Index of /examples/mpi

Name	Last modified	Size

Parent Directory		-
tutorials/	2024-06-10 09:57	-
basic/	2024-05-30 10:06	-
examples/	2013-11-24 09:28	-

RCS EXAMPLES

RCS examples are provided to assist you in learning the software and the development of your applications on the Shared Computing Cluster (SCC). The instructions provided along with the code assume that the underlying OS is Linux. If these examples are run on a different architecture, you might need to make some changes to the code and/or the way the program is built and executed.

Directory Structure:

basic: This directory presents MPI example programs and batch scripts for running the programs on SCC.
tutorials: Examples used in RCS's "Introduction to MPI" live tutorial and online Multiprocessing by Message Passing MPI tutorial.
examples: This directory presents MPI example programs.

Using MPI on the SCC

The OpenMPI library is the standard MPI library for the SCC. To see the available versions use the module command:

scc1% module avail openmpi

OpenMPI version 3.1.4 is the recommended version to use on the SCC and is the one that is used by RCS when building software modules that use MPI. It provides C, C++, and Fortran wrappers for compiling MPI code called mpicc, mpic++, and mpif90, respectively. The OpenMPI module should be matched to the compiler that will be used. The openmpi/3.1.4 module is intended for use with the sytem default gcc, g++, and gfortran compilers which are version 4.8.5. To use other available compilers use the OpenMPI module whose name corresponds to the compiler name and version. For the OpenMPI modules the "gnu" compiler refers to the "gcc" compiler modules. For example, to use the 9.3.0 versions of the GNU compilers use: :

scc1%  module load gcc/9.3.0 

scc1%  module load openmpi/3.1.4_gnu-9.3.0

or, to use the Intel C, C++, and Fortran compilers with OpenMPI use:

scc1%  module load intel/2021.1 

scc1%  module load openmpi/3.1.4_intel-2021

When running your compiled code in a batch job, it is required that you load the compiler and matching OpenMPI module in the batch script before starting the MPI program.

The OpenMPI modules provide the mpirun command to launch MPI jobs. To allocate MPI resources for your job, please see the RCS MPI batch job documentation page. Further details on compiling MPI programs is also available.

If an MPI job is launched across multiple nodes but there are fewer MPI processes than cores the recommended mpirun arguments are as follows. In this example there are 4 28-core nodes in the job but only 4 processes per node will be launched (for a total of 16 processes). The --map-by argument will distribute the MPI processes across the nodes, placing 2 processes per CPU socket to make the best use of memory bandwidth on each node. The --rank-by command will number the MPI ranks so that nodes contain the ranks in order, i.e. node 0 will have ranks 0-3, node 1 will have ranks 4-7, etc.:

scc1% mpirun -np 16 --map-by numa:span --rank-by core mpi_program ...args...

The OpenMPI modules are configured to understand the SGE batch job environment and the mpirun launcher will automatically use the assigned set of nodes. Additionally, mpirun will use the high-speed Infiniband networks available on the MPI nodes for inter-node communication. A utility has been installed, module xth1/1.0, that can be used to experiment with different mpirun options.

Contact Information

help@scc.bu.edu

Note: Research Computing Services (RCS) example programs are provided "as is" without any warranty of any kind. The user assumes the entire risk of quality, performance, and repair of any defects. You are encouraged to copy and modify any of the given examples for your own use.