The NVIDIA Software Development Kit (SDK) Manager is an all-in-one tool that bundles developer software and provides an end-to-end development environment setup solution for NVIDIA SDK. Think “OneAPI” ala Intel. All Nvidia's compilers, math libraries, and communications tools in one easy to set up environment.
I performed a “single system install” in a shared location (see above) using gcc 9.2.0. This will only be good for the exx96 nodes. Bundled with the newest plus three previous CUDA versions (11.6, 11.2, 11.0, 10.2). cuda-10.2 runs the gpus in that queue. So first we set up
# gcc export PATH=/share/apps/CENTOS7/gcc/9.2.0/bin:$PATH export LD_LIBRARY_PATH=/share/apps/CENTOS7/gcc/9.2.0/lib64:\ /share/apps/CENTOS7/gcc/9.2.0/libexec:\ /share/apps/CENTOS7/gcc/9.2.0/lib:\ $LD_LIBRARY_PATH # cuda # comes from SDK
Next we load the 22.2 module and view what all is available. Consult the documentation.
module use /share/apps/CENTOS7/nvidia/hpc_sdk/22.2/modulefiles/nvhpc/
module load 22.2
module list
Currently Loaded Modules:
1) 22.2
module show 22.2
-------------------------------------------------------------------------------------------------------------------------
/share/apps/CENTOS7/nvidia/hpc_sdk/22.2/modulefiles/nvhpc/22.2:
-------------------------------------------------------------------------------------------------------------------------
conflict("nvhpc")
conflict("nvhpc-nompi")
conflict("nvhpc-byo-compiler")
setenv("NVHPC","/share/apps/CENTOS7/nvidia/hpc_sdk/22.2")
setenv("NVHPC_ROOT","/share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2")
setenv("CC","/share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/compilers/bin/nvc")
setenv("CXX","/share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/compilers/bin/nvc++")
setenv("FC","/share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/compilers/bin/nvfortran")
setenv("F90","/share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/compilers/bin/nvfortran")
setenv("F77","/share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/compilers/bin/nvfortran")
setenv("CPP","cpp")
prepend_path("PATH","/share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/cuda/bin")
prepend_path("PATH","/share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/compilers/bin")
prepend_path("PATH","/share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/comm_libs/mpi/bin")
prepend_path("PATH","/share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/compilers/extras/qd/bin")
prepend_path("LD_LIBRARY_PATH","/share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/cuda/lib64")
prepend_path("LD_LIBRARY_PATH","/share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/cuda/extras/CUPTI/lib64")
prepend_path("LD_LIBRARY_PATH","/share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/compilers/extras/qd/lib")
prepend_path("LD_LIBRARY_PATH","/share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/compilers/lib")
prepend_path("LD_LIBRARY_PATH","/share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/math_libs/lib64")
prepend_path("LD_LIBRARY_PATH","/share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/comm_libs/mpi/lib")
prepend_path("LD_LIBRARY_PATH","/share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/comm_libs/nccl/lib")
prepend_path("LD_LIBRARY_PATH","/share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/comm_libs/nvshmem/lib")
prepend_path("CPATH","/share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/math_libs/include")
prepend_path("CPATH","/share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/comm_libs/mpi/include")
prepend_path("CPATH","/share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/comm_libs/nccl/include")
prepend_path("CPATH","/share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/comm_libs/nvshmem/include")
prepend_path("CPATH","/share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/compilers/extras/qd/include/qd")
prepend_path("MANPATH","/share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/compilers/man")
setenv("OPAL_PREFIX","/share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/comm_libs/mpi")
And a whole bunch of other stuff, consult documentation.
[hmeij@n79 ~]$ which nvcc nvc++ /share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/compilers/bin/nvcc /share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/compilers/bin/nvc [hmeij@n79 ~]$ which mpirun mpicc /share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/comm_libs/mpi/bin/mpirun /share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/comm_libs/mpi/bin/mpicc # no cublas for 10.2 [hmeij@n79 ~]$ find /share/apps/CENTOS7/nvidia/hpc_sdk/22.2/ -type f -name libcublas.so /share/apps/CENTOS7/.../math_libs/11.6/targets/x86_64-linux/lib/stubs/libcublas.so /share/apps/CENTOS7/.../math_libs/11.2/targets/x86_64-linux/lib/stubs/libcublas.so /share/apps/CENTOS7/.../math_libs/11.0/targets/x86_64-linux/lib/stubs/libcublas.so
And with the module loaded in your script, it will work on queue exx96
# gcc plus SDK export PATH=/share/apps/CENTOS7/gcc/9.2.0/bin:$PATH export LD_LIBRARY_PATH=/share/apps/CENTOS7/gcc/9.2.0/lib64:\ /share/apps/CENTOS7/gcc/9.2.0/libexec:\ /share/apps/CENTOS7/gcc/9.2.0/lib:\ $LD_LIBRARY_PATH module use /share/apps/CENTOS7/nvidia/hpc_sdk/modulefiles/nvhpc/22.2 module load 22.2 # which and wrapper stdout /share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/compilers/bin/nvcc /share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/comm_libs/mpi/bin/mpirun /share/apps/CENTOS7/python/3.8.3/bin/python GPU allocation n79:1 executing: /share/apps/CENTOS7/nvidia/hpc_sdk/22.2/Linux_x86_64/22.2/comm_libs/mpi/bin/mpirun \ -x LD_LIBRARY_PATH -machinefile /zfshomes/hmeij/.lsbatch/mpi_machines.844864 \ -np 1 pmemd.cuda -O -o mdout.844864 -inf mdinfo.1K10 -x mdcrd.1K10 -r restrt.1K10 -ref inpcrd