OpenMP Benchmarks

This repository contains reproduction scripts and collected results from master thesis titled "Assessment of OpenMP constructs' performance on modern multi-core architectures".

Folder Structure

• compilers/ - this folder contains script which will download, build and install GCC and Clang compilers and download and install ICC and ICX compilers
• offloading_compilers/ - this folder contains script which will download, build and install GCC and Clang compilers with offloading to NVIDIA GPU's
• openmp_benchmarks/ - this folder contains C/C++ benchmarks code, benchmarks descriptors, and scripts to run these benchmarks

Requirment

To run these benchmarks following packages are required (tested on ubuntu docker image):

• build-essential
• git
• wget
• g++/gcc >= 7.3

Building OpenMP benchmarks

Building script has been prepared and is located in the openmp_benchmarks folder: Usage:

./linux_build.sh [-c=<compiler>] [-t]

• -c, --compiler, Build benchmarks with specified compilatior

• -t, --target, Build OpenMP benchmarks with offloading directives - this option requires to have compilers supporting offload

Runing OpenMP benchmarks

To run OpenMP benchmarks use following command:

OMP_NUM_THREADS=<NUM_THREADS> ./build/OpenMPBench <descriptor_path> <outfile_csv_path>

• NUM_THREADS - how many OpenMP threads should be used
• descriptor_path - path to the descriptor file that should be used - one of the already prepare descriptor for master thesis can be used (openmp_benchmarks/descriptors/)
• outfile_csv_path - path to the output file with benchmarks results

Building compilers

To build compilers from source special script has been prepared. It's called prepare.sh and is located in both directories: compilers/ and offloading_compilers/: Usage:

bash prepare.sh

Remember to edit CUDA installation path in prepare.sh script in offloading_compilers/ directory

Results

Results collected for master thesis are in openmp_benchmarks/results/ directory. They can be viewed in raw files, but special Jupyter Notebook was prepared to visualize results as a charts.

Instruction

• Download and install conda environment (skip if already done):

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
sh Miniconda3-latest-Linux-x86_64.sh

• Craete and activate new environment to keep base env. clean:

conda create -n openmp_results
conda activate openmp_results

• Install dependencies required by the script:

conda install -c conda-forge notebook  
conda install plotly matplotlib pandas

• Run jupyter-notebook by following command:

jupyter-notebook

• Open link which appeared in terminal (e.g. http://127.0.0.1:8888/?token=21312....
• In the browser there will be listed all files from folder where notebook was run. Open openmp_benchmarks/processing_results.ipynb file
• Before running the script, modify the following settings to generate results in the desired form. It is possible to generate charts only for a specific benchmark by editing the benchmarks variable in the script or for a specific processor using the CPUS variable. It is also possible to generate charts showing acceleration relative to sequential execution, or showing average execution time with standard deviation - this option is controlled by the speedup variable variable, set to True or False. Additionally, it is possible to separate charts for compilers using the separate_compilers variable, which is especially useful useful when one chart contains too much information.
• Run script by combination of Ctrl+A and Shift + Enter keys or by using the icons on the top of the page