Ada:Compile:MPI

MPI Programs

There are currently two MPI stacks installed on ADA; OpenMPI and Intel MPI. The recommended MPI stack for software development is the Intel MPI software stack and most of this section will focus on this MPI stack.

Intel MPI

To use the Intel MPI environment you need to load the Intel module. This can be done with the following command:

[ netID@cluster ~]$ module load intel/2017A

Note: It is no longer possible to load the default intel module. You must specify a version you are loading for the sake of consistency and clarity. More information about finding and loading modules can be found on our Modules Systems page.

Compiling MPI Code

To compile MPI code a MPI compiler wrapper is used. The wrapper will call the appropriate underlying compiler with additional linker flags specific for MPI programs. The Intel MPI software stack has wrappers for Intel compilers as well as wrappers for gnu compilers. Any argument not recognized by the wrapper will be passed to the underlying compiler. Therefore, any valid compiler flag (Intel or gnu) will also work when using the mpi wrappers

The following table shows the most commonly used MPI wrappers used by Intel MPI.

MPI Wrapper	Compiler	Language	Example
mpiicc	icc	C	mpiicc <compiler_flags> prog.c
mpicc	gcc	C	mpicc <compiler_flags> prog.c
mpiicpc	icpc	C++	mpiicpcp <compiler_flags> prog.cpp
mpicxx	g++	C++	mpicxx <compiler_flags> prog.cpp
mpiifort	ifort	Fortran	mpiifort <compiler_flags> prog.f90
mpif90	gfortran	Fortran	mpif90 <compiler_flags> prog.f90

To see the full compiler command of any of the mpi wrapper scripts use the **-show** flag. This flag does not actually call the compiler, it only prints the full compiler command and exits. This can be useful for debugging purposes and/or when experiencing problems with any of the compiler wrappers

Example: Show the full compiler command for the mpiifort wrapper script

[ netID@cluster ~]$ mpiifort -show 
ifort -I/software/easybuild/software/impi/4.1.3.049/intel64/include -I/software/easybuild/software/impi/4.1.3.049/intel64/include 
-L/software/easybuild/software/impi/4.1.3.049/intel64/lib -Xlinker --enable-new-dtags -Xlinker -rpath -Xlinker 
/software/easybuild/software/impi/4.1.3.049/intel64/lib -Xlinker -rpath -Xlinker -/opt/intel/mpi-rt/4.1 -lmpigf -lmpi -lmpigi -ldl -lrt -lpthread`

Running MPI Code

Running MPI code requires an MPI launcher. The latter will setup the environment and start the requested number of MPI tasks on the needed nodes.
Use the following command to launch an MPI program where [mpi_flags] are options passed to the mpi launcher, <executable> is the name of the mpi program and [executable params] are optional parameters for the mpi program. (We continue to assume here use of the Intel MPI stack.):

 [ netID@cluster ~]$ mpirun [mpi_flags] <executable> [executable params]

Note: <executable> must be on the $PATH otherwise the launcher will not be able to find the executable.

For a list of the most common mpi_flags See table below. This table shows only a very small subset of all flags. To see a full listing type mpirun --help

Flag	Description
-np <n>	The number of mpi tasks to start.
-n <n>	The number of mpi tasks to start (same as -np).
-perhost <n>	Places <n> consecutive (MPI) processes on each host/node.
-ppn <n>	Stands for Process (i.e., task) Per Node (same as -perhost)
-hostfile <file>	The name of the file that contains the list of host/node names the launcher will place tasks on.
-f <file>	Same as -hostfile
-hosts {host list}	comma separated list of specific host/node names.
-help	Shows list of available flags and options

Hybrid MPI/OpenMP Code

To compile hybrid mpi/OpenMP programs (i.e. MPI programs that also contain OpenMP directives) invoke the appropriate mpi wrapper and add the -openmp flag to enable processing of OpenMP primitives.

Running a hybrid program is very similar to running a pure mpi program. To control the number of OpenMP threads to use per task the OMP_NUM_THREADS environmental variable can be set.

Advanced: mapping tasks and threads

Explicitly mapping mpi tasks to processors can result in significantly better performance.This is especially true for hybrid MPI/OpenMP programs where both mpi tasks and OpenMP threads are being mapped on the available cores on a node. The Intel MPI stack provides a way to control the pinning of MPI tasks using the environmental variable I_MPI_PIN_DOMAIN.

[ netID@cluster ~]$ export I_MPI_PIN_DOMAIN=<domain>

where <domain> can have the following values; node, socket, core, cache1, cache2, cache3. The domain tells where to pin the tasks. For example "socket" will pin the tasks on different sockets. To map the OpenMP threads the affinity setting for OpenMP will be used.

NOTE: the above syntax is just one way to describe the pinning. Please visit the Process Pinning documentation or the Intel MPI reference (see Further Information section for link) for alternative ways to pin tasks using the I_MPI_PIN_DOMAIN environmental variable.

Examples

In this section are various examples for compiling and running MPI programs with the Intel toolchain.

Example 1: Compile MPI program written in C, and name it mpi_prog.x. Use the underlying Intel compiler with -O3 optimization

[ netID@cluster ~]$ mpiicc -o mpi_prog.x -O3 mpi_prog.c

Example 2: Same as Example 1, but this time use underlying gnu Fortran compiler.

[ netID@cluster ~]$ mpif90 -o mpi_prog.x  mpi_prog.f90 

Example 3: Run mpi program on local host using 4 tasks

[ netID@cluster ~]$ mpirun -np 4 mpi_prog.x 

Example 4: Run mpi program on a specific host, using 4 tasks

[ netID@cluster ~]$ mpirun -np 4 -hosts login1  mpi_prog.x 

Example 5: Run mpi program on two different hosts, using 4 tasks using host file, assign tasks in round robin fashion

[ netID@cluster ~]$ mpirun -np 4 -perhost 1 -hostfile mylist  mpi_prog.x 

where mylist is a file that contains the following lines:

login1
login2

Note: If you don't specify -pernode all the tasks will be started on login1, even though the hostfile contains multiple entries.

Example 6: Run 4 different programs concurrently using mpirun (MPMD style program)

[ netID@cluster ~]$ mpirun -np 1 prog1.x : -np 1 prog2.x : -np 1 prog3.x : -np 1 prog4.x

Note: For executing a large number of serial (or OpenMP) programs, we recommend using the tamulauncher utility.

Example 7: Compile MPI fortran program named hybrid.f90 that also contains OpenMP primitives, use underlying Intel Fortran compiler

[ netID@cluster ~]$ mpiifort -openmp -o hybrid.x hybrid.f90

Example 8: Run the hybrid program named hybrid.x, use 8 tasks and every task will use 2 threads in its OpenMP regions.

[ netID@cluster ~]$ export OMP_NUM_THREADS=2
[ netID@cluster ~]$ mpirun -np 8 ./hybrid.x

Example 9: run hybrid mpi/OpenMP program using 2 tasks and 10 threads, pin the tasks to different sockets, map all OpenMP threads within the socket

[ netID@cluster ~]$ export I_MPI_PIN_DOMAIN=socket
[ netID@cluster ~]$ export OMP_NUM_THREADS=10
[ netID@cluster ~]$ export OMP_PLACES="socket"
[ netID@cluster ~]$ export OMP_PROC_BIND="master"
[ netID@cluster ~]$ mpirun -np 2 ./hybrid.x

Further Information

For a detailed description of the Intel MPI stack, please visit the Intel MPI Developer Reference Manual. This site contains detailed information about the mpi compiler wrappers, in depth discussion about mpirun and it options, as well as tuning your application for best performance and pinning tasks.

OpenMPI

Using OpenMPI is very similar to using Intel MPI. There are a few minor differences. To use OpenMPI you will need to load one of the OpenMPI modules. ADA has OpenMPI versions build with Intel compilers as well as gnu compilers. The underlying compiler depends on the loaded OpenMPI module

Example 1: Load OpenMPI version 1.8.4 with gnu as underlying compiler

[ netID@cluster ~]$ module load OpenMPI/1.8.4-GCC-4.9.2

To see a list of all available OpenMPI versions type:

[ netID@cluster ~]$ module spider openmpi

Compiling

The table below shows the various mpi compiler wrappers. The names will be the same regardless of the underlying compiler.

MPI wrapper	Language	Example
mpicc	C	mpicc <compiler_flags> prog.c
mpic++	C++	mpic++ <compiler_flags> prog.cpp
mpif90	Fortran	mpif90 <compiler_flags> prog.f90

to see the complete compiler command use the -show flag.

Running

To launch a mpi program you will use the mpirun command. This command is very similar to the Intel MPI mpirun launcher discussed above. However, some of the flags are different for OpenMPI. The table below shows some of the more common flags.

Flag	Description
-np <n>	The number of mpi tasks to start.
-npernode <n>	Places <n> (MPI) processes per node on each allocated node.
-npersocket <n>	Places <n> (MPI) processes per socket on each allocated node
-hostfile <file>	The name of the file that contains the list of host/node names the launcher will place tasks on.
-host {host list}	comma separated list of specific host/node names.

To see all the available options and flags (including short descriptions) use the following command:

[ netID@cluster ~]$ mpirun -help