parlab-tuwien / omp-task-bench

OMPTB Benchmark Suite

OMPTB Benchmark Suite

Although OpenMP is heavily used to parallelize for-loops, it also supports task-parallel programming, which is important for parallelizing irregular applications. In this work, we focus on the performance of OpenMP runtime systems for task-based applications. In particular, we investigate the performance of different OpenMP runtime systems when scheduling a large set independent tasks of different granularity. To that end, we propose the OpenMP benchmark OMPTB, which features profiling and tracing options that help developers to reason about the observed performance differences.

This repository contains the code of our proposed benchmark suite OMPTB. As mentioned above, the goal of this benchmark suite is to compare the task implementation of different OpenMP runtime systems using a workload that is the same in all experiments, making differences in execution time only dependent on a compiler's runtime system implementation. Thus, the repository is structured into two parts:

• tasklib: Workload library, contains the function add_bench(n), used to create an artificial workload, intended to be built with the same compiler for all benchmarked runtime systems.
• taskbenchmark: Benchmark harness, provides the infrastructure for running benchmarks and recording results.

The OMPTB benchmark suite has been confirmed to function as intended for the following compilers:

• tasklib: gcc (11.2.0, 12.1.0)
• taskbenchmark: gcc (11.2.0, 12.1.0), clang (14.0.4), aocc (3.2.0), pgcc (22.5), icc (2021.7.0)

This work is licensed under the GNU General Public License Version 3.

Publications

• A Quantitative Analysis of OpenMP Task Runtime Systems (link TBD) - Sascha Hunold, Klaus Kraßnitzer / Bench 2022

Building and Executing

This section describes how to build the OMPTB benchmark suite to facilitate benchmarking different OpenMP runtime systems. We first build the tasklib library using gcc and then proceed by building the taskbenchmark executable using different CMake-supported compilers while statically linking the previously compiled tasklib library to ensure a uniform workload for runtime system comparison.

Note: In CMake, the used C compiler is set using the argument -DCMAKE_C_COMPILER=

1. Build tasklib with gcc:

$ cd tasklib
$ cmake -DCMAKE_C_COMPILER=gcc .
$ make

2. Build taskbenchmark with compiler COMPILER:

$ cd taskbenchmark
$ cmake -DCMAKE_C_COMPILER=COMPILER .
$ make

After building taskbenchmark, the executable omp_bench is found in taskbenchmark/bin/omp_bench

Acknowledgements

The authors would like to thank Lukas Briem (@briemelchen) for the contributions to this work.