Evaluating and Improving Weak Scalability Analysis of Visualization Algorithms

This is a short description on how to use the accompanying code to replicate the experiments. The default values are chosen s.t. it should be able to run on a smaller machine. Thus, scaling is done only up to 4 cores, but this can be changed. Instructions are written below. This example was tested on a machine running Ubuntu 22.04 and the execution environment, available via docker container, is built for amd64 platforms.

Prerequisites

1. You have to be able to execute docker containers - apptainer works too but may require CPU and CPUSET delegation for non-root users.
2. You have to be able to execute bash scripts
3. Disable hyperthreading/SMT

Running the experiment

1. Execute the following command in this directory bash scripts/get_data.sh to download/create the data necessary for the experiments.
2. Run your desired experiments e.g.\ bash scripts/all_contour_benchmarks.sh,\ bash scripts/all_contourtree_benchmarks.sh,\ bash scripts/all_volumerender_benchmarks.sh
3. Run the following command to generate the plots:\ bash scripts/all_plots.sh

The plots and experiment data are in the output directory.

Changing the number of cores

You can change this in the scripts/all_<algorithm>_benchmark.sh files. There you need to change the threads list.

For extent scaling you need a larger data set if you increase the number of cores. This data set is generated in scripts/get_data.sh. You can adjust the MAX_SCALING_FACTOR parameter to a value equal or larger than the maximum number of cores in your experiments.

Distributed experiment

The scripts provided here use modules, slurm and require a specific version of openmpi (4.1.6).

1. Execute the following command in this directory bash scripts/distributed/get_data_distributed.sh to download/create the data necessary for the experiments.
2. Change the partition that is given in scripts/distributed/<algorithm>_benchmark_distributed.sh files to one of your system.
3. Run your desired experiments e.g.\ bash scripts/distributed/all_contour_benchmarks_distributed.sh
4. Run the following command to generate the plots:\ bash scripts/distributed/all_plots_distributed.sh

The plots and experiment data are in the output_distributed directory.

Changing the number of ranks and or cores-per-rank

You can change this in the scripts/distributed/all_<algorithm>_benchmark_distributed.sh files. There you need to change the threads list.

For extent scaling you need a larger data set if you increase the number of cores. This data set is generated in scripts/distributed/get_data_distributed.sh. You can adjust the MAX_SCALING_FACTOR parameter to a value equal or larger than the maximum number of cores in the experiments.

For the distributed experiments you can also change the cores per rank via CORES_PER_RANK in scripts/distributed/<algorithm>_benchmark_distributed.sh.