This is an artifact for accepted paper 163 of Euro-Par 2020, "Improving mapping for sparse direct solvers: A trade-off between data locality and load balancing" by authors Changjiang Gou, Ali Al Zoobi, Anne Benoit, Mathieu Faverge, Loris Marchal, Grégoire Pichon, and Pierre Ramet.

Contents

• overview, which introduces the basic softwares and details steps on how to reproduce experiments.

• raw_results, which includes all raw results and the R script to clean the results and to plot curves.

• scripts, which contains all bash scripts used to generate raw datas and to download matrices.

• pastix_sharedM, an empty directory, where PaStiX is supposed to be installed.

• pastix_distributed, an empty directory, where PaStiX is supposed to be installed, with MPI option on.

• matrice, where the matrices used in our experiments should be placed.

More informations can be found in file overview to reproduce the results.

Getting started

• Install dependencies

sudo apt-get install git cmake g++ gfortran \
	                 libhwloc-dev libscotch-dev \
					 libopenblas-dev liblapacke-dev \
				     python-numpy openmpi-bin

• Setup the variable for the location of the artifact directory, and let's consider everything will be done from that directory if not specified.

export ARTIFACTS_DIR=path_to_the_artifact_directory
cd $ARTIFACTS_DIR

• Clone the repository and backup the directory for future uses

git clone -b europar2020 --recursive https://gitlab.inria.fr/solverstack/pastix.git

• Let's create the shared memory build and test it

cd $ARTIFACTS_DIR/pastix_sharedM
cmake $ARTIFACTS_DIR/pastix -DPASTIX_WITH_MPI=OFF -DPASTIX_INT64=OFF
make -j 8
./example/simple -9 20:20:20 

• Let's create the distributed memory build and test it

cd $ARTIFACTS_DIR/pastix_distributed
cmake $ARTIFACTS_DIR/pastix -DPASTIX_WITH_MPI=ON -DPASTIX_INT64=OFF
make -j 8
mpirun -np 2 ./example/simple -9 20:20:20 

Reproduce results

• Download the matrices used. This may take some time and a lot of memory.

cd $ARTIFACTS_DIR/scripts
./download_matrices.sh

• Analysis results on a shared memory context on an example matrix

./pastix_sharedM/example/analyze -3 ./matrice/1138_bus.mtx -v4 -i iparm_allcand 0

• Analysis results on a distributed context on an example matrix

mpirun -np 6 ./pastix_distributed/example/analyze -3 ./matrice/1138_bus.mtx -v4 -t 4 -i iparm_allcand 0

• Factorization on a shared memory context on an example matrix

./pastix_sharedM/example/bench_facto -3 ./matrice/1138_bus.mtx -v4 -i iparm_allcand 0

• Analysis results on a shared memory context

cd $ARTIFACTS_DIR/scripts
./basic_forloop_sharedM.sh

• Analysis results on a distributed context

cd $ARTIFACTS_DIR/scripts
./basic_forloop_distributed.sh

• Factorization time on a shared memory context

cd $ARTIFACTS_DIR/scripts
./run_all_matrix_factor.sh

• Generate plots

cd $ARTIFACTS_DIR/raw_results
Rscript ./CombineData_Plot.R