1768 lines (1194 sloc) 70.4 KB
Changes in 3.3.1
# Removed support for logical reduction operations on floating point
Changes in 3.3
# CH4 Device: A new device layer implementation designed for low software
overheads. CH4 has experimental support for OFI and UCX network libraries,
and POSIX shared memory. Thanks to Intel, Mellanox, and RIKEN AICS for
participating in the CH4 coding effort.
# Fixed SLURM integration in Hydra for new node list format.
# Added support for PMIx ( client
library in CH4 netmods. Note that you must use a compatible PMIx
server in this configuration.
# Better organization of collectives in the MPI layer. The new
scheme, which de-couples implementation from selection logic,
enables easier integration of additional algorithms.
# TSP collectives framework: A C++-template style framework for
collective algorithms is added to allow single collective
implementation to move data over generic or device-specific
transport functions.
# Improvements to derived datatype testing (DTPools -
# Added new "non-catastrophic" error codes to expose internal
resource exhaustion.
# Added info hints to MPI_Comm_split_type to support splitting
communicators by machine topology. Both on-node (socket, core,
etc.) and off-node (switch-level) hints are defined.
# Improvements to MPI_THREAD_MULTIPLE in CH4 through new thread safety
models at the Virtual Network Interface (VNI) level. This introduces two
new models that leverage work-queues to offload operations and improve
scalability under contention.
# Message Driven Thread Activation (MDTA). An alternative locking
model is defined for MPI_THREAD_MULTIPLE in CH4.
# Added PMI usage optimizations for business card exchange in CH4
# Improvements on MPI_Abort. MPI_Abort invoked on subcommunicators will
only abort the connected processes within that communicator.
# Cleanup of whitespace (ch3 excluded) using the
maint/ script. For instructions on how to update
PRs/branches based on MPICH before the cleanup, see
# Removed the PAMI device and poe PMI client.
# C99 compiler support is now required to build MPICH.
# Several other minor bug fixes, memory leak fixes, and code cleanup.
A full list of changes is available at the following link:
A list of bugs that have been fixed is available at the following
Changes in 3.2
# Added support for MPI-3.1 features including nonblocking collective I/O,
address manipulation routines, thread-safety for MPI initialization,
pre-init functionality, and new MPI_T routines to look up variables
by name.
# Fortran 2008 bindings are enabled by default and fully supported.
# Added support for the Mellanox MXM InfiniBand interface. (thanks
to Mellanox for the code contribution).
# Added support for the Mellanox HCOLL interface for collectives.
(thanks to Mellanox for the code contribution).
# Significant stability improvements to the MPICH/portals4
# Completely revamped RMA infrastructure including several
scalability improvements, performance improvements, and bug fixes.
# Added experimental support for Open Fabrics Interfaces (OFI) version 1.0.0. (thanks to Intel for code contribution)
# The Myrinet MX network module, which had a life cyle from 1.1 till
3.1.2, has now been deleted.
# Several other minor bug fixes, memory leak fixes, and code cleanup.
A full list of changes is available at the following link:
A full list of bugs that have been fixed is available at the
following link:
Changes in 3.1.3
# Several enhancements to Portals4 support.
# Several enhancements to PAMI (thanks to IBM for the code contribution).
# Several enhancements to the CH3 RMA implementation.
# Several enhancements to ROMIO.
# Fixed deadlock in multi-threaded MPI_Comm_idup.
# Several other minor bug fixes, memory leak fixes, and code cleanup.
A full list of changes is available at the following link:
A full list of bugs that have been fixed is available at the
following link:
Changes in 3.1.2
# Significant enhancements to the BG/Q device, especially for RMA and
shared memory functionality.
# Several enhancements to ROMIO.
# Upgraded to hwloc-1.9.
# Added more Fortran 2008 (F08) tests and fixed a few F08 binding bugs.
Now all MPICH F90 tests have been ported to F08.
# Updated weak alias support to align with gcc-4.x
# Minor enhancements to the CH3 RMA implementation.
# Better implementation of MPI_Allreduce for intercommunicator.
# Added environment variables to control memory tracing overhead.
# Added flags to enable C99 mode with Solaris compilers.
# Updated implementation of MPI-T CVARs of type MPI_CHAR, as interpreted
in MPI-3.0 Errata.
# Several other minor bug fixes, memory leak fixes, and code cleanup.
A full list of changes is available at the following link:
A full list of bugs that have been fixed is available at the
following link:
Changes in 3.1.1
# Blue Gene/Q implementation supports MPI-3. This release contains a
functional and compliant Blue Gene/Q implementation of the MPI-3 standard.
Instructions to build on Blue Gene/Q are on the wiki:
# Fortran 2008 bindings (experimental). Build with --enable-fortran=all. Must have
a Fortran 2008 + TS 29113 capable compiler.
# Significant rework of MPICH library management and which symbols go
into which libraries. Also updated MPICH library names to make
them consistent with Intel MPI, Cray MPI and IBM PE MPI. Backward
compatibility links are provided for older mpich-based build
# The ROMIO "Blue Gene" driver has seen significant rework. We have separated
"file system" features from "platform" features, since GPFS shows up in more
places than just Blue Gene
# New ROMIO options for aggregator selection and placement on Blue Gene
# Optional new ROMIO two-phase algorithm requiring less communication for
certain workloads
# The old ROMIO optimization "deferred open" either stopped working or was
disabled on several platforms.
# Added support for powerpcle compiler. Patched libtool in MPICH to support
little-endian powerpc linux host.
# Fixed the prototype of the Reduce_local C++ binding. The previous
prototype was completely incorrect. Thanks to Jeff Squyres for
reporting the issue.
# The mpd process manager, which was deprecated and unsupported for
the past four major release series (1.3.x till 3.1), has now been
deleted. RIP.
# Several other minor bug fixes, memory leak fixes, and code cleanup.
A full list of changes is available at the following link:
A full list of bugs that have been fixed is available at the
following link:
Changes in 3.1
# Implement runtime compatibility with MPICH-derived implementations as per
the ABI Compatibility Initiative (see for more
# Integrated MPICH-PAMI code base for Blue Gene/Q and other IBM
# Several improvements to the SCIF netmod. (code contribution from
# Major revamp of the MPI_T interface added in MPI-3.
# Added environment variables to control a lot more capabilities for
collectives. See the README.envvar file for more information.
# Allow non-blocking collectives and fault tolerance at the same
time. The option MPIR_PARAM_ENABLE_COLL_FT_RET has been deprecated as
it is no longer necessary.
# Improvements to MPI_WIN_ALLOCATE to internally allocate shared
memory between processes on the same node.
# Performance improvements for MPI RMA operations on shared memory
# Enable shared library builds by default.
# Upgraded hwloc to 1.8.
# Several improvements to the Hydra-SLURM integration.
# Several improvements to the Hydra process binding code. See the
Hydra wiki page for more information:
# MPICH now supports operations on very large datatypes (those that describe
more than 32 bits of data). This work also allows MPICH to fully support
MPI-3's introduction of MPI_Count.
# Several other minor bug fixes, memory leak fixes, and code cleanup.
A full list of changes is available at the following link:
A full list of bugs that have been fixed is available at the
following link:
Changes in 3.0.4
# BUILD SYSTEM: Reordered the default compiler search to prefer Intel
and PG compilers over GNU compilers because of the performance
WARNING: If you do not explicitly specify the compiler you want
through CC and friends, this might break ABI for you relative to
the previous 3.0.x release.
# OVERALL: Added support to manage per-communicator eager-rendezvous
# PM/PMI: Performance improvements to the Hydra process manager on
large-scale systems by allowing for key/value caching.
# Several other minor bug fixes, memory leak fixes, and code cleanup.
A full list of changes is available at the following link:
Changes in 3.0.3
# RMA: Added a new mechanism for piggybacking RMA synchronization operations,
which improves the performance of several synchronization operations,
including Flush.
# RMA: Added an optimization to utilize the MPI_MODE_NOCHECK assertion in
passive target RMA to improve performance by eliminating a lock request
# RMA: Added a default implementation of shared memory windows to CH3. This
adds support for this MPI 3.0 feature to the ch3:sock device.
# RMA: Fix a bug that resulted in an error when RMA operation request handles
where completed outside of a synchronization epoch.
# PM/PMI: Upgraded to hwloc-1.6.2rc1. This version uses libpciaccess
instead of libpci, to workaround the GPL license used by libpci.
# PM/PMI: Added support for the Cobalt process manager.
# BUILD SYSTEM: allow MPI_LONG_DOUBLE_SUPPORT to be disabled with a configure
# FORTRAN: fix MPI_WEIGHTS_EMPTY in the Fortran bindings
# MISC: fix a bug in MPI_Get_elements where it could return incorrect values
# Several other minor bug fixes, memory leak fixes, and code cleanup.
A full list of changes is available at the following link:
Changes in 3.0.2
# PM/PMI: Upgrade to hwloc-1.6.1
# RMA: Performance enhancements for shared memory windows.
# COMPILER INTEGRATION: minor improvements and fixes to the clang static type
checking annotation macros.
# MPI-IO (ROMIO): improved error checking for user errors, contributed by IBM.
# MPI-3 TOOLS INTERFACE: new MPI_T performance variables providing information
about nemesis communication behavior and and CH3 message matching queues.
# TEST SUITE: "make testing" now also outputs a "summary.tap" file that can be
interpreted with standard TAP consumer libraries and tools. The
"summary.xml" format remains unchanged.
# GIT: This is the first release built from the new git repository at A few build system mechanisms have changed because of this
# BUG FIX: resolved a compilation error related to LLONG_MAX that affected
several users (ticket #1776).
# BUG FIX: nonblocking collectives now properly make progress when MPICH is
configured with the ch3:sock channel (ticket #1785).
# Several other minor bug fixes, memory leak fixes, and code cleanup.
A full list of changes is available at the following link:
Changes in 3.0.1
# PM/PMI: Critical bug-fix in Hydra to work correctly in multi-node
# A full list of changes is available using:
svn log -r10790:HEAD
... or at the following link:
Changes in 3.0
# MPI-3: All MPI-3 features are now implemented and the MPI_VERSION
bumped up to 3.0.
# OVERALL: Added support for ARM-v7 native atomics
# MPE: MPE is now separated out of MPICH and can be downloaded/used
as a separate package.
# PM/PMI: Upgraded to hwloc-1.6
# Several other minor bug fixes, memory leak fixes, and code cleanup.
A full list of changes is available using:
svn log -r10344:HEAD
... or at the following link:
Changes in 1.5
# OVERALL: Nemesis now supports an "--enable-yield=..." configure
option for better performance/behavior when oversubscribing
processes to cores. Some form of this option is enabled by default
on Linux, Darwin, and systems that support sched_yield().
# OVERALL: Added support for Intel Many Integrated Core (MIC)
architecture: shared memory, TCP/IP, and SCIF based communication.
# OVERALL: Added support for IBM BG/Q architecture. Thanks to IBM
for the contribution.
# MPI-3: const support has been added to mpi.h, although it is
disabled by default. It can be enabled on a per-translation unit
basis with "#define MPICH2_CONST const".
# MPI-3: Added support for MPIX_Type_create_hindexed_block.
# MPI-3: The new MPI-3 nonblocking collective functions are now
available as "MPIX_" functions (e.g., "MPIX_Ibcast").
# MPI-3: The new MPI-3 neighborhood collective routines are now available as
"MPIX_" functions (e.g., "MPIX_Neighbor_allgather").
# MPI-3: The new MPI-3 MPI_Comm_split_type function is now available
as an "MPIX_" function.
# MPI-3: The new MPI-3 tools interface is now available as "MPIX_T_"
functions. This is a beta implementation right now with several
limitations, including no support for multithreading. Several
performance variables related to CH3's message matching are exposed
through this interface.
# MPI-3: The new MPI-3 matched probe functionality is supported via
the new routines MPIX_Mprobe, MPIX_Improbe, MPIX_Mrecv, and
# MPI-3: The new MPI-3 nonblocking communicator duplication routine,
MPIX_Comm_idup, is now supported. It will only work for
single-threaded programs at this time.
# MPI-3: MPIX_Comm_reenable_anysource support
# MPI-3: Native MPIX_Comm_create_group support (updated version of
the prior MPIX_Group_comm_create routine).
# MPI-3: MPI_Intercomm_create's internal communication no longer interferes
with point-to-point communication, even if point-to-point operations on the
parent communicator use the same tag or MPI_ANY_TAG.
# MPI-3: Eliminated the possibility of interference between
MPI_Intercomm_create and point-to-point messaging operations.
# Build system: Completely revamped build system to rely fully on
autotools. Parallel builds ("make -j8" and similar) are now supported.
# Build system: rename "./maint/updatefiles" --> "./" and
"" --> ""
# JUMPSHOT: Improvements to Jumpshot to handle thousands of
timelines, including performance improvements to slog2 in such
# JUMPSHOT: Added navigation support to locate chosen drawable's ends
when viewport has been scrolled far from the drawable.
# PM/PMI: Added support for memory binding policies.
# PM/PMI: Various improvements to the process binding support in
Hydra. Several new pre-defined binding options are provided.
# PM/PMI: Upgraded to hwloc-1.5
# PM/PMI: Several improvements to PBS support to natively use the PBS
# Several other minor bug fixes, memory leak fixes, and code cleanup.
A full list of changes is available using:
svn log -r8478:HEAD
... or at the following link:
Changes in 1.4.1
# OVERALL: Several improvements to the ARMCI API implementation
within MPICH2.
# Build system: Added beta support for DESTDIR while installing
# PM/PMI: Upgrade hwloc to 1.2.1rc2.
# PM/PMI: Initial support for the PBS launcher.
# Several other minor bug fixes, memory leak fixes, and code cleanup.
A full list of changes is available using:
svn log -r8675:HEAD
... or at the following link:
Changes in 1.4
# OVERALL: Improvements to fault tolerance for collective
operations. Thanks to Rui Wang @ ICT for reporting several of these
# OVERALL: Improvements to the universe size detection. Thanks to
Yauheni Zelenko for reporting this issue.
# OVERALL: Bug fixes for Fortran attributes on some systems. Thanks
to Nicolai Stange for reporting this issue.
# OVERALL: Added new ARMCI API implementation (experimental).
# OVERALL: Added new MPIX_Group_comm_create function to allow
non-collective creation of sub-communicators.
# FORTRAN: Bug fixes in the MPI_DIST_GRAPH_ Fortran bindings.
# PM/PMI: Support for a manual "none" launcher in Hydra to allow for
higher-level tools to be built on top of Hydra. Thanks to Justin
Wozniak for reporting this issue, for providing several patches for
the fix, and testing it.
# PM/PMI: Bug fixes in Hydra to handle non-uniform layouts of hosts
better. Thanks to the MVAPICH group at OSU for reporting this issue
and testing it.
# PM/PMI: Bug fixes in Hydra to handle cases where only a subset of
the available launchers or resource managers are compiled
in. Thanks to Satish Balay @ Argonne for reporting this issue.
# PM/PMI: Support for a different username to be provided for each
host; this only works for launchers that support this (such as
# PM/PMI: Bug fixes for using Hydra on AIX machines. Thanks to
Kitrick Sheets @ NCSA for reporting this issue and providing the
first draft of the patch.
# PM/PMI: Bug fixes in memory allocation/management for environment
variables that was showing up on older platforms. Thanks to Steven
Sutphen for reporting the issue and providing detailed analysis to
track down the bug.
# PM/PMI: Added support for providing a configuration file to pick
the default options for Hydra. Thanks to Saurabh T. for reporting
the issues with the current implementation and working with us to
improve this option.
# PM/PMI: Improvements to the error code returned by Hydra.
# PM/PMI: Bug fixes for handling "=" in environment variable values in
# PM/PMI: Upgrade the hwloc version to 1.2.
# COLLECTIVES: Performance and memory usage improvements for MPI_Bcast
in certain cases.
# VALGRIND: Fix incorrect Valgrind client request usage when MPICH2 is
built for memory debugging.
# BUILD SYSTEM: "--enable-fast" and "--disable-error-checking" are once
again valid simultaneous options to configure.
# TEST SUITE: Several new tests for MPI RMA operations.
# Several other minor bug fixes, memory leak fixes, and code cleanup.
A full list of changes is available using:
svn log -r7838:HEAD
... or at the following link:
Changes in 1.3.2
# OVERALL: MPICH2 now recognizes the OSX mach_absolute_time as a
native timer type.
# OVERALL: Performance improvements to MPI_Comm_split on large
# OVERALL: Several improvements to error returns capabilities in the
presence of faults.
# PM/PMI: Several fixes and improvements to Hydra's process binding
# PM/PMI: Upgrade the hwloc version to 1.1.1.
# PM/PMI: Allow users to sort node lists allocated by resource
managers in Hydra.
# PM/PMI: Improvements to signal handling. Now Hydra respects Ctrl-Z
signals and passes on the signal to the application.
# PM/PMI: Improvements to STDOUT/STDERR handling including improved
support for rank prepending on output. Improvements to STDIN
handling for applications being run in the background.
# PM/PMI: Split the bootstrap servers into "launchers" and "resource
managers", allowing the user to pick a different resource manager
from the launcher. For example, the user can now pick the "SLURM"
resource manager and "SSH" as the launcher.
# PM/PMI: The MPD process manager is deprecated.
# PM/PMI: The PLPA process binding library support is deprecated.
# WINDOWS: Adding support for gfortran and 64-bit gcc libs.
# Several other minor bug fixes, memory leak fixes, and code cleanup.
A full list of changes is available using:
svn log -r7457:HEAD
... or at the following link:
Changes in 1.3.1
# OVERALL: MPICH2 is now fully compliant with the CIFTS FTB standard
MPI events (based on the draft standard).
# OVERALL: Major improvements to RMA performance for long lists of
RMA operations.
# OVERALL: Performance improvements for Group_translate_ranks.
# COLLECTIVES: Collective algorithm selection thresholds can now be controlled
at runtime via environment variables.
# ROMIO: PVFS error codes are now mapped to MPI error codes.
# Several other minor bug fixes, memory leak fixes, and code cleanup.
A full list of changes is available using:
svn log -r7350:HEAD
... or at the following link:
Changes in 1.3
# OVERALL: Initial support for fine-grained threading in
# OVERALL: Support for Asynchronous Communication Progress.
# OVERALL: The ssm and shm channels have been removed.
# OVERALL: Checkpoint/restart support using BLCR.
# OVERALL: Improved tolerance to process and communication failures
when error handler is set to MPI_ERRORS_RETURN. If a communication
operation fails (e.g., due to a process failure) MPICH2 will return
an error, and further communication to that process is not
possible. However, communication with other processes will still
proceed normally. Note, however, that the behavior collective
operations on communicators containing the failed process is
undefined, and may give incorrect results or hang some processes.
# OVERALL: Experimental support for inter-library dependencies.
# PM/PMI: Hydra is now the default process management framework
replacing MPD.
# PM/PMI: Added dynamic process support for Hydra.
# PM/PMI: Added support for LSF, SGE and POE in Hydra.
# PM/PMI: Added support for CPU and memory/cache topology aware
process-core binding.
# DEBUGGER: Improved support and bug fixes in the Totalview support.
# Build system: Replaced F90/F90FLAGS by FC/FCFLAGS. F90/F90FLAGS are
not longer supported in the configure.
# Multi-compiler support: On systems where C compiler that is used to
build mpich2 libraries supports multiple weak symbols and multiple aliases,
the Fortran binding built in the mpich2 libraries can handle different
Fortran compilers (than the one used to build mpich2). Details in README.
# Several other minor bug fixes, memory leak fixes, and code cleanup.
A full list of changes is available using:
svn log -r5762:HEAD
... or at the following link:
Changes in 1.2.1
# OVERALL: Improved support for fine-grained multithreading.
# OVERALL: Improved integration with Valgrind for debugging builds of MPICH2.
# PM/PMI: Initial support for hwloc process-core binding library in
# PM/PMI: Updates to the PMI-2 code to match the PMI-2 API and
wire-protocol draft.
# Several other minor bug fixes, memory leak fixes, and code cleanup.
A full list of changes is available using:
svn log -r5425:HEAD
... or at the following link:
Changes in 1.2
# OVERALL: Support for MPI-2.2
# OVERALL: Several fixes to Nemesis/MX.
# WINDOWS: Performance improvements to Nemesis/windows.
# PM/PMI: Scalability and performance improvements to Hydra using
PMI-1.1 process-mapping features.
# PM/PMI: Support for process-binding for hyperthreading enabled
systems in Hydra.
# PM/PMI: Initial support for PBS as a resource management kernel in
# PM/PMI: PMI2 client code is now officially included in the release.
# TEST SUITE: Support to run the MPICH2 test suite through valgrind.
# Several other minor bug fixes, memory leak fixes, and code cleanup.
A full list of changes is available using:
svn log -r5025:HEAD
... or at the following link:
Changes in 1.1.1p1
- OVERALL: Fixed an invalid read in the dataloop code for zero count types.
- OVERALL: Fixed several bugs in ch3:nemesis:mx (tickets #744,#760;
also change r5126).
- BUILD SYSTEM: Several fixes for functionality broken in 1.1.1 release,
including MPICH2LIB_xFLAGS and extra libraries living in $LIBS instead of
$LDFLAGS. Also, '-lpthread' should no longer be duplicated in link lines.
- BUILD SYSTEM: MPICH2 shared libraries are now compatible with glibc versioned
symbols on Linux, such as those present in the MX shared libraries.
- BUILD SYSTEM: Minor tweaks to improve compilation under the nvcc CUDA
- PM/PMI: Fix mpd incompatibility with python2.3 introduced in mpich2-1.1.1.
- PM/PMI: Several fixes to hydra, including memory leak fixes and process
binding issues.
- TEST SUITE: Correct invalid arguments in the coll2 and coll3 tests.
- Several other minor bug fixes, memory leak fixes, and code cleanup. A full
list of changes is available using:
svn log -r5032:HEAD
... or at the following link:
Changes in 1.1.1
# OVERALL: Improved support for Boost MPI.
# PM/PMI: Significantly improved time taken by MPI_Init with Nemesis and MPD on
large numbers of processes.
# PM/PMI: Improved support for hybrid MPI-UPC program launching with
# PM/PMI: Improved support for process-core binding with Hydra.
# PM/PMI: Preliminary support for PMI-2. Currently supported only
with Hydra.
# Many other bug fixes, memory leak fixes and code cleanup. A full
list of changes is available using:
svn log -r4655:HEAD
... or at the following link:
Changes in 1.1
- OVERALL: Added MPI 2.1 support.
- OVERALL: Nemesis is now the default configuration channel with a
completely new TCP communication module.
- OVERALL: Windows support for nemesis.
- OVERALL: Added a new Myrinet MX network module for nemesis.
- OVERALL: Initial support for shared-memory aware collective
communication operations. Currently MPI_Bcast, MPI_Reduce, MPI_Allreduce,
and MPI_Scan.
- OVERALL: Improved handling of MPI Attributes.
- OVERALL: Support for BlueGene/P through the DCMF library (thanks to
IBM for the patch).
- OVERALL: Experimental support for fine-grained multithreading
- OVERALL: Added dynamic processes support for Nemesis.
- OVERALL: Added automatic as well as statically runtime configurable
receive timeout variation for MPD (thanks to OSU for the patch).
- OVERALL: Improved performance for MPI_Allgatherv, MPI_Gatherv, and MPI_Alltoall.
- PM/PMI: Initial support for the new Hydra process management
framework (current support is for ssh, rsh, fork and a preliminary
version of slurm).
- ROMIO: Added support for MPI_Type_create_resized and
MPI_Type_create_indexed_block datatypes in ROMIO.
- ROMIO: Optimized Lustre ADIO driver (thanks to Weikuan Yu for
initial work and Sun for further improvements).
- Many other bug fixes, memory leak fixes and code cleanup. A full
list of changes is available using:
svn log -r813:HEAD
... or at the following link:
Changes in 1.0.7
- OVERALL: Initial ROMIO device for BlueGene/P (the ADI device is also
added but is not configurable at this time).
- OVERALL: Major clean up for the propagation of user-defined and
other MPICH2 flags throughout the code.
- OVERALL: Support for STI Cell Broadband Engine.
- OVERALL: Added datatype free hooks to be used by devices
- OVERALL: Added device-specific timer support.
- OVERALL: make uninstall works cleanly now.
- ROMIO: Support to take hints from a config file
- ROMIO: more tests and bug fixes for nonblocking I/O
- PM/PMI: Added support to use PMI Clique functionality for process
managers that support it.
- PM/PMI: Added SLURM support to configure to make it transparent to
- PM/PMI: SMPD Singleton Init support.
- WINDOWS: Fortran 90 support added.
- SCTP: Added MPICH_SCTP_NAGLE_ON support.
- MPE: Updated MPE logging API so that it is thread-safe (through
global mutex).
- MPE: Added infrastructure to piggyback argument data to MPI states.
- DOCS: Documentation creation now works correctly for VPATH builds.
- Many other bug fixes, memory leak fixes and code cleanup. A full
list of changes is available using:
svn log -r100:HEAD
Changes in 1.0.6
- Updates to the ch3:nemesis channel including preliminary support for
thread safety.
- Preliminary support for dynamic loading of ch3 channels (sock, ssm,
shm). See the README file for details.
- Singleton init now works with the MPD process manager.
- Fixes in MPD related to MPI-2 connect-accept.
- Improved support for MPI-2 generalized requests that allows true
nonblocking I/O in ROMIO.
- MPE changes:
* Enabled thread-safe MPI logging through global mutex.
* Enhanced Jumpshot to be more thread friendly
+ added simple statistics in the Legend windows.
* Added backtrace support to MPE on Solaris and glibc based systems,
e.g. Linux. This improves the output error message from the
Collective/Datatype checking library.
* Fixed the CLOG2 format so it can be used in serial (non-MPI) logging.
- Performance improvements for derived datatypes (including packing
and communication) through in-built loop-unrolling and buffer
- Performance improvements for MPI_Gather when non-power-of-two
processes are used, and when a non-zero ranked root is performing the
- MPI_Comm_create works for intercommunicators.
- Enabled -O2 and equivalent compiler optimizations for supported
compilers by default (including GNU, Intel, Portland, Sun, Absoft,
- Many other bug fixes, memory leak fixes and code cleanup. A full
list of changes is available at
Changes in 1.0.5
- An SCTP channel has been added to the CH3 device. This was
implemented by Brad Penoff and Mike Tsai, Univ. of British Columbia.
Their group's webpage is located at .
- Bugs related to dynamic processes have been fixed.
- Performance-related fixes have been added to derived datatypes and
collective communication.
- Updates to the Nemesis channel
- Fixes to thread safety for the ch3:sock channel
- Many other bug fixes and code cleanup. A full list of changes is available
at .
Changes in 1.0.4
- For the ch3:sock channel, the default build of MPICH2 supports
thread safety. A separate build is not needed as before. However,
thread safety is enabled only if the user calls MPI_Init_thread with
MPI_THREAD_MULTIPLE. If not, no thread locks are called, so there
is no penalty.
- A new low-latency channel called Nemesis has been added. It can be
selected by specifying the option --with-device=ch3:nemesis.
Nemesis uses shared memory for intranode communication and various
networks for internode communication. Currently available networks
are TCP, GM and MX. Nemesis is still a work in progress. See the
README for more information about the channel.
- Support has been added for providing message queues to debuggers.
Configure with --enable-debuginfo to make this information available.
This is still a "beta" test version and has not been extensively tested.
- For systems with firewalls, the environment variable MPICH_PORT_RANGE can
be used to restrict the range of ports used by MPICH2. See the documentation
for more details.
- Withdrew obsolete modules, including the ib and rdma communication layers.
For Infiniband and MPICH2, please see
For other interconnects, please contact us at .
- Numerous bug fixes and code cleanup. A full list of changes is available
at .
- Numerous new tests in the MPICH2 test suite.
- For developers, the way in which information is passed between the top
level configure and configures in the device, process management, and
related modules has been cleaned up. See the comments at the beginning
of the top-level for details. This change makes it easier
to interface other modules to MPICH2.
Changes in 1.0.3
- There are major changes to the ch3 device implementation. Old and
unsupported channels (essm, rdma) have been removed. The
internal interface between ch3 and the channels has been improved to
similify the process of adding a new channel (sharing existing code
where possible) and to improve performance. Further changes in this
internal interface are expected.
- Numerous bug fixes and code cleanup
Creation of intercommunicators and intracommunicators
from the intercommunicators created with Spawn and Connect/Accept
The computation of the alignment and padding of items within
structures now handles additional cases, including systems
where the alignment an padding depends on the type of the first
item in the structure
MPD recognizes wdir info keyword
gforker's mpiexec supports -env and -genv arguments for controlling
which environment variables are delivered to created processes
- While not a bug, to aid in the use of memory trace packages, MPICH2
tries to free all allocated data no later than when MPI_Finalize
- Support for DESTDIR in install targets
- Enhancements to SMPD
- In order to support special compiler flags for users that may be
different from those used to build MPICH2, the environment variables
to specify the flags used in mpicc, mpif77, mpicxx, and mpif90
respectively. The flags CFLAGS, FFLAGS, CXXFLAGS, and F90FLAGS are
used in the building of MPICH2.
- Many enhancements to MPE
- Enhanced support for features and idiosyncracies of Fortran 77 and
Fortran 90 compilers, including gfortran, g95, and xlf
- Enhanced support for C++ compilers that do not fully support abstract
base classes
- Additional tests in the mpich2/tests/mpi
- New FAQ included (also available at
- Man pages for mpiexec and mpif90
- Enhancements for developers, including a more flexible and general
mechanism for inserting logging and information messages, controlable
with --mpich-dbg-xxx command line arguments or MPICH_DBG_XXX environment
- Note to developers:
This release contains many changes to the structure of the CH3
device implementation (in src/mpid/ch3), including signficant
reworking of the files (many files have been combined into fewer files
representing logical grouping of functions). The next release of
MPICH2 will contain even more significant changes to the device
structure as we introduce a new communication implementation.
Changes in 1.0.2
- Optimizations to the MPI-2 one-sided communication functions for the
sshm (scalable shared memory) channel when window memory is
allocated with MPI_Alloc_mem (for all three synchronization methods).
- Numerous bug fixes and code cleanup.
- Fixed memory leaks.
- Fixed shared library builds.
- Fixed performance problems with MPI_Type_create_subarray/darray
- The following changes have been made to MPE2:
- MPE2 now builds the MPI collective and datatype checking library
by default.
- SLOG-2 format has been upgraded to 2.0.6 which supports event drawables
and provides count of real drawables in preview drawables.
- new slog2 tools, slog2filter and slog2updater, which both are logfile
format convertors. slog2filter removes undesirable categories of
drawables as well as alters the slog2 file structure. slog2updater
is a slog2filter that reads in older logfile format, 2.0.5, and
writes out the latest format 2.0.6.
- The following changes have been made to MPD:
- Nearly all code has been replaced by new code that follows a more
object-oriented approach than before. This has not changed any
fundamental behavior or interfaces.
- There is info support in spawn and spawn_multiple for providing
parts of the environment for spawned processes such as search-path
and current working directory. See the Standard for the required
- mpdcheck has been enhanced to help users debug their cluster and
network configurations.
- CPickle has replaced marshal as the source module for dumps and loads.
- The mpigdb command has been replaced by mpiexec -gdb.
- Alternate interfaces can be used. See the Installer's Guide.
Changes in 1.0.1
- Copyright statements have been added to all code files, clearly identifying
that all code in the distribution is covered by the extremely flexible
copyright described in the COPYRIGHT file.
- The MPICH2 test suite (mpich2/test) can now be run against any MPI
implementation, not just MPICH2.
- The send and receive socket buffers sizes may now be changed by setting
MPICH_SOCKET_BUFFER_SIZE. Note: the operating system may impose a maximum
socket buffer size that prohibits MPICH2 from increasing the buffers to the
desire size. To raise the maximum allowable buffer size, please contact your
system administrator.
- Error handling throughout the MPI routines has been improved. The error
handling in some internal routines has been simplified as well, making the
routines easier to read.
- MPE (Jumpshot and CLOG logging) is now supported on Microsoft Windows.
- C applications built for Microsoft Windows may select the desired channels at
- A program not started with mpiexec may become an MPI program by calling
MPI_Init. It will have an MPI_COMM_WORLD of size one. It may then call
other MPI routines, including MPI_COMM_SPAWN, to become a truly parallel
program. At present, the use of MPI_COMM_SPAWN and MPI_COMM_SPAWN_MULTIPLE
by such a process is only supported by the MPD process manager.
- Memory leaks in communicator allocation and the C++ binding have been fixed.
- Following GNU guidelines, the parts of the install step that checked the
installation have been moved to an installcheck target. Much of the
installation now supports the DESTDIR prefix.
- Microsoft Visual Studio projects have been added to make it possible to build
x86-64 version
- Problems with compilers and linkers that do not support weak symbols, which
are used to support the PMPI profiling interface, have been corrected.
- Handling of Fortran 77 and Fortran 90 compilers has been improved, including
support for g95.
- The Fortran stdcall interface on Microsoft Windows now supports character*.
- A bug in the OS X implementation of poll() caused the sock channel to hang.
A workaround has been put in place.
- Problems with installation under OS/X are now detected and corrected.
(Install breaks libraries that are more than 10 seconds old!)
- The following changes have been made to MPD:
- Sending a SIGINT to mpiexec/mpdrun, such as by typing control-C, now causes
SIGINT to be sent to the processes within the job. Previously, SIGKILL was
sent to the processes, preventing applications from catching the signal
and performing their own signal processing.
- The process for merging output has been improved.
- A new option, -ifhn, has been added to the machine file, allowing the user
to select the destination interface to be used for TCP communication. See
the User's Manual for details.
- The user may now select, via the "-s" option to mpiexec/mpdrun, which
processes receive input through stdin. stdin is immediately closed for all
processes not in set receiving input. This prevents processes not in the
set from hanging should they attempt to read from stdin.
- The MPICH2 Installer's Guide now contains an appendix on troubleshooting
problems with MPD.
- The following changes have been made to SMPD:
- On Windows machines, passwordless authentication (via SSPI) can now be used
to start processes on machines within a domain. This feature is a recent
addition, and should be considered experimental.
- On Windows machines, the -localroot option was added to mpiexec, allowing
processes on the local machines to perform GUI operations on the local
- On Windows machines, network drive mapping is now supported via the -map
option to mpiexec.
- Three new GUI tools have been added for Microsoft Windows. These tools are
wrappers to the command line tools, mpiexec.exe and smpd.exe. wmpiexec
allows the user to run a job much in the way they with mpiexec. wmpiconfig
provides a means of setting various global options to the SMPD process
manager environment. wmpiregister encrypts the user's credentials and
saves them to the Windows Registry.
- The following changes have been made to MPE2:
- MPE2 no longer attempt to compile or link code during 'make install' to
validate the installation. Instead, 'make installcheck' may now be used to
verify that the MPE installation.
- MPE2 now supports DESTDIR.
- The sock channel now has preliminary support for MPI_THREAD_SERIALIZED and
MPI_THREAD_MULTIPLE on both UNIX and Microsoft Windows. We have performed
rudimentary testing; and while overall the results were very positive, known
issues do exist. ROMIO in particular experiences hangs in several places.
We plan to correct that in the next release. As always, please report any
difficulties you encounter.
- Another channel capable of communicating with both over sockets and shared
memory has been added. Unlike the ssm channel which waits for new data to
arrive by continuously polling the system in a busy loop, the essm channel
waits by blocking on an operating system event object. This channel is
experimental, and is only available for Microsoft Windows.
- The topology routines have been modified to allow the device to override the
default implementation. This allows the device to export knowledge of the
underlying physical topology to the MPI routines (Dims_create and the
reorder == true cases in Cart_create and Graph_create).
- New memory allocation macros, MPIU_CHK[PL]MEM_*(), have been added to help
prevent memory leaks. See mpich2/src/include/mpir_mem.h.
- New error reporting macros, MPIU_ERR_*, have been added to simplify the error
handling throughout the code, making the code easier to read. See
- Interprocess communication using the Sock interface (sock and ssm channels)
may now be bound to a particular destination interface using the environment
variable MPICH_INTERFACE_HOSTNAME. The variable needs to be set for each
process for which the destination interface is not the default interface.
(Other mechanisms for destination interface selection will be provided in
future releases.) Both MPD and SMPD provide a more simplistic mechanism for
specifying the interface. See the user documentation.
- Too many bug fixes to describe. Much thanks goes to the users who reported
bugs. Their patience and understanding as we attempted to recreate the
problems and solve them is greatly appreciated.
Changes in 1.0
- MPICH2 now works on Solaris.
- The User's Guide has been expanded considerably. The Installation Guide has
been expanded some as well.
- MPI_COMM_JOIN has been implemented; although like the other dynamic process
routines, it is only supported by the Sock channel.
- MPI_COMM_CONNECT and MPI_COMM_ACCEPT are now allowed to connect with remote
process to which they are already connected.
- Shared libraries can now be built (and used) on IA32 Linux with the GNU
compilers (--enable-sharedlibs=gcc), and on Solaris with the native Sun
Workshop compilers (--enable-sharedlibs=solaris). They may also work on
other operating systems with GCC, but that has not been tested. Previous
restrictions disallowing C++ and Fortran bindings when building shared
libraries have been removed.
- The dataloop and datatype contents code has been improved to address
alignment issues on all platforms.
- A bug in the datatype code, which handled zero block length cases
incorrectly, has been fixed.
- An segmentation fault in the datatype memory management, resulting from
freeing memory twice, has been fixed.
- The following changes were made to the MPD process manager:
- MPI_SPAWN_MULTIPLE now works with MPD.
- The arguments to the 'mpiexec' command supplied by the MPD have changed.
First, the -default option has been removed. Second, more flexible ways to
pass environment variables have been added.
- The commands 'mpdcheck' and 'testconfig' have been to installations using
MPD. These commands test the setup of the machines on which you wish to
run MPICH2 jobs. They help to identify misconfiguration, firewall issues,
and other communication problems.
- Support for MPI_APPNUM and MPI_UNIVERSE_SIZE has been added to the Simple
implementation of PMI and the MPD process manager.
- In general, error detection and recovery in MPD has improved.
- A new process manager, gforker, is now available. Like the forker process
manager, gforker spawns processes using fork(), and thus is quite useful on
SMPs machines. However, unlike forker, gforker supports all of the features
of a standard mpiexec, plus some. Therefore, It should be used in place of
the previous forker process manager, which is now deprecated.
- The following changes were made to ROMIO:
- The amount of duplicated ROMIO code in the close, resize, preallocate,
read, write, asynchronous I/O, and sync routines has been substantially
- A bug in flattening code, triggered by nested datatypes, has been fixed.
- Some small memory leaks have been fixed.
- The error handling has been abstracted allowing different MPI
implementations to handle and report error conditions in their own way.
Using this abstraction, the error handling routines have been made
consistent with rest of MPICH2.
- AIO support has been cleaned up and unified. It now works correctly on
Linux, and is properly detected on old versions of AIX.
- A bug in MPI_File_seek code, and underlying support code, has been fixed.
- Support for PVFS2 has improved.
- Several dead file systems have been removed. Others, including HFS, SFS,
PIOFS, and Paragon, have been deprecated.
- MPE and CLOG have been updated to version 2.1. For more details, please see
- New macros for memory management were added to support function local
allocations (alloca), to rollback pending allocations when error conditions
are detected to avoid memory leaks, and to improve the conciseness of code
performing memory allocations.
- New error handling macros were added to make internal error handling code
more concise.
Changes in 0.971
- Code restricted by copyrights less flexible than the one described in the
COPYRIGHT file has been removed.
- Installation and User Guides have been added.
- The SMPD PMI Wire Protocol Reference Manual has been updated.
- To eliminate portability problems, common blocks in mpif.h that spanned
multiple lines were broken up into multiple common blocks each described on a
single line.
- A new command, mpich2version, was added to allow the user to obtain
information about the MPICH2 installation. This command is currently a
simple shell script. We anticipate that the mpich2version command will
eventually provide additional information such as the patches applied and the
date of the release.
- The following changes were made to MPD2:
- Support was added for MPI's "singleton init", in which a single
process started in the normal way (i.e., not by mpiexec or mpirun)
becomes an MPI process with an MPI_COMM_WORLD of size one by
calling MPI_Init. After this the process can call other MPI
functions, including MPI_Comm_spawn.
- The format for some of the arguments to mpiexec have changed,
especially for passing environment variables to MPI processes.
- In addition to miscellaneous hardening, better error checking and
messages have been added.
- The install process has been improved. In particular, configure
has been updated to check for a working install program and supply
it's own installation script ( if necessary.
- A new program, mpdcheck, has been added to help diagnose machine
configurations that might be erroneous or at least confusing to
- Runtime version checking has been added to insure that the Simple
implementation of PMI linked into the application and the MPD
process manager being used to run that application are compatible.
- Minor improvements have been made to mpdboot.
- Support for the (now deprecated) BNR interface has been added to
allow MPICH1 programs to also be run via MPD2.
- Shared libraries are now supported on Linux systems using the GNU compilers
with the caveat that C++ support must be disabled (--disable-cxx).
- The CH3 interface and device now provide a mechanism for using RDMA (remote
direct memory access) to transfer data between processes.
- Logging capabilities for MPI and internal routines have been readded. See
the documentation in doc/logging for details.
- A "meminit" option was added to --enable-g to force all bytes associated with
a structure or union to be initialized prior to use. This prevents programs
like Valgrind from complaining about uninitialized accesses.
- The dist-with-version and snap targets in the top-level now
properly produce mpich2-<ver>/maint/Version instead of mpich2-<ver>/Version.
In addition, they now properly update the VERSION variable in
without clobbering the sed line that performs the update.
- The dist and snap targets in the top-level now both use the
dist-with-version target to avoid inconsistencies.
- The following changes were made to simplemake:
- The environment variables DEBUG, DEBUG_DIRS, and DEBUG_CONFDIR can now be
used to control debugging output.
- Many fixes were made to make simplemake so that it would run cleanly with
perl -w.
- Installation of *all* files from a directory is now possible (example,
installing all of the man pages).
- The clean targets now remove the cache files produced by newer versions of
- For files that are created by configure, the determination of the
location of that configure has been improved, so that make of those
files (e.g., make Makefile) is more likely to work. There is still
more to do here.
- Short loops over subdirectories are now unrolled.
- The maintainerclean target has been renamed to maintainer-clean to match
GNU guidelines.
- The distclean and maintainer-clean targets have been improved.
- An option was added to perform one ar command per directory instead of one
per file when creating the profiling version of routines (needed only for
systems that do not support weak symbols).
Changes in 0.97
- MPI-2 one-sided communication has been implemented in the CH3 device.
- mpigdb works as a simple parallel debugger for MPI programs started
with mpd. New since MPICH1 is the ability to attach to running
parallel programs. See the README in mpich2/src/pm/mpd for details.
- MPI_Type_create_darray() and MPI_Type_create_subarray() implemented including
the right contents and envelope data.
- ROMIO flattening code now supports subarray and darray combiners.
- Improve scalability and performance of some ROMIO PVFS and PVFS2 routines
- An error message string parameter was added to MPID_Abort(). If the
parameter is non-NULL this string will be used as the message with the abort
output. Otherwise, the output message will be base on the error message
associated with the mpi_errno parameter.
- MPID_Segment_init() now takes an additional boolean parameter that specifies
if the segment processing code is to produce/consume homogeneous (FALSE) or
heterogeneous (TRUE) data.
- The definitions of MPID_VCR and MPID_VCRT are now defined by the device.
- The semantics of MPID_Progress_{Start,Wait,End}() have changed. A typical
blocking progress loop now looks like the following.
if (req->cc != 0)
MPID_Progress_state progress_state;
while (req->cc != 0)
mpi_errno = MPID_Progress_wait(&progress_state);
if (mpi_errno != MPI_SUCCESS)
goto fn_fail;
NOTE: each of these routines now takes a single parameter, a pointer to a
thread local state variable.
- The CH3 device and interface have been modified to better support
writers will notice the following. (This is still a work in progress. See
the note below.)
- The introduction of a process group object (MPIDI_PG_t) and a new
set of routines to manipulate that object.
- The renaming of the MPIDI_VC object to MPIDI_VC_t to make it more
consistent with the naming of other objects in the device.
- The process group information in the MPIDI_VC_t moved from the channel
specific portion to the device layer.
- MPIDI_CH3_Connection_terminate() was added to the CH3 interface to allow
the channel to properly shutdown a connection before the device deletes all
associated data structures.
- A new upcall routine, MPIDI_CH3_Handle_connection(), was added to allow the
device to notify the device when a connection related event has completed.
A present the only event is MPIDI_CH3_VC_EVENT_TERMINATED, which notify the
device that the underlying connection associated with a VC has been
properly shutdown. For every call to MPIDI_CH3_Connection_terminate() that
the device makes, the channel must make a corresponding upcall to
MPIDI_CH3_Handle_connection(). MPID_Finalize() will likely hang if this
rule is not followed.
- MPIDI_CH3_Get_parent_port() was added to provide MPID_Init() with the port
name of the the parent (spawner). This port name is used by MPID_Init()
and MPID_Comm_connect() to create an intercommunicator between the parent
(spawner) and child (spawnee). Eventually, MPID_Comm_spawn_multiple() will
be update to perform the reverse logic; however, the logic is presently
still in the sock channel.
Note: the changes noted are relatively fresh and are the beginning to a set
of future changes. The goal is to minimize the amount of code required by a
channel to support MPI dynamic process functionality. As such, portions of
the device will change dramatically in a future release. A few more changes
to the CH3 interface are also quite likely.
- MPIDI_CH3_{iRead,iWrite}() have been removed from the CH3 interface.
MPIDI_CH3U_Handle_recv_pkt() now returns a receive request with a populated
iovec to receive data associated with the request.
MPIDU_CH3U_Handle_{recv,send}_req() reload the iovec in the request and
return and set the complete argument to TRUE if more data is to read or
written. If data transfer for the request is complete, the complete argument
must be set to FALSE.
Changes in 0.96p2
The shm and ssm channels have been added back into the distribution.
Officially, these channels are supported only on x86 platforms using the gcc
compiler. The necessary assembly instructions to guarantee proper ordering of
memory operations are lacking for other platforms and compilers. That said, we
have seen a high success rate when testing these channels on unsupported
This patch release also includes a new unsupported channel. The scalable
shared memory, or sshm, channel is similar to the shm channel except that it
allocates shared memory communication queues only when necessary instead of
preallocating N-squared queues.
Changes in 0.96p1
This patch release fixes a problem with building MPICH2 on Microsoft Windows
platforms. It also corrects a serious bug in the poll implementation of the
Sock interface.
Changes in 0.96
The 0.96 distribution is largely a bug fix release. In addition to the many
bug fixes, major improvements have been made to the code that supports the
dynamic process management routines (MPI_Comm_{connect,accept,spawn,...}()).
Additional changes are still required to support MPI_Comm_disconnect().
We also added an experimental (and thus completely unsupported) rdma device.
The internal interface is similar to the CH3 interface except that it contains
a couple of extra routines to inform the device about data transfers using the
rendezvous protocol. The channel can use this extra information to pin memory
and perform a zero-copy transfer. If all goes well, the results will be rolled
back into the CH3 device.
Due to last minute difficulties, this release does not contain the shm or ssm
channels. These channels will be included in a subsequent patch release.
Changes in 0.94
Active target one-sided communication is now available for the ch3:sock
channel. This new functionality has undergone some correctness testing but has
not been optimized in terms of performance. Future release will include
performance enhancements, passive target communication, and availability in
channels other than just ch3:sock.
The shared memory channel (ch3:shm), which performs communication using shared
memory on a single machine, is now complete and has been extensively tested.
At present, this channel only supports IA32 based machines (excluding the
Pentium Pro which has a memory ordering bug). In addition, this channel must
be compiled with gcc. Future releases with support additional architectures
and compilers.
A new channel has been added that performs inter-node communication using
sockets (TCP/IP) and intra-node communication using shared memory. This
channel, ch3:ssm, is ideal for clusters of SMPs. Like the shared memory
channel (ch3:shm), this channel only supports IA32 based machines and must be
compiled with gcc. In future releases, the ch3:ssm channel will support
additional architectures and compilers.
The two channels that perform commutation using shared memory, ch3:shm and
ch3:ssm, now support the allocation of shared memory using both the POSIX and
System V interfaces. The POSIX interface will be used if available; otherwise,
the System V interface is used.
In the interest of increasing portability, many enhancements have been made to
both the code and the configure scripts.
And, as always, many bugs have been fixed :-).
The parameters to MPID_Abort() have changed. MPID_Abort() now takes a pointer
to communicator object, an MPI error code, and an exit code.
MPIDI_CH3_Progress() has been split into two functions:
MPIDI_CH3_Progress_wait() and MPIDI_CH3_Progress_test().
Changes in 0.93
Version 0.93 has undergone extensive changes to provide better error reporting.
Part of these changes involved modifications to the ADI3 and CH3 interfaces.
The following routines now return MPI error codes:
Of special note are MPID_Progress_test(), MPID_Progress_wait() and
MPIDI_CH3_Progress() which previously returned an integer value indicating if
one or more requests had completed. They no longer return this value and
instead return an MPI error code (also an integer). The implication being that
while the semantics changed, the type signatures did not.
The function used to create error codes, MPIR_Err_create_code(), has also
changed. It now takes additional parameters, allowing it create a stack of
errors and making it possible for the reporting function to indicate in which
function and on which line the error occurred. It also allows an error to be
designated as fatal or recoverable. Fatal errors always result in program
termination regardless of the error handler installed by the application.
A RDMA channel has been added and includes communication methods for shared
memory and shmem. This is recent development and the RDMA interface is still
in flux.