lammps/examples/accelerate/README

159 lines
5.5 KiB
Plaintext

These are example scripts that can be run with any of
the acclerator packages in LAMMPS:
USER-CUDA, GPU, USER-INTEL, KOKKOS, USER-OMP, OPT
The easiest way to build LAMMPS with these packages
is via the src/Make.py tool described in Section 2.4
of the manual. You can also type "Make.py -h" to see
its options. The easiest way to run these scripts
is by using the appropriate
Details on the individual accelerator packages
can be found in doc/Section_accelerate.html.
---------------------
Build LAMMPS with one or more of the accelerator packages
The following command will invoke the src/Make.py tool with one of the
command-lines from the Make.list file:
../../src/Make.py -r Make.list target
target = one or more of the following:
cpu, omp, opt
cuda_double, cuda_mixed, cuda_single
gpu_double, gpu_mixed, gpu_single
intel_cpu, intel_phi
kokkos_omp, kokkos_cuda, kokkos_phi
If successful, the build will produce the file lmp_target in this
directory.
Note that in addition to any accelerator packages, these packages also
need to be installed to run all of the example scripts: ASPHERE,
MOLECULE, KSPACE, RIGID.
These two targets will build a single LAMMPS executable with all the
CPU accelerator packages installed (USER-INTEL for CPU, KOKKOS for
OMP, USER-OMP, OPT) or all the GPU accelerator packages installed
(USER-CUDA, GPU, KOKKOS for CUDA):
target = all_cpu, all_gpu
Note that the Make.py commands in Make.list assume an MPI environment
exists on your machine and use mpicxx as the wrapper compiler with
whatever underlying compiler it wraps by default. If you add "-cc mpi
wrap=g++" or "-cc mpi wrap=icc" after the target, you can choose the
underlying compiler for mpicxx to invoke. E.g.
../../src/Make.py -r Make.list intel_cpu -cc mpi wrap=icc
You should do this for any build that includes the USER-INTEL
package, since it will perform best with the Intel compilers.
Note that for kokkos_cuda, it needs to be "-cc nvcc" instead of "mpi",
since a KOKKOS for CUDA build requires NVIDIA nvcc as the wrapper
compiler.
Also note that the Make.py commands in Make.list use the default
FFT support which is via the KISS library. If you want to
build with another FFT library, e.g. FFTW3, then you can add
"-fft fftw3" after the target, e.g.
../../src/Make.py -r Make.list gpu -fft fftw3
For any build with USER-CUDA, GPU, or KOKKOS for CUDA, be sure to set
the arch=XX setting to the appropriate value for the GPUs and Cuda
environment on your system. What is defined in the Make.list file is
arch=21 for older Fermi GPUs. This can be overridden as follows,
e.g. for Kepler GPUs:
../../src/Make.py -r Make.list gpu_double -gpu mode=double arch=35
---------------------
Running with each of the accelerator packages
All of the input scripts have a default problem size and number of
timesteps:
in.lj = LJ melt with cutoff of 2.5 = 32K atoms for 100 steps
in.lj.5.0 = same with cutoff of 5.0 = 32K atoms for 100 steps
in.phosphate = 11K atoms for 100 steps
in.rhodo = 32K atoms for 100 steps
in.lc = 33K atoms for 100 steps (after 200 steps equilibration)
These can be reset using the x,y,z and t variables in the command
line. E.g. adding "-v x 2 -v y 2 -v z 4 -t 1000" to any of the run
command below would run a 16x larger problem (2x2x4) for 1000 steps.
Here are example run commands using each of the accelerator packages:
** CPU only
lmp_cpu < in.lj
mpirun -np 4 lmp_cpu -in in.lj
** OPT package
lmp_opt -sf opt < in.lj
mpirun -np 4 lmp_opt -sf opt -in in.lj
** USER-OMP package
lmp_omp -sf omp -pk omp 1 < in.lj
mpirun -np 4 lmp_omp -sf opt -pk omp 1 -in in.lj # 4 MPI, 1 thread/MPI
mpirun -np 2 lmp_omp -sf opt -pk omp 4 -in in.lj # 2 MPI, 4 thread/MPI
** GPU package
lmp_gpu_double -sf gpu < in.lj
mpirun -np 8 lmp_gpu_double -sf gpu < in.lj # 8 MPI, 8 MPI/GPU
mpirun -np 12 lmp_gpu_double -sf gpu -pk gpu 2 < in.lj # 12 MPI, 6 MPI/GPU
mpirun -np 4 lmp_gpu_double -sf gpu -pk gpu 2 tpa 8 < in.lj.5.0 # 4 MPI, 2 MPI/GPU
Note that when running in.lj.5.0 (which has a long cutoff) with the
GPU package, the "-pk tpa" setting should be > 1 (e.g. 8) for best
performance.
** USER-CUDA package
lmp_machine -c on -sf cuda < in.lj
mpirun -np 1 lmp_machine -c on -sf cuda < in.lj # 1 MPI, 1 MPI/GPU
mpirun -np 2 lmp_machine -c on -sf cuda -pk cuda 2 < in.lj # 2 MPI, 1 MPI/GPU
** KOKKOS package for OMP
lmp_kokkos_omp -k on t 1 -sf kk -pk kokkos neigh half < in.lj
mpirun -np 2 lmp_kokkos_omp -k on t 4 -sf kk < in.lj # 2 MPI, 4 thread/MPI
Note that when running with just 1 thread/MPI, "-pk kokkos neigh half"
was speficied to use half neighbor lists which are faster when running
on just 1 thread.
** KOKKOS package for CUDA
lmp_kokkos_cuda -k on t 1 -sf kk < in.lj # 1 thread, 1 GPU
mpirun -np 2 lmp_kokkos_cuda -k on t 6 g 2 -sf kk < in.lj # 2 MPI, 6 thread/MPI, 1 MPI/GPU
** KOKKOS package for PHI
mpirun -np 1 lmp_kokkos_phi -k on t 240 -sf kk -in in.lj # 1 MPI, 240 threads/MPI
mpirun -np 30 lmp_kokkos_phi -k on t 8 -sf kk -in in.lj # 30 MPI, 8 threads/MPI
** USER-INTEL package for CPU
lmp_intel_cpu -sf intel < in.lj
mpirun -np 4 lmp_intl_cpu -sf intel < in.lj # 4 MPI
mpirun -np 4 lmp_intl_cpu -sf intel -pk omp 2 < in.lj # 4 MPI, 2 thread/MPI
** USER-INTEL package for PHI
lmp_intel_phi -sf intel -pk intel 1 omp 16 < in.lc # 1 MPI, 16 CPU thread/MPI, 1 Phi, 240 Phi thread/MPI
mpirun -np 4 lmp_intel_phi -sf intel -pk intel 1 omp 2 < in.lc # 4 MPI, 2 CPU threads/MPI, 1 Phi, 60 Phi thread/MPI
Note that there is currently no Phi support for pair_style lj/cut in
the USER-INTEL package.