vasp需要parallel_studio_xe_2018_update4_cluster_edition.tgz的intel mpi支持,openMPI需要更改makefile参数,麻烦
/etc/profile.d/intel.sh --> source /etc/profile.d/intel.sh ifort --version
source /opt/intel/bin/compilervars.sh intel64 也是链接文件
source /opt/intel/compilers_and_libraries/linux/mkl/bin/mklvars.sh intel64
icpc: error #10417: Problem setting up the Intel(R) Compliler compilation environment, Requires ‘install path‘ setting gatherd from ‘g++‘
yum install gcc gcc-c++
linux 默认无g++ gcc-c++ :c++ support for GCC
编译gpu版本的vasp需要cuda
makefile.include配置文件中:
FFLAGS = -assume byterecl -w -xHOST 改为 FFLAGS = -assume byterecl -w -heap-arrays 64
CUDA_ROOT ?= /usr/local/cuda cuda的默认路径,为一个链接,一般不用改
GENCODE_ARCH := -gencode=arch=compute_30,code=\"sm_30,compute_30\" 根据显卡算力值修改数值 30为显卡算例*10
运行算例:
进到算例目录,nohup mpirun -np 96 vasp_std > out & ##### cpu跑的
gpu跑算例:
1. nohup vasp_gpu > out & 单gpu跑
2. nohup mpirun -np 5 -genv CUDA_VISIBLE_DEVICES=0,1,2,3,4,5 vasp_gpu > out & 用mpi,cuda多显卡跑
算例中:
INCARS LREAL = Auto 才能用gpu,否则报错
KPOINTS 9 9 1 调低,算例变小,前两个一般相同
nvidia-smi [-L]
nvcc --version nvidia cuda compiler
在vasp运行过程中,INCARS KPOINTS POSCAR POTCAR四个原文件不会发生改变,算例不同,可能能使用的显卡数目不同,跟INCARS中参数有关
Gaussian
export g16root=/opt/ohpc/pub/apps
export GAUSS_SCRDIR=/data/tmp 必须存在此目录
source $g16root/g16/bsd/g16.profile g16为解压后的Gaussian根目录
export PATH=$g16root/g16:$PATH
/opt/ohpc/pub/apps/g16/bsd/install
g16 < $g16root/tests/com/*.com > out g16/tests/com 目录下有算例
结合PBS
#PBS -N NAME
#PBS -o test.out
#PBS -e test.err
#PBS -l select=host=cn4:ncpus=20
export GAUSS_EXEDIR=‘/opt/ohpc/pub/apps/g16:/opt/ohpc/pub/apps/g16/bsd‘
export GAUSS_PDEF=‘37‘
/opt/ohpc/pub/apps/g16/g16 test.com