1
1

NEWS: Sync with v2.0.x and v1.10 releases

Pull in content from v1.10 and v2.0.x branches.

[skip ci]
bot:notest

Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
Этот коммит содержится в:
Jeff Squyres 2017-03-14 11:28:17 -07:00
родитель 1709febdea
Коммит 60ca372d60

343
NEWS
Просмотреть файл

@ -8,7 +8,7 @@ Copyright (c) 2004-2008 High Performance Computing Center Stuttgart,
University of Stuttgart. All rights reserved.
Copyright (c) 2004-2006 The Regents of the University of California.
All rights reserved.
Copyright (c) 2006-2016 Cisco Systems, Inc. All rights reserved.
Copyright (c) 2006-2017 Cisco Systems, Inc. All rights reserved.
Copyright (c) 2006 Voltaire, Inc. All rights reserved.
Copyright (c) 2006 Sun Microsystems, Inc. All rights reserved.
Use is subject to license terms.
@ -67,19 +67,184 @@ Master (not on release branches yet)
Please consider Score-P (score-p.org) as an external replacement.
2.0.0 -- DATE
------
2.0.2 -- 26 January 2017
-------------------------
Bug fixes/minor improvements:
- Fix a problem with MPI_FILE_WRITE_SHARED when using MPI_MODE_APPEND and
Open MPI's native MPI-IO implementation. Thanks to Nicolas Joly for
reporting.
- Fix a typo in the MPI_WIN_GET_NAME man page. Thanks to Nicolas Joly
for reporting.
- Fix a race condition with ORTE's session directory setup. Thanks to
@tbj900 for reporting this issue.
- Fix a deadlock issue arising from Open MPI's approach to catching calls to
munmap. Thanks to Paul Hargrove for reporting and helping to analyze this
problem.
- Fix a problem with PPC atomics which caused make check to fail unless builtin
atomics configure option was enabled. Thanks to Orion Poplawski for reporting.
- Fix a problem with use of x86_64 cpuid instruction which led to segmentation
faults when Open MPI was configured with -O3 optimization. Thanks to Mark
Santcroos for reporting this problem.
- Fix a problem when using built in atomics configure options on PPC platforms
when building 32 bit applications. Thanks to Paul Hargrove for reporting.
- Fix a problem with building Open MPI against an external hwloc installation.
Thanks to Orion Poplawski for reporting this issue.
- Remove use of DATE in the message queue version string reported to debuggers to
insure bit-wise reproducibility of binaries. Thanks to Alastair McKinstry
for help in fixing this problem.
- Fix a problem with early exit of a MPI process without calling MPI_FINALIZE
or MPI_ABORT that could lead to job hangs. Thanks to Christof Koehler for
reporting.
- Fix a problem with forwarding of SIGTERM signal from mpirun to MPI processes
in a job. Thanks to Noel Rycroft for reporting this problem
- Plug some memory leaks in MPI_WIN_FREE discovered using Valgrind. Thanks
to Joseph Schuchart for reporting.
- Fix a problems MPI_NEIGHOR_ALLTOALL when using a communicator with an empty topology
graph. Thanks to Daniel Ibanez for reporting.
- Fix a typo in a PMIx component help file. Thanks to @njoly for reporting this.
- Fix a problem with Valgrind false positives when using Open MPI's internal memchecker.
Thanks to Yvan Fournier for reporting.
- Fix a problem with MPI_FILE_DELETE returning MPI_SUCCESS when
deleting a non-existent file. Thanks to Wei-keng Liao for reporting.
- Fix a problem with MPI_IMPROBE that could lead to hangs in subsequent MPI
point to point or collective calls. Thanks to Chris Pattison for reporting.
- Fix a problem when configure Open MPI for powerpc with --enable-mpi-cxx
enabled. Thanks to Alastair McKinstry for reporting.
- Fix a problem using MPI_IALLTOALL with MPI_IN_PLACE argument. Thanks to
Chris Ward for reporting.
- Fix a problem using MPI_RACCUMULATE with the Portals4 transport. Thanks to
@PDeveze for reporting.
- Fix an issue with static linking and duplicate symbols arising from PMIx
Slurm components. Thanks to Limin Gu for reporting.
- Fix a problem when using MPI dynamic memory windows. Thanks to
Christoph Niethammer for reporting.
- Fix a problem with Open MPI's pkgconfig files. Thanks to Alastair McKinstry
for reporting.
- Fix a problem with MPI_IREDUCE when the same buffer is supplied for the
send and recv buffer arguments. Thanks to Valentin Petrov for reporting.
- Fix a problem with atomic operations on PowerPC. Thanks to Paul
Hargrove for reporting.
Known issues (to be addressed in v2.0.3):
- See the list of fixes slated for v2.0.3 here:
https://github.com/open-mpi/ompi/milestone/23
2.0.1 -- 2 September 2016
-----------------------
Bug fixes/minor improvements:
- Short message latency and message rate performance improvements for
all transports.
- Fix shared memory performance when using RDMA-capable networks.
Thanks to Tetsuya Mishima and Christoph Niethammer for reporting.
- Fix bandwith performance degredation in the yalla (MXM) PML. Thanks
to Andreas Kempf for reporting the issue.
- Fix OpenSHMEM crash when running on non-Mellanox MXM-based networks.
Thanks to Debendra Das for reporting the issue.
- Fix a crash occuring after repeated calls to MPI_FILE_SET_VIEW with
predefined datatypes. Thanks to Eric Chamberland and Matthew
Knepley for reporting and helping chase down this issue.
- Fix stdin propagation to MPI processes. Thanks to Jingchao Zhang
for reporting the issue.
- Fix various runtime and portability issues by updating the PMIx
internal component to v1.1.5.
- Fix process startup failures on Intel MIC platforms due to very
large entries in /proc/mounts.
- Fix a problem with use of relative path for specifing executables to
mpirun/oshrun. Thanks to David Schneider for reporting.
- Various improvements when running over portals-based networks.
- Fix thread-based race conditions with GNI-based networks.
- Fix a problem with MPI_FILE_CLOSE and MPI_FILE_SET_SIZE. Thanks
to Cihan Altinay for reporting.
- Remove all use of rand(3) from within Open MPI so as not to perturb
applications use of it. Thanks to Matias Cabral and Noel Rycroft
for reporting.
- Fix crash in MPI_COMM_SPAWN.
- Fix types for MPI_UNWEIGHTED and MPI_WEIGHTS_EMPTY. Thanks to
Lisandro Dalcin for reporting.
- Correctly report the name of MPI_INTEGER16.
- Add some missing MPI constants to the Fortran bindings.
- Fixed compile error when configuring Open MPI with --enable-timing.
- Correctly set the shared library version of libompitrace.so. Thanks
to Alastair McKinstry for reporting.
- Fix errors in the MPI_RPUT, MPI_RGET, MPI_RACCUMULATE, and
MPI_RGET_ACCUMULATE Fortran bindings. Thanks to Alfio Lazzaro and
Joost VandeVondele for tracking this down.
- Fix problems with use of derived datatypes in non-blocking
collectives. Thanks to Yuki Matsumoto for reporting.
- Fix problems with OpenSHMEM header files when using CMake. Thanks to
Paul Kapinos for reporting the issue.
- Fix problem with use use of non-zero lower bound datatypes in
collectives. Thanks to Hristo Iliev for reporting.
- Fix a problem with memory allocation within MPI_GROUP_INTERSECTION.
Thanks to Lisandro Dalcin for reporting.
- Fix an issue with MPI_ALLGATHER for communicators that don't consist
of two ranks. Thanks to David Love for reporting.
- Various fixes for collectives when used with esoteric MPI datatypes.
- Fixed corner cases of handling DARRAY and HINDEXED_BLOCK datatypes.
- Fix a problem with filesystem type check for OpenBSD.
Thanks to Paul Hargrove for reporting.
- Fix some debug input within Open MPI internal functions. Thanks to
Durga Choudhury for reporting.
- Fix a typo in a configury help message. Thanks to Paul Hargrove for
reporting.
- Correctly support MPI_IN_PLACE in MPI_[I]ALLTOALL[V|W] and
MPI_[I]EXSCAN.
- Fix alignment issues on SPARC platforms.
Known issues (to be addressed in v2.0.2):
- See the list of fixes slated for v2.0.2 here:
https://github.com/open-mpi/ompi/milestone/20, and
https://github.com/open-mpi/ompi-release/milestone/19
(note that the "ompi-release" Github repo will be folded/absorbed
into the "ompi" Github repo at some point in the future)
2.0.0 -- 12 July 2016
---------------------
**********************************************************************
* Open MPI is now fully MPI-3.1 compliant
**********************************************************************
- Enhancements to reduce the memory footprint for jobs at scale. A
new MCA parameter, "mpi_add_procs_cutoff", is available to set the
threshold for using this feature.
Major new features:
- Many enhancements to MPI RMA. Open MPI now maps MPI RMA operations
on to native RMA operations for those networks which support this
capability.
- Greatly improved support for MPI_THREAD_MULTIPLE (when configured
with --enable-mpi-thread-multiple).
- Enhancements to reduce the memory footprint for jobs at scale. A
new MCA parameter, "mpi_add_procs_cutoff", is available to set the
threshold for using this feature.
- Completely revamped support for memory registration hooks when using
OS-bypass network transports.
- Significant OMPIO performance improvements and many bug fixes.
- Add support for PMIx - Process Management Interface for Exascale.
Version 1.1.2 of PMIx is included internally in this release.
- Add support for PLFS file systems in Open MPI I/O.
- Add support for UCX transport.
- Simplify build process for Cray XC systems. Add support for
using native SLURM.
- Add a --tune mpirun command line option to simplify setting many
environment variables and MCA parameters.
- Add a new MCA parameter "orte_default_dash_host" to offer an analogue
to the existing "orte_default_hostfile" MCA parameter.
- Add the ability to specify the number of desired slots in the mpirun
--host option.
Changes in behavior compared to prior versions:
- In environments where mpirun cannot automatically determine the
number of slots available (e.g., when using a hostfile that does not
specify "slots", or when using --host without specifying a ":N"
suffix to hostnames), mpirun now requires the use of "-np N" to
specify how many MPI processes to launch.
- The MPI C++ bindings -- which were removed from the MPI standard in
v3.0 -- are no longer built by default and will be removed in some
future version of Open MPI. Use the --enable-mpi-cxx-bindings
@ -87,37 +252,26 @@ Master (not on release branches yet)
- ompi_info now shows all components, even if they do not have MCA
parameters. The prettyprint output now separates groups with a
dashed line.
- Add support for PMIx - Process Management Interface for Exascale.
Version 1.1.2 of PMIx is included internally in this release.
- Add support for PLFS file systems in Open MPI I/O.
- Add support for UCX transport.
- Improved support for MPI_THREAD_MULTIPLE (when configured with
--enable-mpi-thread-multiple).
- Simplify build process for Cray XC systems. Add support for
using native SLURM.
- Updated internal/embedded copies of third-part software:
- Update the internal copy of ROMIO to that which shipped in MPICH
3.1.4.
- Update internal copy of libevent to v2.0.22.
- Update internal copy of hwloc to v1.11.2.
- Add a --tune mpirun command line option to simplify setting many
environment variables and MCA parameters.
- Add a new MCA parameter -
"opal_common_verbs_want_fork_support". This replaces the
"btl_openib_want_fork_support" parameter.
- Add a new MCA parameter "orte_default_dash_host" to offer an analogue
to the existing "orte_default_hostfile" MCA parameter.
- Add --with-platform-patches-dir configure option.
- Add --with-pmi-libdir configure option for environments that install
PMI libs in a non-default location.
- Add the ability to specify the number of desired slots in the mpirun
--host option.
- In environments where mpirun cannot automatically determine the
number of slots available (e.g., when using a hostfile that does not
specify "slots", or when using --host without specifying a ":N"
suffix to hostnames), mpirun now requires the use of "-np N" to
specify how many MPI processes to launch.
- Removed some legacy support:
- OMPIO is now the default implementation of parallel I/O, with the
exception for Lustre parallel filesystems (where ROMIO is still the
default). The default selection of OMPI vs. ROMIO can be controlled
via the "--mca io ompi|romio" command line switch to mpirun.
- Per Open MPI's versioning scheme (see the README), increasing the
major version number to 2 indicates that this version is not
ABI-compatible with prior versions of Open MPI. You will need to
recompile MPI and OpenSHMEM applications to work with this version
of Open MPI.
- Removed checkpoint/restart code due to loss of maintainer. :-(
- Change the behavior for handling certain signals when using PSM and
PSM2 libraries. Previously, the PSM and PSM2 libraries would trap
certain signals in order to generate tracebacks. The mechanism was
found to cause issues with Open MPI's own error reporting mechanism.
If not already set, Open MPI now sets the IPATH_NO_BACKTRACE and
HFI_NO_BACKTRACE environment variables to disable PSM/PSM2's
handling these signals.
Removed legacy support:
- Removed support for OS X Leopard.
- Removed support for Cray XT systems.
- Removed VampirTrace.
@ -125,11 +279,88 @@ Master (not on release branches yet)
- Removed legacy collective module:ML.
- Removed support for Alpha processors.
- Removed --enable-mpi-profiling configure option.
- Removed checkpoint/restart code due to loss of maintainer. :-(
Known issues (to be addressed in v2.0.1):
- See the list of fixes slated for v2.0.1 here:
https://github.com/open-mpi/ompi/milestone/16, and
https://github.com/open-mpi/ompi-release/milestone/16
(note that the "ompi-release" Github repo will be folded/absorbed
into the "ompi" Github repo at some point in the future)
- ompi-release#986: Fix data size counter for large ops with fcoll/static
- ompi-release#987: Fix OMPIO performance on Lustre
- ompi-release#1013: Fix potential inconsistency in btl/openib default settings
- ompi-release#1014: Do not return MPI_ERR_PENDING from collectives
- ompi-release#1056: Remove dead profile code from oshmem
- ompi-release#1081: Fix MPI_IN_PLACE checking for IALLTOALL{V|W}
- ompi-release#1081: Fix memchecker in MPI_IALLTOALLW
- ompi-release#1081: Support MPI_IN_PLACE in MPI_(I)ALLTOALLW and MPI_(I)EXSCAN
- ompi-release#1107: Allow future PMIx support for RM spawn limits
- ompi-release#1108: Fix sparse group process reference counting
- ompi-release#1109: If specified to be oversubcribed, disable binding
- ompi-release#1122: Allow NULL arrays for empty datatypes
- ompi-release#1123: Fix signed vs. unsigned compiler warnings
- ompi-release#1123: Make max hostname length uniform across code base
- ompi-release#1127: Fix MPI_Compare_and_swap
- ompi-release#1127: Fix MPI_Win_lock when used with MPI_Win_fence
- ompi-release#1132: Fix typo in help message for --enable-mca-no-build
- ompi-release#1154: Ensure pairwise coll algorithms disqualify themselves properly
- ompi-release#1165: Fix typos in debugging/verbose message output
- ompi-release#1178: Fix ROMIO filesystem check on OpenBSD 5.7
- ompi-release#1197: Fix Fortran pthread configure check
- ompi-release#1205: Allow using external PMIx 1.1.4 and 2.0
- ompi-release#1215: Fix configure to support the NAG Fortran compiler
- ompi-release#1220: Fix combiner args for MPI_HINDEXED_BLOCK
- ompi-release#1225: Fix combiner args for MPI_DARRAY
- ompi-release#1226: Disable old memory hooks with recent gcc versions
- ompi-release#1231: Fix new "patcher" support for some XLC platforms
- ompi-release#1244: Fix Java error handling
- ompi-release#1250: Ensure TCP is not selected for RDMA operations
- ompi-release#1252: Fix verbose output in coll selection
- ompi-release#1253: Set a default name for user-defined MPI_Op
- ompi-release#1254: Add count==0 checks in some non-blocking colls
- ompi-release#1258: Fix "make distclean" when using external pmix/hwloc/libevent
- ompi-release#1260: Clean up/uniform mca/coll/base memory management
- ompi-release#1261: Remove "patcher" warning message for static builds
- ompi-release#1263: Fix IO MPI_Request for 0-size read/write
- ompi-release#1264: Add blocking fence for SLURM operations
Bug fixes / minor enhancements:
- Updated internal/embedded copies of third-party software:
- Update the internal copy of ROMIO to that which shipped in MPICH
3.1.4.
- Update internal copy of libevent to v2.0.22.
- Update internal copy of hwloc to v1.11.2.
- Notable new MCA parameters:
- opal_progress_lp_call_ration: Control how often low-priority
callbacks are made during Open MPI's main progress loop.
- opal_common_verbs_want_fork_support: This replaces the
btl_openib_want_fork_support parameter.
- Add --with-platform-patches-dir configure option.
- Add --with-pmi-libdir configure option for environments that install
PMI libs in a non-default location.
- Various configure-related compatibility updates for newer versions
of libibverbs and OFED.
- Numerous fixes/improvements to orte-dvm. Special thanks to Mark
Santcroos for his help.
- Fix a problem with timer code on ia32 platforms. Thanks to
Paul Hargrove for reporting this and providing a patch.
- Fix a problem with use of a 64 bit atomic counter. Thanks to
Paul Hargrove for reporting.
- Fix a problem with singleton job launching. Thanks to Lisandro
Dalcin for reporting.
- Fix a problem with use of MPI_UNDEFINED with MPI_COMM_SPLIT_TYPE.
Thanks to Lisandro Dalcin for reporting.
- Silence a compiler warning in PSM MTL. Thanks to Adrian Reber for
reporting this.
- Properly detect Intel TrueScale and OmniPath devices in the ACTIVE
state. Thanks to Durga Choudhury for reporting the issue.
- Fix detection and use of Solaris Studio 12.5 (beta) compilers.
Thanks to Paul Hargrove for reporting and debugging.
- Fix various small memory leaks.
- Allow NULL arrays when creating empty MPI datatypes.
- Replace use of alloca with malloc for certain datatype creation
functions. Thanks to Bogdan Sataric for reporting this.
- Fix use of MPI_LB and MPI_UB in creation of of certain MPI datatypes.
@ -138,6 +369,8 @@ Master (not on release branches yet)
Schnetter for reporting and fixing.
- Improve hcoll library detection in configure. Thanks to David
Shrader and Ake Sandgren for reporting this.
- Miscellaneous minor bug fixes in the hcoll component.
- Miscellaneous minor bug fixes in the ugni component.
- Fix problems with XRC detection in OFED 3.12 and older releases.
Thanks to Paul Hargrove for his analysis of this problem.
- Update (non-standard/experimental) Java MPI interfaces to support
@ -165,6 +398,8 @@ Master (not on release branches yet)
reporting this.
- Fix a problem in neighborhood collectives. Thanks to Lisandro
Dalcin for reporting.
- Fix MPI_IREDUCE_SCATTER_BLOCK for a one-process communicator. Thanks
to Lisandro Dalcin for reporting.
- Add (Open MPI-specific) additional flavors to MPI_COMM_SPLIT_TYPE.
See MPI_Comm_split_type(3) for details. Thanks to Nick Andersen for
supplying this enhancement.
@ -216,6 +451,19 @@ Master (not on release branches yet)
Alastair McKinstry for reporting.
1.10.6 - 17 Feb 2017
------
- Fix bug in timer code that caused problems at optimization settings
greater than 2
- OSHMEM: make mmap allocator the default instead of sysv or verbs
- Support MPI_Dims_create with dimension zero
- Update USNIC support
- Prevent 64-bit overflow on timer counter
- Add support for forwarding signals
- Fix bug that caused truncated messages on large sends over TCP BTL
- Fix potential infinite loop when printing a stacktrace
1.10.5 - 19 Dec 2016
------
- Update UCX APIs
@ -260,6 +508,8 @@ Master (not on release branches yet)
1.10.3 - 15 June 2016
------
- Fix zero-length datatypes. Thanks to Wei-keng Liao for reporting
the issue.
- Minor manpage cleanups
- Implement atomic support in OSHMEM/UCX
- Fix support of MPI_COMBINER_RESIZED. Thanks to James Ramsey
@ -310,6 +560,23 @@ Master (not on release branches yet)
- Fix affinity for MPMD jobs running under LSF
- Fix many Fortran binding bugs
- Fix `MPI_IN_PLACE`-related bugs
- Fix PSM/PSM2 support for singleton operations
- Ensure MPI transports continue to progress during RTE barriers
- Update HWLOC to 1.9.1 end-of-series
- Fix a bug in the Java command line parser when the
-Djava.library.path options was given by the user
- Update the MTL/OFI provider selection behavior
- Add support for clock_gettime on Linux.
- Correctly detect and configure for Solaris Studio 12.5
beta compilers
- Correctly compute #slots when -host is used for MPMD case
- Fix a bug in the hcoll collectives due to an uninitialized field
- Do not set a binding policy when oversubscribing a node
- Fix hang in intercommunicator operations when oversubscribed
- Speed up process termination during MPI_Abort
- Disable backtrace support by default in the PSM/PSM2 libraries to
prevent unintentional conflicting behavior.
1.10.2: 26 Jan 2016