LAMMPS with Intel MPI and Xeon Phi
Compile the latest LAMMPS. Use Intel Xeon Phi co-processor card if available, otherwise fall back to CPU version.
To build the
LAMMPS needs Intel's MPSS libraries (Manycore Platform Software Stack)
which are only available on our Intel Xeon Phi nodes.
Therefore instead of directly running
compile LAMMPS on the Intel Xeon Phi node using
the provided SLURM script:
This will create a
lammps executable that wraps
both the CPU and Phi offload versions of LAMMPS.
Note that the executable loads the intel compiler dependency
for MKL, Intel MPI, etc.
You probably want to add that
lammps script to your
so that you can easily run it other directories without the
# In your ~/.bashrc export PATH=path/to/your/lammps:$PATH
For the CPU version, your SLURM script must subset to Haswell or newer nodes using:
#SBATCH --exclude cn[01-136] mpirun lammps -i your_input_file.in
If you are using the Phi card for offload mode, we don't need the node exclusions as the phi nodes are Haswell. Make sure you have the following in your SLURM script:
#SBATCH --partition phi #SBATCH --ntasks 1 # Offload mode deadlocks with > 1 CPU #SBATCH --gres mic:1 # Offload mode deadlocks with > 1 Phi card lammps -i your_input_file.in
To compare the speed with using the Phi co-processor and without,
run Intel's bundled test inputs.
The Phi card completes in 8 minutes, while the CPU version
times out after 30 minutes.