Recent years have witnessed significant progress in parallel hardware architecture and programming interface. To fully utilize these parallel architectures and APIs, the algorithms need to be parallel as well. However, research in parallel algorithms so far has been mainly focused on methods that guarantee consistent results between sequential and parallel computations. Even though such strict consistencies are necessary for critical application domains, they might be over-strict for other less serious applications such as graphics, imaging, or statistics. We propose the terminology {\em semantic consistency} to summarize such loose application-specific requirements. We describe concrete examples, ranging from statistics, imaging, and graphics, on how semantic consistency facilitates the design of parallel algorithms that are infeasible under traditional stricter consistency models. We believe the notion of semantic consistency would open further possibilities for the development of parallel algorithms and applications.

PowerPoint presentation

ZIP compressed file

In: ISC 2009

Details