Determining the optimal level of parallellism for realtime multimedia applications in a Grid environment

Over the past couple of years, the use of large-scale multimedia data applications has been growing tremendously, and this growth is not likely to come to an end in the near future. Multimedia applications operated in a real-time environment pose very strict requirements on the processing times while offline applications have to classify and index the content of massive multimedia archives within a tolerable time frames. To meet these requirements, multimedia aplications typically run in a parallel processing environment. In this context, it is essential to determine the optimal number of parallel compute nodes, properly dealing with the trade-off between computation power on the one hand and communication overhead on the other hand, which generally depends on the characteristics of the applications and the software and hardware specifics of the computational environment. Motivated by this, in this paper we develop a simple and easy-to-implement method to determine the “optimal” number of parallel compute nodes. The method is based on the classical binary search method for non-linear optimization, and does not depend on the, usually unknown, specifics of the system. Extensive experimental validation in our DAS-3 testbed shows that the method is indeed highly effective