Launching MPMD Jobs
The aprun command can be used in Multiple Program, Multiple Data (MPMD) mode by providing several job placement options/binary combinations separated by a colon:
$> aprun -n pes_1 [placement_options_1] BINARY_0 : \ -n pes_2 [placement_options_2] BINARY_1 : \ -n pes_3 [placement_options_3] BINARY_2
To ensure that your shell parses the colon as a completely separate token, make sure to include a space on both sides of the colon. All other options can be used in the usual fashion.
The example below shows an MPI application that uses all the cores in 8 XE nodes running $MY_XE_BINARY, and all the cores in 16 XK nodes running $MY_XK_BINARY.
#! /bin/bash # #PBS -l nodes=8:ppn=32:xe+16:ppn=16:xk #PBS -l walltime=00:30:00 #PBS -N aprun-mpmd-example-1 aprun -n $((8*32)) -N 32 $MY_XE_BINARY : \ -n $((16*16)) -N 16 $MY_XK_BINARY
The following example demostrates the use of MPMD mode for a hybrid MPI and openMP application.
#! /bin/bash # #PBS -l nodes=16:ppn=32:xe #PBS -l walltime=00:30:00 #PBS -N aprun-mpmd-example-2 export OMP_NUM_THREADS=4 #setenv OMP_NUM_THREADS 4 aprun -n $((32*4)) -N 32 -d 1 $MY_XE_BINARY_1 : \ -n $((8*12)) -N 8 -d 4 $MY_XE_BINARY_2
Note: The binaries in a MPMD job can be the same or different, but will run together under a unified MPI_COMM_WORLD. There's no requirement that 2 different resource specifications be used with the '+' syntax in the first example. In addition, an MPMD job can use all xk or all xe nodes. The important concept is the aprun syntax with the ' : ' separator for the executables.
- Each executable needs a dedicated node. Executables cannot share a single node. i.e. the reservation should be for at least as many nodes as there are executables in the aprun command.
- MPMD mode will not work for system commands; applications require at the least an enclosure with MPI_Init() and MPI_Finalize().
- MPMD applications that use the SHMEM parallel programming model, either standalone or nested within an MPI program, are not supported on BlueWaters.
- Applications submitted in MPMD mode will share the standard output that may result in messy output if several applications will be writing into the output stream.
See the aprun man page for more information.