Low level parallelization of nonlinear diffusion filtering algorithms for cluster computing environments

Abstract

This paper describes different low level parallelization strategies of a nonlinear diffusion filtering algorithm for digital image denoising. The nonlinear diffusion method uses a so-called additive operator splitting (AOS) scheme. This algorithm is very efficient, but requires frequent data exchanges. Our focus was to provide different data decomposition techniques which allow for achieving high efficiency for different hardware platforms. Depending on the available communication performance, our parallelization schemes allow for high scalability when using fast System Area Networks (SAN), but also provide significant performance enhancements on slower interconnects by optimizing data structures and communication patterns. Performance results are presented for a variety of commodity hardware platforms. Our most important result is a speedup factor of 210 using 256 processors of a high end cluster equipped with Myrinet. © Springer-Verlag 2003.