This guide is intended to help Intel® compiler customers build and use Open MPI* library. Open MPI is a standards-compliant, open-source implementation of the Message Passing Interface, a library specification for parallel processes or threads to exchange data in a parallel application.
This application note applies to Open MPI version 1.2, as built with the Intel® Fortran Compiler 10.0 and Intel® C++ Compiler 10.0. The procedure outlined in this document are tested with the 10.0 versions of the Intel Fortran Compiler and Intel C++ compiler. Prior versions of the Intel compilers may work as well, but have not been tested. Older versions or newer versions of Open MPI may work but have not been tested. This document only covers the 10.0 versions of the Intel compilers and Open MPI 1.2.
Open MPI is a standards-compliant, open-source implementation of the Message Passing Interface, a library specification for parallel processes or threads to exchange data in a parallel application. The target environment for Open MPI can vary dramatically based on interconnect, adapters, node types, batch subsystem, etc. Please read all relevant documentation for your target system.
This application note demonstrates the framework for building Open MPI with the Intel compilers but does NOT claim to represent all possible configurations and variations of the build for all possible target environments.
Obtaining the latest version of Intel C++ Compiler and Intel Fortran Compiler
Licensed users of the Intel compilers may download the most recent versions of the compiler from the Download Center which is acce ssed by logging into your account at Registration Center.
Hardware: This note applies to users with stand-alone computers with two or more cores or distributed-memory clusters.
Software: This note applies to Open MPI built on Linux (32 and 64 bit) and Mac OS* X using the 10.0 version of the Intel compilers. Please consult the Open MPI† website for a complete list of supported platforms and operating systems.
Configuration and set-up information
Open MPI uses an Autoconf "configure" script to determine the build environment and tools and create the necessary build configuration. The method to build Open MPI, open a Mac OS* Terminal window or a Linux* shell:
gunzip -c openmpi-1.2.tar.gz | tar xf -
./configure --prefix=/usr/local CC=icc CXX=icpc F77=ifort FC=ifort
... output of configure ...
make all install
... output of build and installation ...
As shown above, the configure options CC, CXX, F77, and FC are used to specify which compilers are used to build Open MPI. The example shown above uses both the Intel C++ Compiler ( CC=icc CXX=icpc ) and the Intel Fortran Compiler (F77=ifort FC=ifort ). Note that the Intel C++ compiler driver is named ' icpc'. Do NOT use ' icc' as the C++ compiler. The Intel compilers are GNU compatible, thus you may mix and match the Intel compilers with GNU compilers for C++ and Fortran. However, the mixing of GNU compilers with Intel compilers has not been tested with this application.
The configure script has a lot of possible options. To learn all of the possible configuration options, see the help provided in the output of "./configure --help" or read the contents of the file README provided in the OpenMPI tar archive file.
Specifying Intel compiler options in the configuration and build
To pass options to the compilers, the following command options to configure are supported:
CFLAGS= arguments to pass to the C compiler ( icc)
CXXFLAGS= arguments to pass to the C++ compiler ( icpc)
FFLAGS= arguments to pass to the F77 compiler ( ifort)
FCFLAGS= arguments to pass to the Fortran 90 compiler ( ifort)
Please see the Intel Compiler Documentation for appropriate options. In general, MPI performance is dominated by interconnect fabric latency and to a less extent on bandwidth. A general rule of thumb is to use the default compiler optimizations and avoid overly aggressive optimizations.
The default optimization with the Intel Compilers, that is if NO compiler option for -O is specified, is -O2. The -O2 optimizations provide a good level of optimization and safety for both the Intel C++ compiler and the Intel Fortran Compiler. The Open MPI build defaults have a reasonable optimization level.
Building MPI applications
Once Open MPI is installed, set your PATH environment variable to point to the <installation directory>/bin directory. Remember, from above, the <installation directory> is set by the --prefix argument you passed to configure. Typically this is /usr/local. Thus, for bash shell on Linux* and Mac* OS, one would:
and on Mac OS X you also should set:
You should consider making this part of your default user environment by adding this export statement to your ~/.profilestartup file.
Once that is done, to compile MPI application you use the Open MPI compiler wrappers.
The C wrapper is named mpicc.
The C++ wrapper has 3 equivalent wrappers, mpicxx, mpiCC, or mpic++. Use any one of these.
The Fortran 77 wrapper is name mpif77.
The Fortran 90 wrapper is named mpif90.
These compiler wrappers will invoke the Intel compilers and link all necessary Open MPI libraries.
mpicc -o mpi_pong mpi_pong.c
See the Open MPI FAQ information on building applications for more extensive details and information.
Running an application under Open MPI
Open MP provides command mpirun and mpiexec to launch parallel applications. Below is a simple example to run a 2 process MPI application:
There are many possible options to mpirun and mpiexec, as well as interactions with batch scheduling systems such as PBS. Please see the Open MPI FAQ for all the specifics for your system or cluster.
This application note has shown how to build Open MPI* using the Intel Compilers. Benefits of using this Intel Compiler build of Open MPI include improved optimization and compatibility with the Intel Compiler generated binary files and GNU generated binary files.