In this page, we show an example of the usage of GENESIS on Intel PC cluster machines. Because the actual usage depends on the machine environment, the following protocol may not be directly applicable to your system. Here, we assume that the machine’s IP address is 192.168.1.2, each computational node has 16 CPU cores (2 CPUs and 8 cores in one CPU), gridengine or its variant is installed as a job scheduler, and the user’s MPI environment is set by mpi-selector-menu.

Without GPGPU

The followings are example of the installation scheme, batch script file for hybrid MPI/OpenMP computation with 16 MPI processes and 4 OpenMP threads (64 CPU cores in total), and a command to submit a job.

Installation

# Login the machine
$ ssh 192.168.1.2

# select one proper option (e.g., ib-openmpi-1.10.1_intel-15.0.4_cuda-6.5)
$ mpi-selector-menu

# login again to update MPI setting
$ exit
$ ssh 192.168.1.2

# install genesis
$ cd /swork/user/genesis
$ ./configure
$ make
$ make install

Batch script

#$ -S /bin/bash
#$ -cwd
#$ -pe ompi 64
#$ -V
#$ -q nogpu.q
BINDIR=/swork/user/genesis/bin
mpirun -machinefile $TMP/machines -np 16 -npernode 4 -npersocket 2 -x OMP_NUM_THREADS=4 ${BINDIR}/spdyn INP > md.log

If “-npernode X” does not work, use “--bind-to socket”.

Example 2:

#$ -S /bin/bash
#$ -cwd
#$ -pe ompi 64
#$ -V
#$ -q nogpu.q
export OMP_NUM_THREADS=4
BINDIR=/swork/user/genesis/bin
mpirun -machinefile $TMP/machines -np 16 -npernode 4 -npersocket 2 ${BINDIR}/spdyn INP > md.log

Usage

# Execute run1.sh to run3.sh sequentially
$ qsub -N R1 run1.sh
$ qsub -N R2 -hold_jid R1 run2.sh
$ qsub -N R3 -hold_jid R2 run3.sh

# Check running jobs
$ qstat -f

# delete a job
$ qdel JOB_ID

After submitting a job, the user should check whether the specified CPU resources are fully utilized by running the top command in some of the computational nodes. If the CPUs are not fully used, the option of mpirun might not be appropriate.  In some cases, --bind-to socket may have to be be used instead of -npernode.

with GPGPU

The followings are example of the installation scheme and batch script file for hybrid MPI/OpenMP computation with 16 MPI processes and 4 OpenMP threads with 2 GPU cards/node (64 CPU cores + 8 GPU cards in total).

Installation

# Login the machine
$ ssh 192.168.1.2

# select one proper option (e.g., ib-openmpi-1.10.1_intel-15.0.4_cuda-6.5)
$ mpi-selector-menu

# login again to update MPI setting
$ exit
$ ssh 192.168.1.2

# install genesis
$ cd /swork/user/genesis
$ ./configure --enable-gpu --enable-single
$ make
$ make install

Batch script

GENESIS 1.1.2 or later

Number of MPI processors and OpenMP threads are specified in run.sh.

run.sh:

#$ -S /bin/bash
#$ -cwd
#$ -pe ompi 64
#$ -V
#$ -q gpu.q
export OMP_NUM_THREADS=4
BINDIR=/swork/user/genesis/bin
mpirun -machinefile $TMP/machines -np 16 -npernode 4 -npersocket 2 ${BINDIR}/spdyn INP > md.log

Number of GPU cards will be automatically detected. If you wanna exclude some specific GPU card or wanna use only a single specific GPU card among multiple cards, please check this page.

GENESIS 1.1.1 or before

Number of MPI processors and OpenMP threads are specified in run.sh, and number of GPU cards are in wrap.sh.

run.sh:

#$ -S /bin/bash
#$ -cwd
#$ -pe ompi 64
#$ -V
#$ -q gpu.q
export OMP_NUM_THREADS=4
BINDIR=/swork/user/genesis/bin
mpirun -machinefile $TMP/machines -np 16 -npernode 4 -npersocket 2 ./wrap.sh ${BINDIR}/spdyn INP > md.log

wrap.sh:

#!/bin/bash
lr=${OMPI_COMM_WORLD_LOCAL_RANK:-0}
gpuid=`expr ${lr} \% 2`
export CUDA_VISIBLE_DEVICES=${gpuid}
$@

Usage of GPU nodes is same with no GPU case. Please see the above explanation.

Updated: