On Summit, a limited number of dependecies are available by default, and the modules
tend to change somewhat frequently. We have a script ./scripts/summitVariables.sh
which loads a combination of system modules and modules built in the ExaSGD
shared project folder which enables a full build of HiOp.
$ git clone [email protected]:LLNL/hiop.git
$ cd hiop
$ source ./scripts/summitVariables.sh
$ mkdir build install && cd build
$ cmake .. -DCMAKE_INSTALL_PREFIX=$PWD/../install
$ ccmake . # Optional, if you would like to customize the build
$ make -j 16 install
$ make test # Optional, if you would like to run tests. This command should be
$ # ran in a BSUB allocation.
Note that this script may become out of date, and is not officially supported.
Different versions of modules may conflict with each other; in this case, one needs to inquire additional compatibility information using module spider package_name
.
HiOp with GPU support can be configured, built, and tested using with the cmake
options HIOP_USE_GPU
and HIOP_USE_CUDA
. For example, you may invoke cmake
like so:
cd build/
cmake -DHIOP_USE_GPU=ON -DHIOP_USE_CUDA=ON ..
make -j 16
make test
make install
In some cases, cmake
picks up a different gcc compiler than the one loaded via module
. This situation can be remedied by preloading the CC
and CXX
for cmake
as it is done above.
Developers are usually required to perform a more comprehensive build like so:
rm -rf *
CC=/sw/summit/gcc/9.2.0/bin/gcc CXX=/sw/summit/gcc/9.2.0/bin/g++ cmake -DHIOP_SPARSE=ON -DHIOP_USE_GPU=ON -HIOP_TEST_WITH_BSUB=ON -DHIOP_USE_MPI=ON -DMETIS_DIR=$INSTALL_DIR/Compiler/gcc-9.2.0/metis/5.1.0/ -DHIOP_COINHSL_DIR=/ccs/home/cpetra/work/installs/coinhsl-2015.06.23/_install -DCMAKE_BUILD_TYPE=DEBUG -DHIOP_TEST_WITH_BSUB=ON -DHIOP_USE_RAJA=ON -DHIOP_USE_UMPIRE=ON -DHIOP_USE_GPU=ON -DHIOP_DEEPCHECKS=ON ..
make -j
Depending on the LAPACK/BLAS CPU library loaded by the module
command, runtime crashes can occur (intermintently for different problems/linear systems sizes) if the nvblas.conf
cannot be find or is not specifying the so-called NVBLAS CPU BLAS library required by CUDA. This can be remedied by specifying the location of the BLAS library in NVBLAS_CPU_BLAS_LIB
in the nvblas.conf
file; for example, this can be done by adding the following lines to nvblas.conf
file
NVBLAS_CPU_BLAS_LIB /gpfs/alpine/proj-shared/csc359/installs/ExaSGD/Compiler/gcc-9.2.0/openblas/0.3.10/lib/libopenblas.so
Be aware that this assumes that the openblas has been module load
ed.
For non-standard locations of nvblas.conf
, one can specify the path to nvblas.conf
via the environment variable NVBLAS_CONFIG_FILE
; for example, this can be achieved by exporting the variable in your submission your_job_file.lsf
file:
# ##LSF directives here
# other commands here
# example assumes a bash shell
export NVBLAS_CONFIG_FILE=/ccs/home/cpetra/work/projects/hiop/runs/nvblas.conf
# jsrun command here
An Spack recipe is available in the builtin spack repository. To build HiOp on Summit using Spack requires some experience with Spack. Users are referred to the Spack documentation here.