The bad casting to int was losing information when the message size is
larger than 2GB.

Signed-off-by: Wesley Bland <wbland@mcs.anl.gov>

Signed-off-by: Wesley Bland <wbland@mcs.anl.gov>

This works around problems with writev on some platforms (such as Mac
OSX, 10.9.1) where writev hangs for large messages.

Signed-off-by: Wesley Bland <wbland@mcs.anl.gov>

Using writev on larger than 2GB messages seems to cause some platforms
to hang (e.g., Mac OSX, at least as of 10.9.1).  This patch creates a
new function that reduces the message size being transmitted.  The
function does not attempt to send all data, in order to avoid making
it a blocking function.  A higher-level function would need to check
how much data is sent and retry later if needed.

Signed-off-by: Wesley Bland <wbland@mcs.anl.gov>

Su Huang authored Feb 05, 2014 and Pavan Balaji committed Feb 06, 2014

In the pamid layer, mpir_nbc is set to 0 as the default e.g. the scheduled NBC
work will not be advanced which would cause the job hang. The NBC would work
only if the schedule is advanced. To do so, a user requires to set the
environment variable PAMID_MPIR_NBC to 1.

With MPICH 3.1 being released and NBC is one of support items in the release,
the default for mpir_nbc should be changed from 0 to 1.

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>
Signed-off-by: Pavan Balaji <balaji@mcs.anl.gov>

Pavan Balaji authored Feb 06, 2014 and Kenneth Raffenetti committed Feb 06, 2014

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

Pavan Balaji authored Feb 04, 2014 and Kenneth Raffenetti committed Feb 06, 2014

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

Pavan Balaji authored Feb 04, 2014 and Kenneth Raffenetti committed Feb 06, 2014

No new interfaces were added or removed.  Just the source code
changed.

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

No reviewer.

No reviewer.

No reviewer.

Huiwei Lu authored Feb 02, 2014 and Kenneth Raffenetti committed Feb 03, 2014

'make testing' now will generate another output "summary.junit.xml"
using junit format.

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

No reviewer.

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

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

This test does not do anything special for 4 processes.  rank 0
communicates with each other process sequentially, and waits for all
communication to be over before communicating with a different target.
Reducing the number of processes allows us to reduce the time taken by
the test without affecting any tested parameter.

Signed-off-by: Wesley Bland <wbland@mcs.anl.gov>

Return MPI_ERR_SERVICE when the name being unpublished is not found.

No reviewer.

Signed-off-by: Wesley Bland <wbland@mcs.anl.gov>

The test suite already recognizes --disable-error-checking which is
the right way to tell it to not enable error checking tests.

Signed-off-by: Wesley Bland <wbland@mcs.anl.gov>

Signed-off-by: Wesley Bland <wbland@mcs.anl.gov>

1. enable-fast=all|none should really reflect "all"/"none", not
"some".  We were enable some of the optimizations, but not all, e.g.,
mpit pvars disabling.

2. Making naming consistent and move related checks closer to each
other, so they are easier to verify.

3. Don't control enable-timing, enable-mpit-pvars, and
enable-error-checking from enable-fast.  They have their own configure
options.  enable-fast is kind of weird in that it sets some of its own
configure variables, but also resets variables set by other configure
options making it very confusing for users.  Instead we should point
out in the README what users should do for performance tests.

4. Allow optimization levels like O3 to be used with other enable-fast
options, such as ndebug.

5. Remove some incorrect and/or unnecessary comments.

6. Don't force default compiler optimizations with --disable-fast is
given.

Signed-off-by: Wesley Bland <wbland@mcs.anl.gov>

No reviewer.

Signed-off-by: Jeff Hammond <jeff.science@gmail.com>

Pavan Balaji authored Feb 01, 2014 and Kenneth Raffenetti committed Feb 01, 2014

When the autotools versions are not as expected error out instead of
throwing a warning.  In reality, only the libtool version should be
critical for ABI, not the autoconf or automake version.  But we want
to avoid going backward in autotools versions in releases.

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

No reviewer.

This patch only removes the most obvious pieces of windows code.
There is certainly more windows-related code remaining.

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

Kenneth Raffenetti authored Jan 31, 2014 and Pavan Balaji committed Jan 31, 2014

Check that attempting to unpublish a service the nameserver does
not know about returns MPI_ERR_SERVICE. See 10.4.4 in the MPI standard.

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

Kenneth Raffenetti authored Jan 31, 2014 and Pavan Balaji committed Jan 31, 2014

The error class should be MPI_ERR_SERVICE, not MPI_ERR_NAME, as
defined by the MPI-3 standard section 10.4.4.

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

Kenneth Raffenetti authored Jan 31, 2014 and Pavan Balaji committed Jan 31, 2014

Improves error reporting in the pmi nameservice by checking return
codes from the pmi server. Debug messages are also printed when
enabled.

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

Kenneth Raffenetti authored Jan 31, 2014 and Pavan Balaji committed Jan 31, 2014

When MPI_Unpublish_name is called on the pmi nameserver before any
services have been published, a segfault will occur. Add a check to
determine if anything is published so we can exit early instead.

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

Kenneth Raffenetti authored Jan 30, 2014 and Pavan Balaji committed Jan 30, 2014

Change the default name service system from file to pmi. The file
method can be unreliable on network filesystems with client side
caching.

Closes #2007

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

No reviewer.

By default we still print warnings, but if the
HYDRA_LAUNCHER_SSH_ENABLE_WARNINGS environment variable is set to
false, we disable it.

Refs #2007.

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

Gilles Gouaillardet authored Jan 29, 2014 and Rob Latham committed Jan 30, 2014

Use type (const void *) instead of (void *) and explicitly cast
to (void *) when needed in order to makes gcc -Wall happy.

See #2010

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

Gilles Gouaillardet authored Jan 29, 2014 and Rob Latham committed Jan 30, 2014

Use MPI_Type_size_x instead of MPI_Type_size and make sure the
second argument has type MPI_Count.

Fixes #2010

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

Gilles Gouaillardet authored Jan 29, 2014 and Rob Latham committed Jan 30, 2014

Ensure the third argument of MPIR_Status_set_bytes
has type MPI_Count in order to avoid integer overflow.

Fixes #2010

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

During finalize, we were destroying the COMM_WORLD, COMM_SELF and
COMM_IWORLD communicator objects, and all other associated resources
internally, before waiting for the final progress checks for incoming
messages finished.  This resulted in the following sequence of cleanup:

1. COMM_WORLD got cleaned up.  Internally, there is a check to see if
a group object has been allocated for COMM_WORLD.  If there is one, it
is freed up.

2. We waited for other messages to arrive.  We noticed a failure at
this time, so we try to create a failed process group.  This uses the
COMM_WORLD group internally, causing it to be created again, but with
a reference count of 2, since the code assumes that the first
reference count is always for the original COMM_WORLD.

3. When we try to free the world group, we notice that the reference
count is 2, so we decrement the reference count and not actually free
the object.

Moving the check for incoming messages to happen before the
communicator free, fixes this problem.  Note that the PG finalization
still needs to be the last step since that cleans up all the VCs as
well.

See #1996

Signed-off-by: Wesley Bland <wbland@mcs.anl.gov>

If the PMI process mapping string wraps around to node 0, we were
creating a bad node list of which processes are local and which are
not.  This patch provides a hacky fix for this case by only repeating
the part of the PMI mapping string from the point where it wrapped
around.

The patch is hacky because it assumes that seeing a start node ID of 0
means a wrap around.  This is not necessarily true.  A user-defined
node list can use the node ID 0 without actually creating a wrap
around.  The reason this patch still works in this case is because
Hydra creates a new node list starting from node ID 0 for
user-specified nodes during MPI_Comm_spawn{_multiple}.  If a different
process manager searches for allocated nodes in the user-specified
list, this patch will break.

Fixes #2007.

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

Kenneth Raffenetti authored Jan 29, 2014 and Pavan Balaji committed Jan 29, 2014

When qsort is not available, don't define comparision function and
fallback to simple insertion sort implementation. In the future, a
more general function with fallback should be added in MPL so it
can be used in other cases like comm_split.

Refs #2007

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

The original PMI process mapping parsing code had a number of
assumptions that would allow it to only work on COMM_WORLD.  This
patch corrects these to work for dynamic processes as well.

It also corrects the evaluation of the number of nodes used to be
correct in the general case.

Refs #2007.

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