Skip to content

Latest commit

 

History

History
2296 lines (1832 loc) · 93.4 KB

README.md

File metadata and controls

2296 lines (1832 loc) · 93.4 KB

Open MPI

The Open MPI Project is an open source Message Passing Interface (MPI) implementation that is developed and maintained by a consortium of academic, research, and industry partners. Open MPI is therefore able to combine the expertise, technologies, and resources from all across the High Performance Computing community in order to build the best MPI library available. Open MPI offers advantages for system and software vendors, application developers and computer science researchers.

See the MPI Forum web site for information about the MPI API specification.

Quick start

In many cases, Open MPI can be built and installed by simply indicating the installation directory on the command line:

$ tar xf openmpi-<version>.tar.bz2
$ cd openmpi-<version>
$ ./configure --prefix=<path> |& tee config.out
...lots of output...
$ make -j 8 |& tee make.out
...lots of output...
$ make install |& tee install.out
...lots of output...

Note that there are many, many configuration options to the ./configure step. Some of them may be needed for your particular environmnet; see below for desciptions of the options available.

If your installation prefix path is not writable by a regular user, you may need to use sudo or su to run the make install step. For example:

$ sudo make install |& tee install.out
[sudo] password for jsquyres: <enter your password here>
...lots of output...

Finally, note that VPATH builds are fully supported. For example:

$ tar xf openmpi-<version>.tar.bz2
$ cd openmpi-<version>
$ mkdir build
$ cd build
$ ../configure --prefix=<path> |& tee config.out
...etc.

Table of contents

The rest of this file contains:

Also, note that much, much more information is also available in the Open MPI FAQ.

General notes

The following abbreviated list of release notes applies to this code base as of this writing (April 2020):

  • Open MPI now includes two public software layers: MPI and OpenSHMEM. Throughout this document, references to Open MPI implicitly include both of these layers. When distinction between these two layers is necessary, we will reference them as the "MPI" and "OpenSHMEM" layers respectively.

  • OpenSHMEM is a collaborative effort between academia, industry, and the U.S. Government to create a specification for a standardized API for parallel programming in the Partitioned Global Address Space (PGAS). For more information about the OpenSHMEM project, including access to the current OpenSHMEM specification, please visit http://openshmem.org/.

    This OpenSHMEM implementation will only work in Linux environments with a restricted set of supported networks.

  • Open MPI includes support for a wide variety of supplemental hardware and software package. When configuring Open MPI, you may need to supply additional flags to the configure script in order to tell Open MPI where the header files, libraries, and any other required files are located. As such, running configure by itself may not include support for all the devices (etc.) that you expect, especially if their support headers / libraries are installed in non-standard locations. Network interconnects are an easy example to discuss -- Libfabric and OpenFabrics networks, for example, both have supplemental headers and libraries that must be found before Open MPI can build support for them. You must specify where these files are with the appropriate options to configure. See the listing of configure command-line switches, below, for more details.

  • The majority of Open MPI's documentation is here in this file, the included man pages, and on the web site FAQ.

  • Note that Open MPI documentation uses the word "component" frequently; the word "plugin" is probably more familiar to most users. As such, end users can probably completely substitute the word "plugin" wherever you see "component" in our documentation. For what it's worth, we use the word "component" for historical reasons, mainly because it is part of our acronyms and internal API function calls.

  • The run-time systems that are currently supported are:

    • rsh / ssh
    • PBS Pro, Torque
    • Platform LSF (tested with v9.1.1 and later)
    • SLURM
    • Cray XE, XC, and XK
    • Oracle Grid Engine (OGE) 6.1, 6.2 and open source Grid Engine
  • Systems that have been tested are:

    • Linux (various flavors/distros), 64 bit (x86, ppc, aarch64), with gcc (>=4.8.x+), clang (>=3.6.0), Absoft (fortran), Intel, and Portland (*)
    • macOS (10.14-10.15), 64 bit (x86_64) with XCode compilers

    (*) Be sure to read the Compiler Notes, below.

  • Other systems have been lightly (but not fully) tested:

    • Linux (various flavors/distros), 32 bit, with gcc
    • Cygwin 32 & 64 bit with gcc
    • ARMv6, ARMv7
    • Other 64 bit platforms.
    • OpenBSD. Requires configure options --enable-mca-no-build=patcher and --disable-dlopen with this release.
    • Problems have been reported when building Open MPI on FreeBSD 11.1 using the clang-4.0 system compiler. A workaround is to build Open MPI using the GNU compiler.
  • Open MPI has taken some steps towards Reproducible Builds. Specifically, Open MPI's configure and make process, by default, records the build date and some system-specific information such as the hostname where Open MPI was built and the username who built it. If you desire a Reproducible Build, set the $SOURCE_DATE_EPOCH, $USER and $HOSTNAME environment variables before invoking configure and make, and Open MPI will use those values instead of invoking whoami and/or hostname, respectively. See https://reproducible-builds.org/docs/source-date-epoch/ for information on the expected format and content of the $SOURCE_DATE_EPOCH variable.

Platform Notes

  • N/A

Compiler Notes

  • Open MPI requires a C99-capable compiler to build.

  • On platforms other than x86-64, AArch64 (64-bit ARM), and PPC64, Open MPI requires a compiler that either supports C11 atomics or the GCC __atomic atomics (e.g., GCC >= v4.8.1).

  • 32-bit platforms are only supported with a recent compiler that supports C11 atomics. This includes gcc 4.9.x+ (gcc 6.x or newer recommened), icc 16, clang 3.1+, etc.

  • Mixing compilers from different vendors when building Open MPI (e.g., using the C/C++ compiler from one vendor and the Fortran compiler from a different vendor) has been successfully employed by some Open MPI users (discussed on the Open MPI user's mailing list), but such configurations are not tested and not documented. For example, such configurations may require additional compiler / linker flags to make Open MPI build properly.

    A not-uncommon case for this is when building on MacOS with the system-default GCC compiler (i.e., /usr/bin/gcc), but a 3rd party gfortran (e.g., provided by Homebrew, in /usr/local/bin/gfortran). Since these compilers are provided by different organizations, they have different default search paths. For example, if Homebrew has also installed a local copy of Libevent (a 3rd party package that Open MPI requires), the MacOS-default gcc linker will find it without any additional command line flags, but the Homebrew-provided gfortran linker will not. In this case, it may be necessary to provide the following on the configure command line:

    $ ./configure FCFLAGS=-L/usr/local/lib ...
    

    This -L flag will then be passed to the Fortran linker when creating Open MPI's Fortran libraries, and it will therefore be able to find the installed Libevent.

  • In general, the latest versions of compilers of a given vendor's series have the least bugs. We have seen cases where Vendor XYZ's compiler version A.B fails to compile Open MPI, but version A.C (where C>B) works just fine. If you run into a compile failure, you might want to double check that you have the latest bug fixes and patches for your compiler.

  • Users have reported issues with older versions of the Fortran PGI compiler suite when using Open MPI's (non-default) --enable-debug configure option. Per the above advice of using the most recent version of a compiler series, the Open MPI team recommends using the latest version of the PGI suite, and/or not using the --enable-debug configure option. If it helps, here's what we have found with some (not comprehensive) testing of various versions of the PGI compiler suite:

    • pgi-8 : NO known good version with --enable-debug
    • pgi-9 : 9.0-4 known GOOD
    • pgi-10: 10.0-0 known GOOD
    • pgi-11: NO known good version with --enable-debug
    • pgi-12: 12.10 known BAD with -m32, but known GOOD without -m32 (and 12.8 and 12.9 both known BAD with --enable-debug)
    • pgi-13: 13.9 known BAD with -m32, 13.10 known GOOD without -m32
    • pgi-15: 15.10 known BAD with -m32
  • Similarly, there is a known Fortran PGI compiler issue with long source directory path names that was resolved in 9.0-4 (9.0-3 is known to be broken in this regard).

  • Open MPI does not support the PGI compiler suite on OS X or MacOS. See issues below for more details:

  • OpenSHMEM Fortran bindings do not support the "no underscore" Fortran symbol convention. IBM's xlf compilers build in that mode by default. As such, IBM's xlf compilers cannot build/link the OpenSHMEM Fortran bindings by default. A workaround is to pass FC="xlf -qextname" at configure time to force a trailing underscore. See this issue for more details.

  • MPI applications that use the mpi_f08 module on PowerPC platforms (tested ppc64le) will likely experience runtime failures if:

    • they are using a GNU linker (ld) version after v2.25.1 and before v2.28, and
    • they compiled with PGI (tested 17.5) or XL (tested v15.1.5) compilers. This was noticed on Ubuntu 16.04 which uses the 2.26.1 version of ld by default. However, this issue impacts any OS using a version of ld noted above. This GNU linker regression will be fixed in version 2.28. Here is a link to the GNU bug on this issue. The XL compiler will include a fix for this issue in a future release.
  • On NetBSD-6 (at least AMD64 and i386), and possibly on OpenBSD, Libtool misidentifies properties of f95/g95, leading to obscure compile-time failures if used to build Open MPI. You can work around this issue by ensuring that libtool will not use f95/g95 (e.g., by specifying FC=<some_other_compiler>, or otherwise ensuring a different Fortran compiler will be found earlier in the path than f95/g95), or by disabling the Fortran MPI bindings with --disable-mpi-fortran.

  • On OpenBSD/i386, if you configure with --enable-mca-no-build=patcher, you will also need to add --disable-dlopen. Otherwise, odd crashes can occur nondeterministically.

  • Absoft 11.5.2 plus a service pack from September 2012 (which Absoft says is available upon request), or a version later than 11.5.2 (e.g., 11.5.3), is required to compile the Fortran mpi_f08 module.

  • Open MPI does not support the Sparc v8 CPU target. However, as of Solaris Studio 12.1, and later compilers, one should not specify -xarch=v8plus or -xarch=v9. The use of the options -m32 and -m64 for producing 32 and 64 bit targets, respectively, are now preferred by the Solaris Studio compilers. GCC may require either -m32 or -mcpu=v9 -m32, depending on GCC version.

  • If one tries to build OMPI on Ubuntu with Solaris Studio using the C++ compiler and the -m32 option, you might see a warning:

    CC: Warning: failed to detect system linker version, falling back to custom linker usage
    

    And the build will fail. One can overcome this error by either setting LD_LIBRARY_PATH to the location of the 32 bit libraries (most likely /lib32), or giving LDFLAGS="-L/lib32 -R/lib32" to the configure command. Officially, Solaris Studio is not supported on Ubuntu Linux distributions, so additional problems might be incurred.

  • Open MPI does not support the gccfss compiler (GCC For SPARC Systems; a now-defunct compiler project from Sun).

  • At least some versions of the Intel 8.1 compiler seg fault while compiling certain Open MPI source code files. As such, it is not supported.

  • It has been reported that the Intel 9.1 and 10.0 compilers fail to compile Open MPI on IA64 platforms. As of 12 Sep 2012, there is very little (if any) testing performed on IA64 platforms (with any compiler). Support is "best effort" for these platforms, but it is doubtful that any effort will be expended to fix the Intel 9.1 / 10.0 compiler issuers on this platform.

  • Early versions of the Intel 12.1 Linux compiler suite on x86_64 seem to have a bug that prevents Open MPI from working. Symptoms including immediate segv of the wrapper compilers (e.g., mpicc) and MPI applications. As of 1 Feb 2012, if you upgrade to the latest version of the Intel 12.1 Linux compiler suite, the problem will go away.

  • Users have reported that the Intel Fortran compiler will fail to link Fortran-based MPI applications on macOS with linker errors similar to this:

    Undefined symbols for architecture x86_64:
      "_ompi_buffer_detach_f08", referenced from:
          import-atom in libmpi_usempif08.dylib
    ld: symbol(s) not found for architecture x86_64
    

    It appears that setting the environment variable lt_cx_ld_force_load=no before invoking Open MPI's configure script works around the issue. For example:

    shell$ lt_cv_ld_force_load=no ./configure ...
    
  • The Portland Group compilers prior to version 7.0 require the -Msignextend compiler flag to extend the sign bit when converting from a shorter to longer integer. This is is different than other compilers (such as GNU). When compiling Open MPI with the Portland compiler suite, the following flags should be passed to Open MPI's configure script:

    shell$ ./configure CFLAGS=-Msignextend CXXFLAGS=-Msignextend \
           --with-wrapper-cflags=-Msignextend \
           --with-wrapper-cxxflags=-Msignextend ...
    

    This will both compile Open MPI with the proper compile flags and also automatically add "-Msignextend" when the C and C++ MPI wrapper compilers are used to compile user MPI applications.

  • It has been reported that Pathscale 5.0.5 and 6.0.527 compilers give an internal compiler error when trying to build Open MPI.

  • As of July 2017, the Pathscale compiler suite apparently has no further commercial support, and it does not look like there will be further releases. Any issues discovered regarding building / running Open MPI with the Pathscale compiler suite therefore may not be able to be resolved.

  • Using the Absoft compiler to build the MPI Fortran bindings on Suse 9.3 is known to fail due to a Libtool compatibility issue.

  • MPI Fortran API support has been completely overhauled since the Open MPI v1.5/v1.6 series.

    There is now only a single Fortran MPI wrapper compiler and a single Fortran OpenSHMEM wrapper compiler: mpifort and oshfort, respectively. mpif77 and mpif90 still exist, but they are symbolic links to mpifort.

    Similarly, Open MPI's configure script only recognizes the FC and FCFLAGS environment variables (to specify the Fortran compiler and compiler flags, respectively). The F77 and FFLAGS environment variables are IGNORED.

    As a direct result, it is STRONGLY recommended that you specify a Fortran compiler that uses file suffixes to determine Fortran code layout (e.g., free form vs. fixed). For example, with some versions of the IBM XLF compiler, it is preferable to use FC=xlf instead of FC=xlf90, because xlf will automatically determine the difference between free form and fixed Fortran source code.

    However, many Fortran compilers allow specifying additional command-line arguments to indicate which Fortran dialect to use. For example, if FC=xlf90, you may need to use mpifort --qfixed ... to compile fixed format Fortran source files.

    You can use either ompi_info or oshmem_info to see with which Fortran compiler Open MPI was configured and compiled.

    There are up to three sets of Fortran MPI bindings that may be provided (depending on your Fortran compiler):

    1. mpif.h: This is the first MPI Fortran interface that was defined in MPI-1. It is a file that is included in Fortran source code. Open MPI's mpif.h does not declare any MPI subroutines; they are all implicit.

    2. mpi module: The mpi module file was added in MPI-2. It provides strong compile-time parameter type checking for MPI subroutines.

    3. mpi_f08 module: The mpi_f08 module was added in MPI-3. It provides many advantages over the mpif.h file and mpi module. For example, MPI handles have distinct types (vs. all being integers). See the MPI-3 document for more details.

    NOTE: The mpi_f08 module is STRONGLY recommended for all new MPI Fortran subroutines and applications. Note that the mpi_f08 module can be used in conjunction with the other two Fortran MPI bindings in the same application (only one binding can be used per subroutine/function, however). Full interoperability between mpif.h/mpi module and mpi_f08 module MPI handle types is provided, allowing mpi_f08 to be used in new subroutines in legacy MPI applications.

    Per the OpenSHMEM specification, there is only one Fortran OpenSHMEM binding provided:

    • shmem.fh: All Fortran OpenSHMEM programs should include shmem.f, and Fortran OpenSHMEM programs that use constants defined by OpenSHMEM MUST include shmem.fh.

    The following notes apply to the above-listed Fortran bindings:

    • All Fortran compilers support the mpif.h/shmem.fh-based bindings, with one exception: the MPI_SIZEOF interfaces will only be present when Open MPI is built with a Fortran compiler that supports the INTERFACE keyword and ISO_FORTRAN_ENV. Most notably, this excludes the GNU Fortran compiler suite before version 4.9.

    • The level of support provided by the mpi module is based on your Fortran compiler.

      If Open MPI is built with a non-GNU Fortran compiler, or if Open MPI is built with the GNU Fortran compiler >= v4.9, all MPI subroutines will be prototyped in the mpi module. All calls to MPI subroutines will therefore have their parameter types checked at compile time.

      If Open MPI is built with an old gfortran (i.e., < v4.9), a limited mpi module will be built. Due to the limitations of these compilers, and per guidance from the MPI-3 specification, all MPI subroutines with "choice" buffers are specifically not included in the mpi module, and their parameters will not be checked at compile time. Specifically, all MPI subroutines with no "choice" buffers are prototyped and will receive strong parameter type checking at run-time (e.g., MPI_INIT, MPI_COMM_RANK, etc.).

      Similar to the mpif.h interface, MPI_SIZEOF is only supported on Fortran compilers that support INTERFACE and ISO_FORTRAN_ENV.

    • The mpi_f08 module has been tested with the Intel Fortran compiler and gfortran >= 4.9. Other modern Fortran compilers likely also work.

      Many older Fortran compilers do not provide enough modern Fortran features to support the mpi_f08 module. For example, gfortran < v4.9 does provide enough support for the mpi_f08 module.

    You can examine the output of the following command to see all the Fortran features that are/are not enabled in your Open MPI installation:

    shell$ ompi_info | grep -i fort
    
  • As of October 2021, the NVIDIA 'nvfortran' compiler version 21.1 fails to link libmpi_usempif08.la (see open-mpi#8919). One can work-around this issue by explicitely setting FCFLAGS=-fPIC during configure.

General Run-Time Support Notes

  • The Open MPI installation must be in your PATH on all nodes (and potentially LD_LIBRARY_PATH or DYLD_LIBRARY_PATH, if libmpi/libshmem is a shared library), unless using the --prefix or --enable-mpirun-prefix-by-default functionality (see below).

  • Open MPI's run-time behavior can be customized via Modular Component Architecture (MCA) parameters (see below for more information on how to get/set MCA parameter values). Some MCA parameters can be set in a way that renders Open MPI inoperable (see notes about MCA parameters later in this file). In particular, some parameters have required options that must be included.

    • If specified, the btl parameter must include the self component, or Open MPI will not be able to deliver messages to the same rank as the sender. For example: mpirun --mca btl tcp,self ...
    • If specified, the btl_tcp_if_exclude parameter must include the loopback device (lo on many Linux platforms), or Open MPI will not be able to route MPI messages using the TCP BTL. For example: mpirun --mca btl_tcp_if_exclude lo,eth1 ...
  • Running on nodes with different endian and/or different datatype sizes within a single parallel job is supported in this release. However, Open MPI does not resize data when datatypes differ in size (for example, sending a 4 byte MPI_DOUBLE and receiving an 8 byte MPI_DOUBLE will fail).

MPI Functionality and Features

  • All MPI-3.1 functionality is supported.

  • Note that starting with Open MPI v4.0.0, prototypes for several legacy MPI-1 symbols that were deleted in the MPI-3.0 specification (which was published in 2012) are no longer available by default in mpi.h. Specifically, several MPI-1 symbols were deprecated in the 1996 publishing of the MPI-2.0 specification. These deprecated symbols were eventually removed from the MPI-3.0 specification in 2012.

    The symbols that now no longer appear by default in Open MPI's mpi.h are:

    • MPI_Address (replaced by MPI_Get_address)
    • MPI_Errhandler_create (replaced by MPI_Comm_create_errhandler)
    • MPI_Errhandler_get (replaced by MPI_Comm_get_errhandler)
    • MPI_Errhandler_set (replaced by MPI_Comm_set_errhandler)
    • MPI_Type_extent (replaced by MPI_Type_get_extent)
    • MPI_Type_hindexed (replaced by MPI_Type_create_hindexed)
    • MPI_Type_hvector (replaced by MPI_Type_create_hvector)
    • MPI_Type_lb (replaced by MPI_Type_get_extent)
    • MPI_Type_struct (replaced by MPI_Type_create_struct)
    • MPI_Type_ub (replaced by MPI_Type_get_extent)
    • MPI_LB (replaced by MPI_Type_create_resized)
    • MPI_UB (replaced by MPI_Type_create_resized)
    • MPI_COMBINER_HINDEXED_INTEGER
    • MPI_COMBINER_HVECTOR_INTEGER
    • MPI_COMBINER_STRUCT_INTEGER
    • MPI_Handler_function (replaced by MPI_Comm_errhandler_function)

    Although these symbols are no longer prototyped in mpi.h, they are still present in the MPI library in Open MPI v4.0.x. This enables legacy MPI applications to link and run successfully with Open MPI v4.0.x, even though they will fail to compile.

    WARNING: Future releases of Open MPI beyond the v4.0.x series may remove these symbols altogether.

    WARNING: The Open MPI team STRONGLY encourages all MPI application developers to stop using these constructs that were first deprecated over 20 years ago, and finally removed from the MPI specification in MPI-3.0 (in 2012).

    WARNING: The Open MPI FAQ contains examples of how to update legacy MPI applications using these deleted symbols to use the "new" symbols.

    All that being said, if you are unable to immediately update your application to stop using these legacy MPI-1 symbols, you can re-enable them in mpi.h by configuring Open MPI with the --enable-mpi1-compatibility flag.

  • Rank reordering support is available using the TreeMatch library. It is activated for the graph and dist_graph communicator topologies.

  • When using MPI deprecated functions, some compilers will emit warnings. For example:

    shell$ cat deprecated_example.c
    #include <mpi.h>
    void foo(void) {
        MPI_Datatype type;
        MPI_Type_struct(1, NULL, NULL, NULL, &type);
    }
    shell$ mpicc -c deprecated_example.c
    deprecated_example.c: In function 'foo':
    deprecated_example.c:4: warning: 'MPI_Type_struct' is deprecated (declared at /opt/openmpi/include/mpi.h:1522)
    shell$
    
  • MPI_THREAD_MULTIPLE is supported with some exceptions.

    The following PMLs support MPI_THREAD_MULTIPLE:

    1. cm (see list (1) of supported MTLs, below)
    2. ob1 (see list (2) of supported BTLs, below)
    3. ucx

    (1) The cm PML and the following MTLs support MPI_THREAD_MULTIPLE:

    1. ofi (Libfabric)
    2. portals4

    (2) The ob1 PML and the following BTLs support MPI_THREAD_MULTIPLE:

    1. self
    2. sm
    3. smcuda
    4. tcp
    5. ugni
    6. usnic

    Currently, MPI File operations are not thread safe even if MPI is initialized for MPI_THREAD_MULTIPLE support.

  • MPI_REAL16 and MPI_COMPLEX32 are only supported on platforms where a portable C datatype can be found that matches the Fortran type REAL*16, both in size and bit representation.

  • The "libompitrace" library is bundled in Open MPI and is installed by default (it can be disabled via the --disable-libompitrace flag). This library provides a simplistic tracing of select MPI function calls via the MPI profiling interface. Linking it in to your application via (e.g., via -lompitrace) will automatically output to stderr when some MPI functions are invoked:

    shell$ cd examples/
    shell$ mpicc hello_c.c -o hello_c -lompitrace
    shell$ mpirun -np 1 hello_c
    MPI_INIT: argc 1
    Hello, world, I am 0 of 1
    MPI_BARRIER[0]: comm MPI_COMM_WORLD
    MPI_FINALIZE[0]
    shell$
    

    Keep in mind that the output from the trace library is going to stderr, so it may output in a slightly different order than the stdout from your application.

    This library is being offered as a "proof of concept" / convenience from Open MPI. If there is interest, it is trivially easy to extend it to printf for other MPI functions. Pull requests on github.com would be greatly appreciated.

OpenSHMEM Functionality and Features

All OpenSHMEM-1.3 functionality is supported.

MPI Collectives

  • The cuda coll component provides CUDA-aware support for the reduction type collectives with GPU buffers. This component is only compiled into the library when the library has been configured with CUDA-aware support. It intercepts calls to the reduction collectives, copies the data to staging buffers if GPU buffers, then calls underlying collectives to do the work.

OpenSHMEM Collectives

  • The fca scoll component: the Mellanox Fabric Collective Accelerator (FCA) is a solution for offloading collective operations from the MPI process onto Mellanox QDR InfiniBand switch CPUs and HCAs.

  • The basic scoll component: Reference implementation of all OpenSHMEM collective operations.

Network Support

  • There are several main MPI network models available: ob1, cm, and ucx. ob1 uses BTL ("Byte Transfer Layer") components for each supported network. cm uses MTL ("Matching Transport Layer") components for each supported network. ucx uses the OpenUCX transport.

    • ob1 supports a variety of networks that can be used in combination with each other:

      • OpenFabrics: InfiniBand, iWARP, and RoCE
      • Loopback (send-to-self)
      • Shared memory
      • TCP
      • SMCUDA
      • Cisco usNIC
      • uGNI (Cray Gemini, Aries)
      • shared memory (XPMEM, Linux CMA, Linux KNEM, and copy-in/copy-out shared memory)
    • cm supports a smaller number of networks (and they cannot be used together), but may provide better overall MPI performance:

      • Intel Omni-Path PSM2 (version 11.2.173 or later)
      • Intel True Scale PSM (QLogic InfiniPath)
      • OpenFabrics Interfaces ("libfabric" tag matching)
      • Portals 4
    • UCX is the Unified Communication X (UCX) communication library. This is an open-source project developed in collaboration between industry, laboratories, and academia to create an open-source production grade communication framework for data centric and high-performance applications. The UCX library can be downloaded from repositories (e.g., Fedora/RedHat yum repositories). The UCX library is also part of Mellanox OFED and Mellanox HPC-X binary distributions.

      UCX currently supports:

      • OpenFabrics Verbs (including InfiniBand and RoCE)
      • Cray's uGNI
      • TCP
      • Shared memory
      • NVIDIA CUDA drivers

    While users can manually select any of the above transports at run time, Open MPI will select a default transport as follows:

    1. If InfiniBand devices are available, use the UCX PML.
    2. If PSM, PSM2, or other tag-matching-supporting Libfabric transport devices are available (e.g., Cray uGNI), use the cm PML and a single appropriate corresponding mtl module.
    3. Otherwise, use the ob1 PML and one or more appropriate btl modules.

    Users can override Open MPI's default selection algorithms and force the use of a specific transport if desired by setting the pml MCA parameter (and potentially the btl and/or mtl MCA parameters) at run-time:

    shell$ mpirun --mca pml ob1 --mca btl [comma-delimted-BTLs] ...
    or
    shell$ mpirun --mca pml cm --mca mtl [MTL] ...
    or
    shell$ mpirun --mca pml ucx ...
    

    There is a known issue when using UCX with very old Mellanox Infiniband HCAs, in particular HCAs preceding the introduction of the ConnectX product line, which can result in Open MPI crashing in MPI_Finalize. This issue is addressed by UCX release 1.9.0 and newer.

  • The main OpenSHMEM network model is ucx; it interfaces directly with UCX.

  • In prior versions of Open MPI, InfiniBand and RoCE support was provided through the openib BTL and ob1 PML plugins. Starting with Open MPI 4.0.0, InfiniBand support through the openib plugin is both deprecated and superseded by the ucx PML component. The openib BTL was removed in Open MPI v5.0.0.

    While the openib BTL depended on libibverbs, the UCX PML depends on the UCX library.

    Once installed, Open MPI can be built with UCX support by adding --with-ucx to the Open MPI configure command. Once Open MPI is configured to use UCX, the runtime will automatically select the ucx PML if one of the supported networks is detected (e.g., InfiniBand). It's possible to force using UCX in the mpirun or oshrun command lines by specifying any or all of the following mca parameters: --mca pml ucx for MPI point-to-point operations, --mca spml ucx for OpenSHMEM support, and --mca osc ucx for MPI RMA (one-sided) operations.

  • The usnic BTL is support for Cisco's usNIC device ("userspace NIC") on Cisco UCS servers with the Virtualized Interface Card (VIC). Although the usNIC is accessed via the OpenFabrics Libfabric API stack, this BTL is specific to Cisco usNIC devices.

  • uGNI is a Cray library for communicating over the Gemini and Aries interconnects.

  • Linux knem support is used when the sm (shared memory) BTL is compiled with knem support (see the --with-knem configure option) and the knem Linux module is loaded in the running kernel. If the knem Linux kernel module is not loaded, the knem support is (by default) silently deactivated during Open MPI jobs.

    See https://knem.gforge.inria.fr/ for details on Knem.

  • Linux Cross-Memory Attach (CMA) or XPMEM is used by the sm shared memory BTL when the CMA/XPMEM libraries are installed, respectively. Linux CMA and XPMEM are similar (but different) mechanisms for Open MPI to utilize single-copy semantics for shared memory.

  • The OFI MTL does not support sending messages larger than the active Libfabric provider's max_msg_size. If you receive an error message about sending too large of a message when using the OFI MTL, please reach out to your networking vendor to ask them to support a larger max_msg_size for tagged messages.

Open MPI Extensions

An MPI "extensions" framework is included in Open MPI, but is not enabled by default. See the "Open MPI API Extensions" section below for more information on compiling and using MPI extensions.

The following extensions are included in this version of Open MPI:

  1. pcollreq: Provides routines for persistent collective communication operations and persistent neighborhood collective communication operations, which are planned to be included in MPI-4.0. The function names are prefixed with MPIX_ instead of MPI_, like MPIX_Barrier_init, because they are not standardized yet. Future versions of Open MPI will switch to the MPI_ prefix once the MPI Standard which includes this feature is published. See their man page for more details.
  2. shortfloat: Provides MPI datatypes MPIX_C_FLOAT16, MPIX_SHORT_FLOAT, MPIX_SHORT_FLOAT, and MPIX_CXX_SHORT_FLOAT_COMPLEX if corresponding language types are available. See ompi/mpiext/shortfloat/README.txt for details.
  3. affinity: Provides the OMPI_Affinity_str() API, which returns a string indicating the resources which a process is bound. For more details, see its man page.
  4. cuda: When the library is compiled with CUDA-aware support, it provides two things. First, a macro MPIX_CUDA_AWARE_SUPPORT. Secondly, the function MPIX_Query_cuda_support() that can be used to query for support.
  5. example: A non-functional extension; its only purpose is to provide an example for how to create other extensions.

Building Open MPI

If you have checked out a developer's copy of Open MPI (i.e., you cloned from Git), you really need to read the HACKING file before attempting to build Open MPI. Really.

If you have downloaded a tarball, then things are much simpler. Open MPI uses a traditional configure script paired with make to build. Typical installs can be of the pattern:

shell$ ./configure [...options...]
shell$ make [-j N] all install
      (use an integer value of N for parallel builds)

There are many available configure options (see ./configure --help for a full list); a summary of the more commonly used ones is included below.

NOTE: if you are building Open MPI on a network filesystem, the machine you on which you are building must be time-synchronized with the file server. Specifically: Open MPI's build system requires accurate filesystem timestamps. If your make output includes warning about timestamps in the future or runs GNU Automake, Autoconf, and/or Libtool, this is not normal, and you may have an invalid build. Ensure that the time on your build machine is synchronized with the time on your file server, or build on a local filesystem. Then remove the Open MPI source directory and start over (e.g., by re-extracting the Open MPI tarball).

Note that for many of Open MPI's --with-FOO options, Open MPI will, by default, search for header files and/or libraries for FOO. If the relevant files are found, Open MPI will built support for FOO; if they are not found, Open MPI will skip building support for FOO. However, if you specify --with-FOO on the configure command line and Open MPI is unable to find relevant support for FOO, configure will assume that it was unable to provide a feature that was specifically requested and will abort so that a human can resolve out the issue.

Additionally, if a search directory is specified in the form --with-FOO=DIR, Open MPI will:

  1. Search for FOO's header files in DIR/include.
  2. Search for FOO's library files:
    1. If --with-FOO-libdir=<libdir> was specified, search in <libdir>.
    2. Otherwise, search in DIR/lib, and if they are not found there, search again in DIR/lib64.
  3. If both the relevant header files and libraries are found:
    1. Open MPI will build support for FOO.
    2. If the root path where the FOO libraries are found is neither /usr nor /usr/local, Open MPI will compile itself with RPATH flags pointing to the directory where FOO's libraries are located. Open MPI does not RPATH /usr/lib[64] and /usr/local/lib[64] because many systems already search these directories for run-time libraries by default; adding RPATH for them could have unintended consequences for the search path ordering.

Installation Options

  • --prefix=DIR: Install Open MPI into the base directory named DIR. Hence, Open MPI will place its executables in DIR/bin, its header files in DIR/include, its libraries in DIR/lib, etc.

  • --disable-shared: By default, Open MPI and OpenSHMEM build shared libraries, and all components are built as dynamic shared objects (DSOs). This switch disables this default; it is really only useful when used with --enable-static. Specifically, this option does not imply --enable-static; enabling static libraries and disabling shared libraries are two independent options.

  • --enable-static: Build MPI and OpenSHMEM as static libraries, and statically link in all components. Note that this option does not imply --disable-shared; enabling static libraries and disabling shared libraries are two independent options.

    Be sure to read the description of --without-memory-manager, below; it may have some effect on --enable-static.

  • --disable-wrapper-rpath: By default, the wrapper compilers (e.g., mpicc) will enable "rpath" support in generated executables on systems that support it. That is, they will include a file reference to the location of Open MPI's libraries in the application executable itself. This means that the user does not have to set LD_LIBRARY_PATH to find Open MPI's libraries (e.g., if they are installed in a location that the run-time linker does not search by default).

    On systems that utilize the GNU ld linker, recent enough versions will actually utilize "runpath" functionality, not "rpath". There is an important difference between the two:

    1. "rpath": the location of the Open MPI libraries is hard-coded into the MPI/OpenSHMEM application and cannot be overridden at run-time.
    2. "runpath": the location of the Open MPI libraries is hard-coded into the MPI/OpenSHMEM application, but can be overridden at run-time by setting the LD_LIBRARY_PATH environment variable.

    For example, consider that you install Open MPI vA.B.0 and compile/link your MPI/OpenSHMEM application against it. Later, you install Open MPI vA.B.1 to a different installation prefix (e.g., /opt/openmpi/A.B.1 vs. /opt/openmpi/A.B.0), and you leave the old installation intact.

    In the rpath case, your MPI application will always use the libraries from your A.B.0 installation. In the runpath case, you can set the LD_LIBRARY_PATH environment variable to point to the A.B.1 installation, and then your MPI application will use those libraries.

    Note that in both cases, however, if you remove the original A.B.0 installation and set LD_LIBRARY_PATH to point to the A.B.1 installation, your application will use the A.B.1 libraries.

    This rpath/runpath behavior can be disabled via --disable-wrapper-rpath.

    If you would like to keep the rpath option, but not enable runpath a different configure option is avalabile --disable-wrapper-runpath.

  • --enable-dlopen: Enable loading of Open MPI components as standalone Dynamic Shared Objects (DSOs) that are loaded at run-time. This option is enabled by default.

    The opposite of this option, --disable-dlopen, causes the following:

    1. Open MPI will not attempt to open any DSOs at run-time.
    2. configure behaves as if the --enable-mca-static argument was set.
    3. configure will ignore the --enable-mca-dso argument.

    See the description of --enable-mca-static / --enable-mca-dso for more information.

    Note that this option does not change how Open MPI's libraries (libmpi, for example) will be built. You can change whether Open MPI builds static or dynamic libraries via the --enable|disable-static and --enable|disable-shared arguments.

  • --enable-mca-dso[=LIST] and --enable-mca-static[=LIST] These two options, along with --enable-mca-no-build, govern the behavior of how Open MPI's frameworks and components are built.

    The --enable-mca-dso option specifies which frameworks and/or components are built as Dynamic Shared Objects (DSOs). Specifically, DSOs are built as "plugins" outside of the core Open MPI libraries, and are loaded by Open MPI at run time.

    The --enable-mca-static option specifies which frameworks and/or components are built as part of the core Open MPI libraries (i.e., they are not built as DSOs, and therefore do not need to be separately discovered and opened at run time).

    Both options can be used one of two ways:

    1. --enable-mca-OPTION (with no value)
    2. --enable-mca-OPTION=LIST

    --enable-mca-OPTION=no or --disable-mca-OPTION are both legal options, but have no impact on the selection logic described below. Only affirmative options change the selection process.

    LIST is a comma-delimited list of Open MPI frameworks and/or framework+component tuples. Examples:

    • "btl" specifies the entire BTL framework
    • "btl-tcp" specifies just the TCP component in the BTL framework
    • "mtl,btl-tcp" specifies the entire MTL framework and the TCP component in the BTL framework

    Open MPI's configure script uses the values of these two options when evaluating each component to determine how it should be built by evaluating these conditions in order:

    1. If an individual component's build behavior has been specified via these two options, configure uses that behavior.
    2. Otherwise, if the component is in a framework whose build behavior has been specified via these two options, configure uses that behavior.
    3. Otherwise, configure uses the global default build behavior.

    At each level of the selection process, if the component is specified to be built as both a static and dso component, the static option will win.

    Note that as of Open MPI v5.0.0, configure's global default is to build all components as static (i.e., part of the Open MPI core libraries, not as DSO's). Prior to Open MPI v5.0.0, the global default behavior was to build most components as DSOs.

    Also note that if the --disable-dlopen option is specified, then Open MPI will not be able to search for DSOs at run time, and the value of the --enable-mca-dso option will be silently ignored.

    Some examples:

    1. Default to building all components as static (i.e., as part of the Open MPI core libraries -- no DSOs):

      $ ./configure

    2. Build all components as static, except the TCP BTL, which will be built as a DSO:

      $ ./configure --enable-mca-dso=btl-tcp

    3. Build all components as static, except all BTL components, which will be built as DSOs:

      $ ./configure --enable-mca-dso=btl

    4. Build all components as static, except all MTL components and the TCP BTL component, which will be built as DSOs:

      $ ./configure --enable-mca-dso=mtl,btl-tcp

    5. Build all BTLs as static, except the TCP BTL, as the option is more specific than the option:

      $ ./configure --enable-mca-dso=btl --enable-mca-static=btl-tcp

    6. Build the TCP BTL as static, because the static option at the same level always wins:

      $ ./configure --enable-mca-dso=btl-tcp --enable-mca-static=btl-tcp

  • --enable-mca-no-build=LIST: Comma-separated list of <framework>-<component> pairs that will not be built. For example, --enable-mca-no-build=btl-portals,oob-ud will disable building the portals BTL and the ud OOB component.

  • --disable-show-load-errors-by-default: Set the default value of the mca_base_component_show_load_errors MCA variable: the --enable form of this option sets the MCA variable to true, the --disable form sets the MCA variable to false. The MCA mca_base_component_show_load_errors variable can still be overridden at run time via the usual MCA-variable-setting mechanisms; this configure option simply sets the default value.

    The --disable form of this option is intended for Open MPI packagers who tend to enable support for many different types of networks and systems in their packages. For example, consider a packager who includes support for both the FOO and BAR networks in their Open MPI package, both of which require support libraries (libFOO.so and libBAR.so). If an end user only has BAR hardware, they likely only have libBAR.so available on their systems -- not libFOO.so. Disabling load errors by default will prevent the user from seeing potentially confusing warnings about the FOO components failing to load because libFOO.so is not available on their systems.

    Conversely, system administrators tend to build an Open MPI that is targeted at their specific environment, and contains few (if any) components that are not needed. In such cases, they might want their users to be warned that the FOO network components failed to load (e.g., if libFOO.so was mistakenly unavailable), because Open MPI may otherwise silently failover to a slower network path for MPI traffic.

  • --with-platform=FILE: Load configure options for the build from FILE. Options on the command line that are not in FILE are also used. Options on the command line and in FILE are replaced by what is in FILE.

  • --with-libmpi-name=STRING: Replace libmpi.* and libmpi_FOO.* (where FOO is one of the fortran supporting libraries installed in lib) with libSTRING.* and libSTRING_FOO.*. This is provided as a convenience mechanism for third-party packagers of Open MPI that might want to rename these libraries for their own purposes. This option is not intended for typical users of Open MPI.

Networking support / options

  • --with-fca=DIR: Specify the directory where the Mellanox FCA library and header files are located.

    FCA is the support library for Mellanox switches and HCAs.

  • --with-hcoll=DIR: Specify the directory where the Mellanox hcoll library and header files are located. This option is generally only necessary if the hcoll headers and libraries are not in default compiler/linker search paths.

    hcoll is the support library for MPI collective operation offload on Mellanox ConnectX-3 HCAs (and later).

  • --with-knem=DIR: Specify the directory where the knem libraries and header files are located. This option is generally only necessary if the knem headers and libraries are not in default compiler/linker search paths.

    knem is a Linux kernel module that allows direct process-to-process memory copies (optionally using hardware offload), potentially increasing bandwidth for large messages sent between messages on the same server. See the Knem web site for details.

  • --with-libfabric=DIR: Specify the directory where the OpenFabrics Interfaces libfabric library and header files are located. This option is generally only necessary if the libfabric headers and libraries are not in default compiler/linker search paths.

    Libfabric is the support library for OpenFabrics Interfaces-based network adapters, such as Cisco usNIC, Intel True Scale PSM, Cray uGNI, etc.

  • --with-libfabric-libdir=DIR: Look in directory for the libfabric libraries. By default, Open MPI will look in DIR/lib and DIR/lib64, which covers most cases. This option is only needed for special configurations.

  • --with-portals4=DIR: Specify the directory where the Portals4 libraries and header files are located. This option is generally only necessary if the Portals4 headers and libraries are not in default compiler/linker search paths.

    Portals is a low-level network API for high-performance networking on high-performance computing systems developed by Sandia National Laboratories, Intel Corporation, and the University of New Mexico. The Portals 4 Reference Implementation is a complete implementation of Portals 4, with transport over InfiniBand verbs and UDP.

  • --with-portals4-libdir=DIR: Location of libraries to link with for Portals4 support.

  • --with-portals4-max-md-size=SIZE and --with-portals4-max-va-size=SIZE: Set configuration values for Portals 4

  • --with-psm=<directory>: Specify the directory where the QLogic InfiniPath / Intel True Scale PSM library and header files are located. This option is generally only necessary if the PSM headers and libraries are not in default compiler/linker search paths.

    PSM is the support library for QLogic InfiniPath and Intel TrueScale network adapters.

  • --with-psm-libdir=DIR: Look in directory for the PSM libraries. By default, Open MPI will look in DIR/lib and DIR/lib64, which covers most cases. This option is only needed for special configurations.

  • --with-psm2=DIR: Specify the directory where the Intel Omni-Path PSM2 library and header files are located. This option is generally only necessary if the PSM2 headers and libraries are not in default compiler/linker search paths.

    PSM is the support library for Intel Omni-Path network adapters.

  • --with-psm2-libdir=DIR: Look in directory for the PSM2 libraries. By default, Open MPI will look in DIR/lib and DIR/lib64, which covers most cases. This option is only needed for special configurations.

  • --with-ucx=DIR: Specify the directory where the UCX libraries and header files are located. This option is generally only necessary if the UCX headers and libraries are not in default compiler/linker search paths.

  • --with-ucx-libdir=DIR: Look in directory for the UCX libraries. By default, Open MPI will look in DIR/lib and DIR/lib64, which covers most cases. This option is only needed for special configurations.

  • --with-usnic: Abort configure if Cisco usNIC support cannot be built.

Run-time system support

  • --enable-mpirun-prefix-by-default: This option forces the mpirun command to always behave as if --prefix $prefix was present on the command line (where $prefix is the value given to the --prefix option to configure). This prevents most rsh/ssh-based users from needing to modify their shell startup files to set the PATH and/or LD_LIBRARY_PATH for Open MPI on remote nodes. Note, however, that such users may still desire to set PATH -- perhaps even in their shell startup files -- so that executables such as mpicc and mpirun can be found without needing to type long path names.

  • --with-alps: Force the building of for the Cray Alps run-time environment. If Alps support cannot be found, configure will abort.

  • --with-lsf=DIR: Specify the directory where the LSF libraries and header files are located. This option is generally only necessary if the LSF headers and libraries are not in default compiler/linker search paths.

    LSF is a resource manager system, frequently used as a batch scheduler in HPC systems.

  • --with-lsf-libdir=DIR: Look in directory for the LSF libraries. By default, Open MPI will look in DIR/lib and DIR/lib64, which covers most cases. This option is only needed for special configurations.

  • --with-slurm: Force the building of SLURM scheduler support.

  • --with-sge: Specify to build support for the Oracle Grid Engine (OGE) resource manager and/or the Open Grid Engine. OGE support is disabled by default; this option must be specified to build OMPI's OGE support.

    The Oracle Grid Engine (OGE) and open Grid Engine packages are resource manager systems, frequently used as a batch scheduler in HPC systems. It used to be called the "Sun Grid Engine", which is why the option is still named --with-sge.

  • --with-tm=DIR: Specify the directory where the TM libraries and header files are located. This option is generally only necessary if the TM headers and libraries are not in default compiler/linker search paths.

    TM is the support library for the Torque and PBS Pro resource manager systems, both of which are frequently used as a batch scheduler in HPC systems.

Miscellaneous support libraries

  • --with-libevent(=VALUE) This option specifies where to find the libevent support headers and library. The following VALUEs are permitted:

    • internal: Use Open MPI's internal copy of libevent.
    • external: Use an external Libevent installation (rely on default compiler and linker paths to find it)
    • <no value>: Same as internal.
    • DIR: Specify the location of a specific libevent installation to use

    By default (or if --with-libevent is specified with no VALUE), Open MPI will build and use the copy of libevent that it has in its source tree. However, if the VALUE is external, Open MPI will look for the relevant libevent header file and library in default compiler / linker locations. Or, VALUE can be a directory tree where the libevent header file and library can be found. This option allows operating systems to include Open MPI and use their default libevent installation instead of Open MPI's bundled libevent.

    libevent is a support library that provides event-based processing, timers, and signal handlers. Open MPI requires libevent to build; passing --without-libevent will cause configure to abort.

  • --with-libevent-libdir=DIR: Look in directory for the libevent libraries. This option is only usable when building Open MPI against an external libevent installation. Just like other --with-FOO-libdir configure options, this option is only needed for special configurations.

  • --with-hwloc(=VALUE): hwloc is a support library that provides processor and memory affinity information for NUMA platforms. It is required by Open MPI. Therefore, specifying --with-hwloc=no (or --without-hwloc) is disallowed.

    By default (i.e., if --with-hwloc is not specified, or if --with-hwloc is specified without a value), Open MPI will first try to find/use an hwloc installation on the current system. If Open MPI cannot find one, it will fall back to build and use the internal copy of hwloc included in the Open MPI source tree.

    Alternatively, the --with-hwloc option can be used to specify where to find the hwloc support headers and library. The following VALUEs are permitted:

    • internal: Only use Open MPI's internal copy of hwloc.
    • external: Only use an external hwloc installation (rely on default compiler and linker paths to find it).
    • DIR: Only use the specific hwloc installation found in the specified directory.
  • --with-hwloc-libdir=DIR: Look in directory for the hwloc libraries. This option is only usable when building Open MPI against an external hwloc installation. Just like other --with-FOO-libdir configure options, this option is only needed for special configurations.

  • --disable-hwloc-pci: Disable building hwloc's PCI device-sensing capabilities. On some platforms (e.g., SusE 10 SP1, x86-64), the libpci support library is broken. Open MPI's configure script should usually detect when libpci is not usable due to such brokenness and turn off PCI support, but there may be cases when configure mistakenly enables PCI support in the presence of a broken libpci. These cases may result in make failing with warnings about relocation symbols in libpci. The --disable-hwloc-pci switch can be used to force Open MPI to not build hwloc's PCI device-sensing capabilities in these cases.

    Similarly, if Open MPI incorrectly decides that libpci is broken, you can force Open MPI to build hwloc's PCI device-sensing capabilities by using --enable-hwloc-pci.

    hwloc can discover PCI devices and locality, which can be useful for Open MPI in assigning message passing resources to MPI processes.

  • --with-libltdl=DIR: Specify the directory where the GNU Libtool libltdl libraries and header files are located. This option is generally only necessary if the libltdl headers and libraries are not in default compiler/linker search paths.

    Note that this option is ignored if --disable-dlopen is specified.

  • --disable-libompitrace: Disable building the simple libompitrace library (see note above about libompitrace)

  • --with-valgrind(=DIR): Directory where the valgrind software is installed. If Open MPI finds Valgrind's header files, it will include additional support for Valgrind's memory-checking debugger.

    Specifically, it will eliminate a lot of false positives from running Valgrind on MPI applications. There is a minor performance penalty for enabling this option.

MPI Functionality

  • --with-mpi-param-check(=VALUE): Whether or not to check MPI function parameters for errors at runtime. The following VALUEs are permitted:

    • always: MPI function parameters are always checked for errors
    • never: MPI function parameters are never checked for errors
    • runtime: Whether MPI function parameters are checked depends on the value of the MCA parameter mpi_param_check (default: yes).
    • yes: Synonym for "always" (same as --with-mpi-param-check).
    • no: Synonym for "never" (same as --without-mpi-param-check).

    If --with-mpi-param is not specified, runtime is the default.

  • --disable-mpi-thread-multiple: Disable the MPI thread level MPI_THREAD_MULTIPLE (it is enabled by default).

  • --enable-mpi-java: Enable building of an EXPERIMENTAL Java MPI interface (disabled by default). You may also need to specify --with-jdk-dir, --with-jdk-bindir, and/or --with-jdk-headers. See README.JAVA.md for details.

    Note that this Java interface is INCOMPLETE (meaning: it does not support all MPI functionality) and LIKELY TO CHANGE. The Open MPI developers would very much like to hear your feedback about this interface. See README.JAVA.md for more details.

  • --enable-mpi-fortran(=VALUE): By default, Open MPI will attempt to build all 3 Fortran bindings: mpif.h, the mpi module, and the mpi_f08 module. The following VALUEs are permitted:

    • all: Synonym for yes.
    • yes: Attempt to build all 3 Fortran bindings; skip any binding that cannot be built (same as --enable-mpi-fortran).
    • mpifh: Only build mpif.h support.
    • usempi: Only build mpif.h and mpi module support.
    • usempif08: Build mpif.h, mpi module, and mpi_f08 module support.
    • none: Synonym for no.
    • no: Do not build any MPI Fortran support (same as --disable-mpi-fortran). This is mutually exclusive with building the OpenSHMEM Fortran interface.
  • --enable-mpi-ext(=LIST): Enable Open MPI's non-portable API extensions. LIST is a comma-delmited list of extensions. If no LIST is specified, all of the extensions are enabled.

    See the "Open MPI API Extensions" section for more details.

  • --disable-mpi-io: Disable built-in support for MPI-2 I/O, likely because an externally-provided MPI I/O package will be used. Default is to use the internal framework system that uses the ompio component and a specially modified version of ROMIO that fits inside the romio component

  • --disable-io-romio: Disable the ROMIO MPI-IO component

  • --with-io-romio-flags=FLAGS: Pass FLAGS to the ROMIO distribution configuration script. This option is usually only necessary to pass parallel-filesystem-specific preprocessor/compiler/linker flags back to the ROMIO system.

  • --disable-io-ompio: Disable the ompio MPI-IO component

  • --enable-sparse-groups: Enable the usage of sparse groups. This would save memory significantly especially if you are creating large communicators. (Disabled by default)

OpenSHMEM Functionality

  • --disable-oshmem: Disable building the OpenSHMEM implementation (by default, it is enabled).

  • --disable-oshmem-fortran: Disable building only the Fortran OpenSHMEM bindings. Please see the "Compiler Notes" section herein which contains further details on known issues with various Fortran compilers.

Miscellaneous Functionality

  • --without-memory-manager: Disable building Open MPI's memory manager. Open MPI's memory manager is usually built on Linux based platforms, and is generally only used for optimizations with some OpenFabrics-based networks (it is not necessary for OpenFabrics networks, but some performance loss may be observed without it).

    However, it may be necessary to disable the memory manager in order to build Open MPI statically.

  • --with-ft=TYPE: Specify the type of fault tolerance to enable. Options: LAM (LAM/MPI-like), cr (Checkpoint/Restart). Fault tolerance support is disabled unless this option is specified.

  • --enable-peruse: Enable the PERUSE MPI data analysis interface.

  • --enable-heterogeneous: Enable support for running on heterogeneous clusters where data types are equivalent sizes across nodes, but may have differing endian representations. Heterogeneous support is disabled by default because it imposes a minor performance penalty.

    Note that the MPI standard does not guarantee that all heterogeneous communication will function properly, especially when the conversion between the different representations leads to loss of accuracy or range. For example, if a message with a 16-bit integer datatype is sent with value 0x10000 to a receiver where the same integer datatype is only 8 bits, the value will be truncated at the receiver. Similarly, problems can occur if a floating point datatype in one MPI process uses X1 bits for its mantissa and Y1 bits for its exponent, but the same floating point datatype in another MPI process uses X2 and Y2 bits, respectively (where X1 != X2 and/or Y1 != Y2). Type size differences like this can lead to unexpected behavior.

    Open MPI's heterogeneous support correctly handles endian differences between datatype representations that are otherwise compatible.

  • --with-wrapper-cflags=CFLAGS

  • --with-wrapper-cxxflags=CXXFLAGS

  • --with-wrapper-fflags=FFLAGS

  • --with-wrapper-fcflags=FCFLAGS

  • --with-wrapper-ldflags=LDFLAGS

  • --with-wrapper-libs=LIBS: Add the specified flags to the default flags that are used in Open MPI's "wrapper" compilers (e.g., mpicc -- see below for more information about Open MPI's wrapper compilers). By default, Open MPI's wrapper compilers use the same compilers used to build Open MPI and specify a minimum set of additional flags that are necessary to compile/link MPI applications. These configure options give system administrators the ability to embed additional flags in OMPI's wrapper compilers (which is a local policy decision). The meanings of the different flags are:

    CFLAGS: Flags passed by the mpicc wrapper to the C compiler CXXFLAGS: Flags passed by the mpic++ wrapper to the C++ compiler FCFLAGS: Flags passed by the mpifort wrapper to the Fortran compiler LDFLAGS: Flags passed by all the wrappers to the linker LIBS: Flags passed by all the wrappers to the linker

    There are other ways to configure Open MPI's wrapper compiler behavior; see the Open MPI FAQ for more information.

There are many other options available -- see ./configure --help.

Changing the compilers that Open MPI uses to build itself uses the standard Autoconf mechanism of setting special environment variables either before invoking configure or on the configure command line. The following environment variables are recognized by configure:

  • CC: C compiler to use
  • CFLAGS: Compile flags to pass to the C compiler
  • CPPFLAGS: Preprocessor flags to pass to the C compiler
  • CXX: C++ compiler to use
  • CXXFLAGS: Compile flags to pass to the C++ compiler
  • CXXCPPFLAGS: Preprocessor flags to pass to the C++ compiler
  • FC: Fortran compiler to use
  • FCFLAGS: Compile flags to pass to the Fortran compiler
  • LDFLAGS: Linker flags to pass to all compilers
  • LIBS: Libraries to pass to all compilers (it is rarely necessary for users to need to specify additional LIBS)
  • PKG_CONFIG: Path to the pkg-config utility

For example:

shell$ ./configure CC=mycc CXX=myc++ FC=myfortran ...

NOTE: We generally suggest using the above command line form for setting different compilers (vs. setting environment variables and then invoking ./configure). The above form will save all variables and values in the config.log file, which makes post-mortem analysis easier if problems occur.

Note that if you intend to compile Open MPI with a make other than the default one in your PATH, then you must either set the $MAKE environment variable before invoking Open MPI's configure script, or pass MAKE=your_make_prog to configure. For example:

shell$ ./configure MAKE=/path/to/my/make ...

This could be the case, for instance, if you have a shell alias for make, or you always type gmake out of habit. Failure to tell configure which non-default make you will use to compile Open MPI can result in undefined behavior (meaning: don't do that).

Note that you may also want to ensure that the value of LD_LIBRARY_PATH is set appropriately (or not at all) for your build (or whatever environment variable is relevant for your operating system). For example, some users have been tripped up by setting to use a non-default Fortran compiler via the FC environment variable, but then failing to set LD_LIBRARY_PATH to include the directory containing that non-default Fortran compiler's support libraries. This causes Open MPI's configure script to fail when it tries to compile / link / run simple Fortran programs.

It is required that the compilers specified be compile and link compatible, meaning that object files created by one compiler must be able to be linked with object files from the other compilers and produce correctly functioning executables.

Open MPI supports all the make targets that are provided by GNU Automake, such as:

  • all: build the entire Open MPI package
  • install: install Open MPI
  • uninstall: remove all traces of Open MPI from the $prefix
  • clean: clean out the build tree

Once Open MPI has been built and installed, it is safe to run make clean and/or remove the entire build tree.

VPATH and parallel builds are fully supported.

Generally speaking, the only thing that users need to do to use Open MPI is ensure that PREFIX/bin is in their PATH and PREFIX/lib is in their LD_LIBRARY_PATH. Users may need to ensure to set the PATH and LD_LIBRARY_PATH in their shell setup files (e.g., .bashrc, .cshrc) so that non-interactive rsh/ssh-based logins will be able to find the Open MPI executables.

Open MPI Version Numbers and Binary Compatibility

Open MPI has two sets of version numbers that are likely of interest to end users / system administrator:

  1. Software version number
  2. Shared library version numbers

Both are predicated on Open MPI's definition of "backwards compatibility."

NOTE: The version numbering conventions were changed with the release of v1.10.0. Most notably, Open MPI no longer uses an "odd/even" release schedule to indicate feature development vs. stable releases. See the README in releases prior to v1.10.0 for more information (e.g., https://github.com/open-mpi/ompi/blob/v1.8/README#L1392-L1475).

Backwards Compatibility

Open MPI version Y is backwards compatible with Open MPI version X (where Y>X) if users can:

  • Compile an MPI/OpenSHMEM application with version X, mpirun/oshrun it with version Y, and get the same user-observable behavior.
  • Invoke ompi_info with the same CLI options in versions X and Y and get the same user-observable behavior.

Note that this definition encompasses several things:

  • Application Binary Interface (ABI)
  • MPI / OpenSHMEM run time system
  • mpirun / oshrun command line options
  • MCA parameter names / values / meanings

However, this definition only applies when the same version of Open MPI is used with all instances of the runtime and MPI / OpenSHMEM processes in a single MPI job. If the versions are not exactly the same everywhere, Open MPI is not guaranteed to work properly in any scenario.

Backwards compatibility tends to work best when user applications are dynamically linked to one version of the Open MPI / OSHMEM libraries, and can be updated at run time to link to a new version of the Open MPI / OSHMEM libraries.

For example, if an MPI / OSHMEM application links statically against the libraries from Open MPI vX, then attempting to launch that application with mpirun / oshrun from Open MPI vY is not guaranteed to work (because it is mixing vX and vY of Open MPI in a single job).

Similarly, if using a container technology that internally bundles all the libraries from Open MPI vX, attempting to launch that container with mpirun / oshrun from Open MPI vY is not guaranteed to work.

Software Version Number

Official Open MPI releases use the common "A.B.C" version identifier format. Each of the three numbers has a specific meaning:

  • Major: The major number is the first integer in the version string Changes in the major number typically indicate a significant change in the code base and/or end-user functionality, and also indicate a break from backwards compatibility. Specifically: Open MPI releases with different major version numbers are not backwards compatibale with each other.

    CAVEAT: This rule does not extend to versions prior to v1.10.0. Specifically: v1.10.x is not guaranteed to be backwards compatible with other v1.x releases.

  • Minor: The minor number is the second integer in the version string. Changes in the minor number indicate a user-observable change in the code base and/or end-user functionality. Backwards compatibility will still be preserved with prior releases that have the same major version number (e.g., v2.5.3 is backwards compatible with v2.3.1).

  • Release: The release number is the third integer in the version string. Changes in the release number typically indicate a bug fix in the code base and/or end-user functionality. For example, if there is a release that only contains bug fixes and no other user-observable changes or new features, only the third integer will be increased (e.g., from v4.3.0 to v4.3.1).

The "A.B.C" version number may optionally be followed by a Quantifier:

  • Quantifier: Open MPI version numbers sometimes have an arbitrary string affixed to the end of the version number. Common strings include:
    • aX: Indicates an alpha release. X is an integer indicating the number of the alpha release (e.g., v1.10.3a5 indicates the 5th alpha release of version 1.10.3).
    • bX: Indicates a beta release. X is an integer indicating the number of the beta release (e.g., v1.10.3b3 indicates the 3rd beta release of version 1.10.3).
    • rcX: Indicates a release candidate. X is an integer indicating the number of the release candidate (e.g., v1.10.3rc4 indicates the 4th release candidate of version 1.10.3).

Nightly development snapshot tarballs use a different version number scheme; they contain three distinct values:

  • The git branch name from which the tarball was created.
  • The date/timestamp, in YYYYMMDDHHMM format.
  • The hash of the git commit from which the tarball was created.

For example, a snapshot tarball filename of openmpi-v2.x-201703070235-e4798fb.tar.gz indicates that this tarball was created from the v2.x branch, on March 7, 2017, at 2:35am GMT, from git hash e4798fb.

Shared Library Version Number

The GNU Libtool official documentation details how the versioning scheme works. The quick version is that the shared library versions are a triple of integers: (current,revision,age), or c:r:a. This triple is not related to the Open MPI software version number. There are six simple rules for updating the values (taken almost verbatim from the Libtool docs):

  1. Start with version information of 0:0:0 for each shared library.
  2. Update the version information only immediately before a public release of your software. More frequent updates are unnecessary, and only guarantee that the current interface number gets larger faster.
  3. If the library source code has changed at all since the last update, then increment revision (c:r:a becomes c:r+1:a).
  4. If any interfaces have been added, removed, or changed since the last update, increment current, and set revision to 0.
  5. If any interfaces have been added since the last public release, then increment age.
  6. If any interfaces have been removed since the last public release, then set age to 0.

Here's how we apply those rules specifically to Open MPI:

  1. The above rules do not apply to MCA components (a.k.a. "plugins"); MCA component .so versions stay unspecified.
  2. The above rules apply exactly as written to the following libraries starting with Open MPI version v1.5 (prior to v1.5, libopen-pal and libopen-rte were still at 0:0:0 for reasons discussed in bug ticket #2092 https://svn.open-mpi.org/trac/ompi/ticket/2092):
    • libopen-rte
    • libopen-pal
    • libmca_common_*
  3. The following libraries use a slightly modified version of the above rules: rules 4, 5, and 6 only apply to the official MPI and OpenSHMEM interfaces (functions, global variables). The rationale for this decision is that the vast majority of our users only care about the official/public MPI/OpenSHMEM interfaces; we therefore want the .so version number to reflect only changes to the official MPI/OpenSHMEM APIs. Put simply: non-MPI/OpenSHMEM API / internal changes to the MPI-application-facing libraries are irrelevant to pure MPI/OpenSHMEM applications.
    • libmpi
    • libmpi_mpifh
    • libmpi_usempi_tkr
    • libmpi_usempi_ignore_tkr
    • libmpi_usempif08
    • libmpi_cxx
    • libmpi_java
    • liboshmem

Checking Your Open MPI Installation

The ompi_info command can be used to check the status of your Open MPI installation (located in PREFIX/bin/ompi_info). Running it with no arguments provides a summary of information about your Open MPI installation.

Note that the ompi_info command is extremely helpful in determining which components are installed as well as listing all the run-time settable parameters that are available in each component (as well as their default values).

The following options may be helpful:

  • --all: Show a lot of information about your Open MPI installation.
  • --parsable: Display all the information in an easily grep/cut/awk/sed-able format.
  • --param FRAMEWORK COMPONENT: A FRAMEWORK value of all and a COMPONENT value of all will show all parameters to all components. Otherwise, the parameters of all the components in a specific framework, or just the parameters of a specific component can be displayed by using an appropriate FRAMEWORK and/or COMPONENT name.
  • --level LEVEL: By default, ompi_info only shows "Level 1" MCA parameters -- parameters that can affect whether MPI processes can run successfully or not (e.g., determining which network interfaces to use). The --level option will display all MCA parameters from level 1 to LEVEL (the max LEVEL value is 9). Use ompi_info --param FRAMEWORK COMPONENT --level 9 to see all MCA parameters for a given component. See "The Modular Component Architecture (MCA)" section, below, for a fuller explanation.

Changing the values of these parameters is explained in the "The Modular Component Architecture (MCA)" section, below.

When verifying a new Open MPI installation, we recommend running six tests:

  1. Use mpirun to launch a non-MPI program (e.g., hostname or uptime) across multiple nodes.
  2. Use mpirun to launch a trivial MPI program that does no MPI communication (e.g., the hello_c program in the examples/ directory in the Open MPI distribution).
  3. Use mpirun to launch a trivial MPI program that sends and receives a few MPI messages (e.g., the ring_c program in the examples/ directory in the Open MPI distribution).
  4. Use oshrun to launch a non-OpenSHMEM program across multiple nodes.
  5. Use oshrun to launch a trivial MPI program that does no OpenSHMEM communication (e.g., hello_shmem.c program in the examples/ directory in the Open MPI distribution.)
  6. Use oshrun to launch a trivial OpenSHMEM program that puts and gets a few messages (e.g., the ring_shmem.c in the examples/ directory in the Open MPI distribution.)

If you can run all six of these tests successfully, that is a good indication that Open MPI built and installed properly.

Open MPI API Extensions

Open MPI contains a framework for extending the MPI API that is available to applications. Each extension is usually a standalone set of functionality that is distinct from other extensions (similar to how Open MPI's plugins are usually unrelated to each other). These extensions provide new functions and/or constants that are available to MPI applications.

WARNING: These extensions are neither standard nor portable to other MPI implementations!

Compiling the extensions

Open MPI extensions are all enabled by default; they can be disabled via the --disable-mpi-ext command line switch.

Since extensions are meant to be used by advanced users only, this file does not document which extensions are available or what they do. Look in the ompi/mpiext/ directory to see the extensions; each subdirectory of that directory contains an extension. Each has a README file that describes what it does.

Using the extensions

To reinforce the fact that these extensions are non-standard, you must include a separate header file after <mpi.h> to obtain the function prototypes, constant declarations, etc. For example:

#include <mpi.h>
#if defined(OPEN_MPI) && OPEN_MPI
#include <mpi-ext.h>
#endif

int main() {
    MPI_Init(NULL, NULL);

#if defined(OPEN_MPI) && OPEN_MPI
    {
        char ompi_bound[OMPI_AFFINITY_STRING_MAX];
        char current_binding[OMPI_AFFINITY_STRING_MAX];
        char exists[OMPI_AFFINITY_STRING_MAX];
        OMPI_Affinity_str(OMPI_AFFINITY_LAYOUT_FMT, ompi_bound,
                          current_bindings, exists);
    }
#endif
    MPI_Finalize();
    return 0;
}

Notice that the Open MPI-specific code is surrounded by the #if statement to ensure that it is only ever compiled by Open MPI.

The Open MPI wrapper compilers (mpicc and friends) should automatically insert all relevant compiler and linker flags necessary to use the extensions. No special flags or steps should be necessary compared to "normal" MPI applications.

Compiling Open MPI Applications

Open MPI provides "wrapper" compilers that should be used for compiling MPI and OpenSHMEM applications:

  • C: mpicc, oshcc
  • C++: mpiCC, oshCC (or mpic++ if your filesystem is case-insensitive)
  • Fortran: mpifort, oshfort

For example:

shell$ mpicc hello_world_mpi.c -o hello_world_mpi -g
shell$

For OpenSHMEM applications:

shell$ oshcc hello_shmem.c -o hello_shmem -g
shell$

All the wrapper compilers do is add a variety of compiler and linker flags to the command line and then invoke a back-end compiler. To be specific: the wrapper compilers do not parse source code at all; they are solely command-line manipulators, and have nothing to do with the actual compilation or linking of programs. The end result is an MPI executable that is properly linked to all the relevant libraries.

Customizing the behavior of the wrapper compilers is possible (e.g., changing the compiler [not recommended] or specifying additional compiler/linker flags); see the Open MPI FAQ for more information.

Alternatively, Open MPI also installs pkg-config(1) configuration files under $libdir/pkgconfig. If pkg-config is configured to find these files, then compiling / linking Open MPI programs can be performed like this:

shell$ gcc hello_world_mpi.c -o hello_world_mpi -g \
            `pkg-config ompi-c --cflags --libs`
shell$

Open MPI supplies multiple pkg-config(1) configuration files; one for each different wrapper compiler (language):

  • ompi: Synonym for ompi-c; Open MPI applications using the C MPI bindings
  • ompi-c: Open MPI applications using the C MPI bindings
  • ompi-cxx: Open MPI applications using the C MPI bindings
  • ompi-fort: Open MPI applications using the Fortran MPI bindings

The following pkg-config(1) configuration files may be installed, depending on which command line options were specified to Open MPI's configure script. They are not necessary for MPI applications, but may be used by applications that use Open MPI's lower layer support libraries.

  • opal: Open Portable Access Layer applications

Running Open MPI Applications

Open MPI supports both mpirun and mpiexec (they are exactly equivalent) to launch MPI applications. For example:

shell$ mpirun -np 2 hello_world_mpi
or
shell$ mpiexec -np 1 hello_world_mpi : -np 1 hello_world_mpi

are equivalent.

The rsh launcher (which defaults to using ssh) accepts a --hostfile parameter (the option --machinefile is equivalent); you can specify a --hostfile parameter indicating a standard mpirun-style hostfile (one hostname per line):

shell$ mpirun --hostfile my_hostfile -np 2 hello_world_mpi

If you intend to run more than one process on a node, the hostfile can use the "slots" attribute. If "slots" is not specified, a count of 1 is assumed. For example, using the following hostfile:

shell$ cat my_hostfile
node1.example.com
node2.example.com
node3.example.com slots=2
node4.example.com slots=4
shell$ mpirun --hostfile my_hostfile -np 8 hello_world_mpi

will launch MPI_COMM_WORLD rank 0 on node1, rank 1 on node2, ranks 2 and 3 on node3, and ranks 4 through 7 on node4.

Other starters, such as the resource manager / batch scheduling environments, do not require hostfiles (and will ignore the hostfile if it is supplied). They will also launch as many processes as slots have been allocated by the scheduler if no "-np" argument has been provided. For example, running a SLURM job with 8 processors:

shell$ salloc -n 8 mpirun a.out

The above command will reserve 8 processors and run 1 copy of mpirun, which will, in turn, launch 8 copies of a.out in a single MPI_COMM_WORLD on the processors that were allocated by SLURM.

Note that the values of component parameters can be changed on the mpirun / mpiexec command line. This is explained in the section below, "The Modular Component Architecture (MCA)".

Open MPI supports oshrun to launch OpenSHMEM applications. For example:

shell$ oshrun -np 2 hello_world_oshmem

OpenSHMEM applications may also be launched directly by resource managers such as SLURM. For example, when OMPI is configured --with-pmix and --with-slurm, one may launch OpenSHMEM applications via srun:

shell$ srun -N 2 hello_world_oshmem

The Modular Component Architecture (MCA)

The MCA is the backbone of Open MPI -- most services and functionality are implemented through MCA components.

MPI layer frameworks

Here is a list of all the component frameworks in the MPI layer of Open MPI:

  • bml: BTL management layer
  • coll: MPI collective algorithms
  • fbtl: file byte transfer layer: abstraction for individual blocking and non-blocking read and write operations
  • fcoll: Collective read and write operations for MPI I/O.
  • fs: File system functions for MPI I/O.
  • hook: Make calls at various points of MPI process life-cycle.
  • io: MPI I/O
  • mtl: Matching transport layer, used for MPI point-to-point messages on some types of networks
  • op: Back end computations for intrinsic MPI_Op operators
  • osc: MPI one-sided communications
  • pml: MPI point-to-point management layer
  • part: MPI Partitioned communication.
  • sharedfp: shared file pointer operations for MPI I/O
  • topo: MPI topology routines
  • vprotocol: Protocols for the "v" PML

OpenSHMEM component frameworks

  • atomic: OpenSHMEM atomic operations
  • memheap: OpenSHMEM memory allocators that support the PGAS memory model
  • scoll: OpenSHMEM collective operations
  • spml: OpenSHMEM "pml-like" layer: supports one-sided, point-to-point operations
  • sshmem: OpenSHMEM shared memory backing facility

Miscellaneous frameworks:

  • allocator: Memory allocator
  • backtrace: Debugging call stack backtrace support
  • btl: Point-to-point Byte Transfer Layer
  • dl: Dynamic loading library interface
  • hwloc: Hardware locality (hwloc) versioning support
  • if: OS IP interface support
  • installdirs: Installation directory relocation services
  • memchecker: Run-time memory checking
  • memcpy: Memory copy support
  • memory: Memory management hooks
  • mpool: Memory pooling
  • patcher: Symbol patcher hooks
  • pmix: Process management interface (exascale)
  • rcache: Memory registration cache
  • reachable: Reachability matrix between endpoints of a given pair of hosts
  • shmem: Shared memory support (NOT related to OpenSHMEM)
  • smsc: Shared Memory Single Copy
  • threads: Thread management and support.
  • timer: High-resolution timers

Back-end run-time environment (PRTE):

See: https://github.com/openpmix/prrte

Framework notes

Each framework typically has one or more components that are used at run-time. For example, the btl framework is used by the MPI layer to send bytes across different types underlying networks. The tcp btl, for example, sends messages across TCP-based networks; the ucx pml sends messages across InfiniBand-based networks.

Each component typically has some tunable parameters that can be changed at run-time. Use the ompi_info command to check a component to see what its tunable parameters are. For example:

shell$ ompi_info --param btl tcp

shows some of the parameters (and default values) for the tcp btl component (use --level to show all the parameters; see below).

Note that ompi_info only shows a small number a component's MCA parameters by default. Each MCA parameter has a "level" value from 1 to 9, corresponding to the MPI-3 MPI_T tool interface levels. In Open MPI, we have interpreted these nine levels as three groups of three:

  1. End user / basic
  2. End user / detailed
  3. End user / all
  4. Application tuner / basic
  5. Application tuner / detailed
  6. Application tuner / all
  7. MPI/OpenSHMEM developer / basic
  8. MPI/OpenSHMEM developer / detailed
  9. MPI/OpenSHMEM developer / all

Here's how the three sub-groups are defined:

  1. End user: Generally, these are parameters that are required for correctness, meaning that someone may need to set these just to get their MPI/OpenSHMEM application to run correctly.
  2. Application tuner: Generally, these are parameters that can be used to tweak MPI application performance.
  3. MPI/OpenSHMEM developer: Parameters that either don't fit in the other two, or are specifically intended for debugging / development of Open MPI itself.

Each sub-group is broken down into three classifications:

  1. Basic: For parameters that everyone in this category will want to see.
  2. Detailed: Parameters that are useful, but you probably won't need to change them often.
  3. All: All other parameters -- probably including some fairly esoteric parameters.

To see all available parameters for a given component, specify that ompi_info should use level 9:

shell$ ompi_info --param btl tcp --level 9

These values can be overridden at run-time in several ways. At run-time, the following locations are examined (in order) for new values of parameters:

  1. PREFIX/etc/openmpi-mca-params.conf: This file is intended to set any system-wide default MCA parameter values -- it will apply, by default, to all users who use this Open MPI installation. The default file that is installed contains many comments explaining its format.

  2. $HOME/.openmpi/mca-params.conf: If this file exists, it should be in the same format as PREFIX/etc/openmpi-mca-params.conf. It is intended to provide per-user default parameter values.

  3. environment variables of the form OMPI_MCA_<name> set equal to a VALUE:

    Where <name> is the name of the parameter. For example, set the variable named OMPI_MCA_btl_tcp_frag_size to the value 65536 (Bourne-style shells):

    shell$ OMPI_MCA_btl_tcp_frag_size=65536
    shell$ export OMPI_MCA_btl_tcp_frag_size
    
  4. the mpirun/oshrun command line: --mca NAME VALUE

    Where is the name of the parameter. For example:

    shell$ mpirun --mca btl_tcp_frag_size 65536 -np 2 hello_world_mpi
    

These locations are checked in order. For example, a parameter value passed on the mpirun command line will override an environment variable; an environment variable will override the system-wide defaults.

Each component typically activates itself when relevant. For example, the usNIC component will detect that usNIC devices are present and will automatically be used for MPI communications. The SLURM component will automatically detect when running inside a SLURM job and activate itself. And so on.

Components can be manually activated or deactivated if necessary, of course. The most common components that are manually activated, deactivated, or tuned are the btl components -- components that are used for MPI point-to-point communications on many types common networks.

For example, to only activate the tcp and self (process loopback) components are used for MPI communications, specify them in a comma-delimited list to the btl MCA parameter:

shell$ mpirun --mca btl tcp,self hello_world_mpi

To add shared memory support, add sm into the command-delimited list (list order does not matter):

shell$ mpirun --mca btl tcp,sm,self hello_world_mpi

(there used to be a vader BTL for shared memory support; it was renamed to sm in Open MPI v5.0.0, but the alias vader still works as well)

To specifically deactivate a specific component, the comma-delimited list can be prepended with a ^ to negate it:

shell$ mpirun --mca btl ^tcp hello_mpi_world

The above command will use any other btl component other than the tcp component.

Questions? Problems?

Found a bug? Got a question? Want to make a suggestion? Want to contribute to Open MPI? Please let us know!

When submitting questions and problems, be sure to include as much extra information as possible. See the community help web page for details on all the information that we request in order to provide assistance:

The best way to report bugs, send comments, or ask questions is to sign up on the user's and/or developer's mailing list (for user-level and developer-level questions; when in doubt, send to the user's list):

Because of spam, only subscribers are allowed to post to these lists (ensure that you subscribe with and post from exactly the same e-mail address -- [email protected] is considered different than [email protected]!). Visit these pages to subscribe to the lists:

Make today an Open MPI day!