to get updated values on target side using synchronizations
with MPI_MODE_NOCHECK.

Here we modify it by making runtime wait for ACKs from all
targets before returning from MPI_Win_fence and MPI_Win_complete.

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

The release.pl code was essentially a simple svn to git conversion.
With our current testing model, some of this code is irrelevant.

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

Igor Ivanov authored Sep 12, 2014 and Pavan Balaji committed Sep 20, 2014

- Changed MXM version that is supported to 3.1+
- Added additional debug output for iov
- Added check for datatype mismatch during noncontiguous data

Signed-off-by: Igor Ivanov <Igor.Ivanov@itseez.com>

Enable using MPI_Win_allocate, which by default turns
on SHM window, in PSCW tests, so that we can test PSCW
on SHM window.

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

If MPI_MODE_NOCHECK is passed to MPI_Win_start, we should not
wait for synchronization from targets.

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

recv_post_msgs function on origin process is used to wait
for synchronization from target processes, and it can be
called from MPI_Win_start (for targets on SHM) and
MPI_Win_complete (for targets not on SHM). Here we fix the
bug so that origin will not wait for SHM targets again in
MPI_Win_complete.

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

Fixed a bug in MPID_Win_allgather to avoid passing 0 for memory registration

(ibm) D199265

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

long ago we thought there might be more than one process flushing data,
but that has not borne out.  the much faster MPI_Bcast will suffice, and
we can avoid an MPI_Allreduce.

Signed-off-by: Paul Coffman <pkcoff@us.ibm.com>

comm_split might scale poorly on some systems, and we don't even use the
resulting communicator.  it was used as a marker, but we have enough
other information.

Ken was right: I was being too clever before.  now, simply and more
explicitly check if anyone has a NULL systemwide-hint info object.

Junchao Zhang authored Sep 11, 2014 and Rob Latham committed Sep 11, 2014

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

MPI_Pcontrol does not have the ierror argument.

Fixes #2173

Signed-off-by: William Gropp <wgropp@illinois.edu>

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

In MPI, the predefined named constant handles like MPI_INT are link-time
but not necessarily compile-time constants, so they should not be used in
switch/case clauses. If you do so, this kind of code only compiles with
MPICH but not OpenMPI.

Fixes #2169, #2171

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

Deregister until the total amount released exceeds the requested length.
If a registration fails after deregistering some memory regions, then
deregister a next memory region and try to re-register.

There is a maximum message size supported by HCA port. So it's necessary
to divide a memory area to segments when a message size is larger than it.

Resolved warnings reported by --enable-strict configuration.

Signed-off-by: Igor Ivanov <Igor.Ivanov@itseez.com>

This patch implements the "eager" protocol used in MPI_Comm_idup.
Without this patch, test/mpi/threads/comm/comm_idup will crash due to
segmentation fault caused by livelock.

It uses the same algorithm as multi-threaded
MPIR_Get_contextid_sparse_group to avoid deadlock in multi-threaded
context_id allocation, but in a nonblocking way. The nonblocking
mechanism is explained in the comment of function
sched_get_cid_nonblock.

Fixes #2069

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

No reviewer.

No reviewer.

First, cache every SHM window created by Win_allocate or
Win_allocate_shared into a global list, and unlink it in Win_free.

Then, when user calls Win_create for a new window, check user specified
buffer and comm. Enable local SHM communicaiton in the new window if it
matches a cached SHM window. It is noted that all the shared resources
are still freed by the original SHM window.

Matching a SHM window must satisfy following two conditions:
1. The new node comm is equal to, or a subset of the SHM node comm.
(Note that in the other cases where two node comms are overlapped,
although the overlapped processes could be logically shared, it is not
supported for now. To support this, we need to fist modify the implementation
of RMA operations in order to remember shared status per target but not
just compare its node_id).
2. The buffer is in the range of the SHM segment across local processes
in original SHM window (a contigunous segment is mapped across local
processes regardless of whether alloc_shared_noncontig is set).

Resolves #2161

Signed-off-by: Xin Zhao <xinzhao3@illinois.edu>

No Reviewer

Paul Coffman authored Aug 31, 2014 and Rob Latham committed Sep 03, 2014

The code that chose the aggs for gpfsmpio_bridgeringagg was written very
inefficiently resulting in a large performance degradation during the
MPI_File_open.  That code has been rewritten to squash the five torus
dimensions into one dimension and to loop over that new representation
fewer times. Now there is no performance degradation.

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

We do not often enough compile with --enable-strict on blue gene

No Reviewer

Pavan Balaji authored Sep 03, 2014 and Kenneth Raffenetti committed Sep 03, 2014

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

Pavan Balaji authored Sep 03, 2014 and Kenneth Raffenetti committed Sep 03, 2014

We had some settings in the nemesis subconfigure, which seems
misplaced.

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

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

We were mixing up external libraries, which are prebuilt, and internal
libraries on which we need to add dependencies.  This patch separates
these out and only adds build dependencies on the internal libraries.

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

The original MPICH code was directly adding external library
dependencies into LIBS.  This forces configure to use these libraries,
thus requiring the user to point to these libraries through
LD_LIBRARY_PATH (for shared libraries).  This is unnecessary since
those libraries are only needed when building the executables and not
the rest of MPICH.

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