Wei-keng Liao authored Apr 30, 2014 and Rob Latham committed May 06, 2014

The problem is when defining a filetype using MPI_Type_indexed and the
first few elements of argument blocklens[] are zeros, a collective write
will miss writing some data.

The test program first fills a file with 9 integers with values all
-999.  It then defines a filetype and writes to the file in parallel
with user buffers with value all 1s. Lastly, the file is read back and
checked for contents.

(it's Wei-keng's test case. I just hooked it into ROMIO's test suite)

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

MPI-IO presumes concurrent parallel access, so make that the default.
Applications in strange situations where an MPI-IO program is
manipulating a file while another program manipulates the file can
disable with a hint, but let's make the common, likely case fast.

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

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

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

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

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

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

This reverts commit 12531694.

Jenkins, for one platform/toolchain, still fails these datatype tests. The
results are still correct and performance is no worse, so we need not revert
the whole batch of changes, but the tests are still expected to fail.

No Reviewer

William Gropp authored Mar 10, 2014 and Rob Latham committed May 02, 2014

The best approach for some of these optimizations is through the new
system, so this commit removes some of the work in progress to allow
this patch for dataloop performance to be integrated with MPICH
until the dataloop code is replaced.

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

William Gropp authored Mar 08, 2014 and Rob Latham committed May 02, 2014

A probably copy and paste error in the description of the block indexed
structure.

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

William Gropp authored Mar 08, 2014 and Rob Latham committed May 02, 2014

Statement was placed before one of the values was defined.

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

William Gropp authored Mar 08, 2014 and Rob Latham committed May 02, 2014

This test was created from another one that used communication - in that
process, the use of status.MPI_SOURCE needed to be changed and wasn't in
the original version of this test.

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

William Gropp authored Feb 19, 2014 and Rob Latham committed May 02, 2014

This adds a step that optimizes the dataloop representation, primarily
merging CONTIG blocks with the parent dataloop, such as a VECTOR loop.
It also performs a change of VECTOR of CONTIG with extent > size to
VECTOR of VECTOR; this reduces the stack operations needed to perform
the move.

This is a temporary fix for the dataloop performance.  See the DAME
wiki page (http://wiki.mpich.org/mpich/index.php/DAME

) for current work
on a replacement, higher performance datatype system.

A partial, but not complete, fix for ticket #1788

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

William Gropp authored Feb 19, 2014 and Rob Latham committed May 02, 2014

These tests were inspired by tests in the MPICH-1 test suite
(structpack2.c) or the Intel test suite (vecblklen.c and hvecblklen.c)
that failed with early versions of the dataloop optimization code.

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

William Gropp authored Feb 19, 2014 and Rob Latham committed May 02, 2014

While debugging dataloop optimizations, I needed a little more information
about the errors from this test.  No change to output when there are
no errors.

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

William Gropp authored Feb 19, 2014 and Rob Latham committed May 02, 2014

Mostly minor changes in copyright message, but also retained correction
to erroneous use of the status returned by a send request, and correctly
freeing only allocated buffers.

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

William Gropp authored Feb 19, 2014 and Rob Latham committed May 02, 2014

While adding dataloop optimizations, a number of places where there
were errors in text or in nomunclature were found and fixed.

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

William Gropp authored Feb 03, 2014 and Rob Latham committed May 02, 2014

With these optimizations, most of the datatype performance tests should
pass.  However, indexperf still fails, since these optimizations to not
address the pattern in indexperf.

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

  MPI_Comm_split hang w/ collective_offload=yes on plinux w/o fca

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

No review since F08 binding is experimental now.

No review since F08 binding is experimental now.

No review since F08 binding is experimental now.

Copies a sampling of tests from the f90 directory and converts them
to use the f08 bindings.

No review since it is experimental.

Adds configure logic and a stub Makefile for Fortran 2008 tests.

No review since F08 binding is experimental now.

No review since F08 binding is experimental now.

This includes a configure test to check if the compiler supports F08.

No review since F08 binding is experimental now.

This script is used to generate C wrapper files

No review since F08 binding is experimental now.

The script will
1) generate PMPI files. For PMPI specific procedure names, we use PMPIR_* instead
   of PMPI_*, since the later is reserved by MPI.
2) generate use_mpi_f08/Makefile.mk

No review since F08 binding is experimental now.

Redesigned the binding infrastructer and gave an almost complete F08 binding.
Remaining tasks include :

1) supporting non-contiguous subarray on the C wrapper side.
2) Use a script to autoset the now hard-coded compile time constants.

The directory structure is as following:

use_mpi_f08/                      --- To put F08 binding header files
use_mpi_f08/wrappers_f/           --- To put Fortran wrapper files
use_mpi_f08/wrappers_f/profiling/ --- To put PMPI Fortran wrapper files
use_mpi_f08/wrappers_c/           --- To put C wrapper files (auto-gen'ed or not)

Note we use *.F90 suffix instead of *.f90, since .F90 supports macro preprocess.
This leaves room for future revision. Additionally, the tests directory is removed
since it should not be there.

No review since F08 binding is experimental now.

The MPI-3 standard is broken in a bunch of ways with respect to
subarrays support.  For example, it is unclear as to which functions
can be passed a subarray and which cannot.  This is being fixed by the
WG.  In the meanwhile, we'll disable subarray types.

No review since F08 binding is experimental now.

The code is provided by Steve Oyanagi and Bill Long from Cray.

No review since F08 binding is experimental now.

for one-sided communication, the completion of a local operation is when
the completion handler of the operation is activated. The datatype
associated with the operation can not be freed until the completion
handler is activated which is MPIDI_Win_DoneCB. In the current code, the
mapped datatype is unmapped immediately freed after the routine gets the
control from PAMI calls. The calls could be PAMI_Send, PAMI_Get, PAMI_Rget,
PAMI_Put and PAMI_Rget. Please note that the datatype is needed for PAMI
to process the data.

The fix is to unmap the datatype in MPIDI_Win_DoneCB() and remove all
"unmap" datatype operations from put and get operations.

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

MPID_Win_allocated_shared(), the routine could obtain memory space via either
shmget() or MPIU_Malloc() if one task on a node. In MPID_Win_free(), the
routine only frees shared memory space.

The fix is to free none-shared space if there is one.

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

Request based RMA communication operations require that each operation
allocates a communication request object and associates it with the request
handle (the argument request) that can be used to wait or test for
completion. Both the structures for the request handle and the window
structure are not being freed. The problem is when the request handle
is created, the ref_count of the object is set to 2. When
MPID_Request_releas_inline() is called by a request based RMA operation,
the ref_count of the handle is decremented to 1. The allocated memory spaces
are freed only if ref_count is equal to 0.

To fix the problem, set the ref_count to 1 when the request handle is
created. In MPID_Request_release_inline, the routine decrements the
ref_count to 0 which leads the memory space be freed.

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

In MPID_Win_accumulate, when processing non-contiguous datatype
for origin, the ref_count of datatype is incremented by one which
prevents the space allocated for contents and datatloop from being freed.

  MPID_Datatype_add_ref(req->origin.dt.pointer);

In the routine, the data is processed by MPIR_Localcopy and no need to
increment the reference count (ref_count) of the object by 1.

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

No reviewer

Tested using the mpich testsuite and the following command:

  make testing MPITEST_PROGRAM_WRAPPER=" --ranks-per-node=8 --block R00-M1-N10 --timeout 60 --envs BG_MAPCOMMONHEAP=1 : " MPIEXEC=runjob V=1