Comparison of OpenMP and Classical Multi-Threading Parallelization for Regular and Irregular Algorithms
Claude Timsit, Plateau de Moulon · 2000
The new emerging Distributed Shared Memory architecture promises to be more scalable than Symmetric Multiprocessor architecture, and leads to a regain of interest for parallel shared-memory programming paradigms. This paper compares two such important paradigms: classical multithreading and multi-threading based on compiler directives (with OpenMP). Several implementations of regular and irregular algorithms, taken from artificial intelligence field, were made on an SGI-Origin2000 (a DSM architecture) and compared both in terms of development time and of execution time. Finally, we identify the most appropriate paradigm for each kind of algorithm.