cluster:161
This is an old revision of the document!
Table of Contents
lammps-11Aug17
lammps-11Aug17 (n78)
and now lammps-22Aug18 (n33-n37)
and now lammps-5Jun19 (microway)
Update:
- n78/gtx1080 lammps 11aug17 (centos7, mpich3/mpic++, cuda 8/sm_61, /usr/local)
- n37/k20 lammps 22aug18 (centos7, openmpi 1.8.4/mpic++, cuda 9.2/sm_35, /usr/local)
- n33/k20 lammps 11aug17 + 22aug18 (centos6, openmpi 1.8.4, cuda 5/sm35, /share/apps/CENTOS6)
- this page
- n33/k20 older lammps (centos6, mvapich, cuda 5/sm_??, /cm/shared/apps)
- /home/hmeij/sharptail/run.gpu or gromacs.sub files
— Henk 2018/09/20 09:30
N78 # devel packages for libjpeg, blas, lapack, boost which nvcc export PATH=/usr/local/mpich-3.1.4/bin:$PATH export LD_LIBRARY_PATH=/usr/local/mpich-3.1.4/lib:$LD_LIBRARY_PATH which mpicxx # untar stable.gz cd lib/gpu vi Makefile.linux.double # edits CUDA_HOME = /usr/local/cuda # best fitting ARCH for GTX 1080 with CUDA 8 CUDA_ARCH = -arch=sm_61 # best fitting ARCH for RTX/T4 with CUDA 10 CUDA_ARCH = -arch=sm_75 http://arnon.dk/matching-sm-architectures-arch-and-gencode-for-various-nvidia-cards/ # generate DOUBLE_DOUBLE (case matters) CUDA_PRECISION = -D_DOUBLE_DOUBLE make -f Makefile.linux.double ls -lrt # new files: libgpu.a and Makefile.lammps cd ../../src/ vi MAKE/Makefile.mpi # edits LMP_INC = -DLAMMPS_GZIP -DLAMMPS_MEMALIGN=64 -DLAMMPS_JPEG JPG_INC = -I/usr/include JPG_PATH = -L/usr/lib64 JPG_LIB = -ljpeg make purge make yes-rigid make yes-gpu make yes-colloid make yes-class2 make yes-kspace make yes-misc make yes-molecule make package-update which mpicxx make mpi mv lmp_mpi /usr/local/lammps-11Aug17/lmp_mpi-double-double-with-gpu[-rigid] # repeat from scratch for SINGLE_SINGLE and SINGLE_DOUBLE # different makefiles same edits
Lammps CPU/GPU
Because it took me so long to get the GPU side of things compiled. I've compiled it two ways, and the packages will differ between both versions.
- Build my own recipe
- Followed, and slightly altered somebody else's recipe
Initially the GPU binary would hang after launching in a nanosleep loop. My bad. In my submit script I had an incomplete PATH, we need both /cm/shared/…/bin and /cm/local/…/bin for the GPU to work properly.
Submit scripts for serial, cpu and gpu follow. Centos 6.x and Openlava 3.x and 31Mar17 Lammps.
My recipe
# N33, a gpu node mkdir /tmp/lammps; cd !$ gunzip lammps-stable.tar.gz tar xvf lammps-stable.tar chown -R hmeij:its lammps-31Mar17 su - hmeij [hmeij@n33 src]$ which icc /share/apps/intel/parallel_studio_xe_2016_update3/compilers_and_libraries_2016.3.210/linux/bin/intel64/icc [hmeij@n33 src]$ which ifort /share/apps/intel/parallel_studio_xe_2016_update3/compilers_and_libraries_2016.3.210/linux/bin/intel64/ifort [hmeij@n33 src]$ echo $INTEL_LICENSE_FILE /share/apps/intel/parallel_studio_xe_2016_update3/flexlm/server.lic:/share/apps/intel/parallel_studio_xe_2016_update3/compilers_and_libraries_2016.3.210/linux/licenses:/opt/intel/licenses:/home/hmeij/intel/licenses [hmeij@n33 src]$ echo $MKL_HOME # not using icc/icpc/ifort but g++ only #no export MKL_HOME=/share/apps/intel/parallel_studio_xe_2016_update3/compilers_and_libraries_2016.3.210/linux/mkl cd /tmp/lammps/lammps-31Mar17/ cd src make yes-colloid make yes-class2 make yes-kspace make yes-misc make yes-molecule make no-user-misc make no-user-vtk make no-user-omp cd STUBS/ make cd .. #no cp MAKE/Makefile.serial MAKE/Makefile.serial_icc # edit MAKE/MAkefile.serial # leave gcc++ icc/icpc -DLAMMPS_JPEG for LMP_INC # JPG_INC = -I/usr/include # JPG_PATH = -L/usr/lib64 # JPG_LIB = -ljpeg # leave FFT_INC, FFT_PATH, FFT_LIB blank for KISS FFTW #no make serial_icc make serial size ../lmp_serial_icc text data bss dec hex filename 8472387 248912 2464 8723763 851d33 ../lmp_serial_icc #no mv lmp_serial_icc /share/apps/CENTOS6/lammps/31Mar17/ mv lmp_serial /share/apps/CENTOS6/lammps/31Mar17/ # note compile fails for mvapich2, trying with openmpi #/cm/shared/apps/mvapich2/gcc/64/1.6/bin/mpicxx <---------- has to be for wrapper export LD_LIBRARY_PATH=/share/apps/CENTOS6/openmpi/1.8.4/lib/:$LD_LIBRARY_PATH export PATH=/share/apps/CENTOS6/openmpi/1.8.4/bin:$PATH which mpicxx /share/apps/CENTOS6/openmpi/1.8.4/bin/mpicxx <------ works, will need logic of mvapich2 wrapper into new openmpi wrapper # perform edits in makefile as above but leave mpicxx # leave MPI_INC, MPI_PATH, MPI_LIB as is, blank make mpi size ../lmp_mpi text data bss dec hex filename 5630638 8320 16784 5655742 564cbe ../lmp_mpi mv lmp_mpi /share/apps/CENTOS6/lammps/31Mar17/ ### CUDA, use icc/mpi env from above export PATH=/share/apps/CENTOS6/python/2.7.9/bin:$PATH export LD_LIBRARY_PATH=/share/apps/CENTOS6/python/2.7.9/lib:$LD_LIBRARY_PATH which python export PATH=/cm/shared/apps/cuda50/toolkit/5.0.35/bin:$PATH export PATH=/cm/local/apps/cuda50/libs/304.54/bin:$PATH export LD_LIBRARY_PATH="/cm/shared/apps/cuda50/toolkit/5.0.35/lib64:/cm/shared/apps/cuda50/toolkit/5.0.35/lib:$LD_LIBRARY_PATH" # this is for libcuda.so export LD_LIBRARY_PATH="/cm/local/apps/cuda50/libs/304.54/lib64:/cm/local/apps/cuda50/libs/304.54/lib/:$LD_LIBRARY_PATH" cd /tmp/lammps/lammps-31Mar17/ cd src make yes-gpu make yes-colloid make yes-class2 make yes-kspace make yes-misc make yes-molecule which mpicxx which nvcc # edit lib/gpu/Makefile.auto set CUDA_HOME, this had no effect export CUDA_HOME=/cm/shared/apps/cuda50/toolkit/current export CUDA_INSTALL_PATH=/cm/shared/apps/cuda50/toolkit/5.0.35 # had to make links # /usr/local/cuda -> /cm/shared/apps/cuda50/toolkit/current # /cm/shared/apps/cuda50/toolkit/current/lib64/libcuda.so -> # /cm/local/apps/cuda50/libs/304.54/lib64/libcuda.so # note the -a flag is missing in Make.list examples, how can that be? # also -o is ignored it generates lmp_auto, totally weirdo make clean ./Make.py -v -j 2 -p colloid class2 kspace misc molecule gpu -gpu mode=double arch=35 -o gpu_double -a lib-gpu file clean mpi size ../lmp_auto text data bss dec hex filename 11831705 10608 299016 12141329 b94311 ../lmp_auto mv lmp_auto /share/apps/CENTOS6/lammps/31Mar17/lmp_gpu_double # repeat for single and mixed modes # nanosleep problem due to incomple $PATH, fixed, # need both /cm/shared... and /cm/local ... # the malaysian approach also works, see install file
Other recipe
Same environment as above
#source, many thanks. #http://comsics.usm.my/tlyoon/configrepo/howto/customise_centos/inst_lammps_31Mar17_gnu.txt #lammps installation (version 31Mar2017) # hmeij - edit the Makefiles and add -DLAMMPS-JPEG etc, fix path in *.lammps.installed # for -I/usr/lib64/atlas, leave FFTW lines blank cd /share/apps wget http://www2.fizik.usm.my/configrepo/howto/mylammps/lammps-31Mar17.tar tar -xvf lammps-31Mar17.tar cd lammps-31Mar17 mylammpsdir=$(echo $PWD) cd $mylammpsdir/src/MAKE wget http://comsics.usm.my/tlyoon/configrepo/howto/customise_centos/lammps/Makefile.gnu # atc # cd $mylammpsdir/lib/atc rm -f *.a *.o wget http://comsics.usm.my/tlyoon/configrepo/howto/customise_centos/lammps/Makefile.lammps.installed.atc rm -rf Makefile.lammps.installed cp Makefile.lammps.installed.atc Makefile.lammps.installed make -f Makefile.mpic++ cd $mylammpsdir/lib/awpmd rm -f *.a *.o wget http://comsics.usm.my/tlyoon/configrepo/howto/customise_centos/lammps/Makefile.lammps.installed.awpmd rm -rf Makefile.lammps.installed cp Makefile.lammps.installed.awpmd Makefile.lammps.installed make -f Makefile.mpicc # linalg # cd $mylammpsdir/lib/linalg rm -f *.a *.o make -f Makefile.gfortran # gpu NVIDIA CUDA on n33 comppute node cd $mylammpsdir/lib/gpu rm -f *.a *.o make -f Makefile.linux.double # meam # cd $mylammpsdir/lib/meam rm -f *.a *.o make -f Makefile.gfortran # poems # cd $mylammpsdir/lib/poems rm -f *.a *.o make -f Makefile.g++ # reax # cd $mylammpsdir/lib/reax rm -f *.a *.o make -f Makefile.gfortran cd $mylammpsdir/lib/qmmm rm -f *.a *.o make -f Makefile.gfortran cd $mylammpsdir/lib/colvars rm -f *.a *.o make -f Makefile.g++ # Now exit /lib cd $mylammpsdir/src make yes-all make no-user-quip make no-user-smd make no-user-vtk make no-kim make no-voronoi make no-kokkos make no-mscg make no-user-h5md # then serial & double make gnu make gpu
CPU/GPU Jobs
# serial and mpi
#!/bin/bash
# submit via 'bsub < run.gpu'
rm -f out err auout.[0-9]*
#BSUB -e err
#BSUB -o out
#BSUB -q mw128
#BSUB -J "GPU test"
#BSUB -n 4
# unique job scratch dirs
MYSANSCRATCH=/sanscratch/$LSB_JOBID
MYLOCALSCRATCH=/localscratch/$LSB_JOBID
export MYSANSCRATCH MYLOCALSCRATCH
cd $MYSANSCRATCH
# LAMMPS
export LD_LIBRARY_PATH=/share/apps/CENTOS6/openmpi/1.8.4/lib/:$LD_LIBRARY_PATH
export PATH=/share/apps/CENTOS6/openmpi/1.8.4/bin:$PATH
export PATH=/cm/shared/apps/cuda50/toolkit/5.0.35/bin:$PATH
export LD_LIBRARY_PATH="/cm/shared/apps/cuda50/toolkit/5.0.35/lib64:/cm/shared/apps/cuda50/toolkit/5.0.35/lib:$LD_LIBRARY_PATH"
export LD_LIBRARY_PATH="/cm/local/apps/cuda50/libs/304.54/lib64:/cm/local/apps/cuda50/libs/304.54/lib/:$LD_LIBRARY_PATH"
export CUDA_HOME=/cm/shared/apps/cuda50/toolkit/current
export CUDA_INSTALL_PATH=/cm/shared/apps/cuda50/toolkit/5.0.35
export PATH=/share/apps/CENTOS6/python/2.7.9/bin:$PATH
export LD_LIBRARY_PATH=/share/apps/CENTOS6/python/2.7.9/lib:$LD_LIBRARY_PATH
ldd /share/apps/CENTOS6/lammps/31Mar17/lmp_mpi
# GPUIDX=1 use allocated GPU(s), GPUIDX=0 cpu run only (view header au.inp)
export GPUIDX=0 # <---------- flag, change lmp_serial to lmp_mpi below
# stage the data
scp -r ~/lammps/colloid-gpu/* .
# feed the binary
/share/apps/CENTOS6/lammps/31Mar17/lmp_mpi -h
/share/apps/CENTOS6/lammps/31Mar17/lmp_mpi \
-var GPUIDX $GPUIDX -in in.colloid -l out.colloid
# save results
scp out.colloid $HOME/lammps/colloid-gpu/out.colloid.$LSB_JOBID
And …
# gpu
#!/bin/bash
# submit via 'bsub < run.gpu'
rm -f out err
#BSUB -e err
#BSUB -o out
#BSUB -q mwgpu
#BSUB -J "GPU test"
#BSUB -m n33
#BSUB -n 1
#BSUB -R "rusage[gpu=1:mem=6144],span[hosts=1]"
# unique job scratch dirs
MYSANSCRATCH=/sanscratch/$LSB_JOBID
MYLOCALSCRATCH=/localscratch/$LSB_JOBID
export MYSANSCRATCH MYLOCALSCRATCH
cd $MYSANSCRATCH
# LAMMPS
export LD_LIBRARY_PATH=/share/apps/CENTOS6/openmpi/1.8.4/lib/:$LD_LIBRARY_PATH
export PATH=/share/apps/CENTOS6/openmpi/1.8.4/bin:$PATH
export PATH=/cm/shared/apps/cuda50/toolkit/5.0.35/bin:$PATH
export PATH=/cm/local/apps/cuda50/libs/304.54/bin:$PATH
export LD_LIBRARY_PATH="/cm/shared/apps/cuda50/toolkit/5.0.35/lib64:/cm/shared/apps/cuda50/toolkit/5.0.35/lib:$LD_LIBRARY_PATH"
export LD_LIBRARY_PATH="/cm/local/apps/cuda50/libs/304.54/lib64:/cm/local/apps/cuda50/libs/304.54/lib/:$LD_LIBRARY_PATH"
export CUDA_HOME=/cm/shared/apps/cuda50/toolkit/current
export CUDA_INSTALL_PATH=/cm/shared/apps/cuda50/toolkit/5.0.35
export PATH=/share/apps/CENTOS6/python/2.7.9/bin:$PATH
export LD_LIBRARY_PATH=/share/apps/CENTOS6/python/2.7.9/lib:$LD_LIBRARY_PATH
ldd /share/apps/CENTOS6/lammps/31Mar17/lmp_gpu_double
# GPUIDX=1 use allocated GPU(s), GPUIDX=0 cpu run only (view header au.inp)
export GPUIDX=1
# stage the data
scp -r ~/lammps/colloid-gpu/* .
# feed the wrapper
openmpi-mpirun-gpu /share/apps/CENTOS6/lammps/31Mar17/lmp_gpu -h
openmpi-mpirun-gpu /share/apps/CENTOS6/lammps/31Mar17/lmp_gpu \
-suffix gpu -var GPUIDX $GPUIDX -in in.colloid -l lmp_gpu.out.colloid
openmpi-mpirun-gpu /share/apps/CENTOS6/lammps/31Mar17/lmp_gpu_double -h
openmpi-mpirun-gpu /share/apps/CENTOS6/lammps/31Mar17/lmp_gpu_double \
-suffix gpu -var GPUIDX $GPUIDX -in in.colloid -l lmp_gpu_double.out.colloid
## save results
scp lmp_gpu.out.colloid $HOME/lammps/colloid-gpu/lmp_gpu.out.colloid.$LSB_JOBID
scp lmp_gpu_double.out.colloid $HOME/lammps/colloid-gpu/lmp_gpu_double.out.colloid.$LSB_JOBID
cluster/161.1565025731.txt.gz · Last modified: 2019/08/05 17:22 by hmeij07
Except where otherwise noted, content on this wiki is licensed under the following license: CC Attribution-Share Alike 4.0 International
