Distribution of image processing applications on a heterogeneous workstation network: modeling, load balancing, and experimental results

Abstract

This work analyzes the computation distribution in applications generated by a multilevel knowledge-based system for image processing called SVEX. This distribution has been carried out on a heterogeneous workstation network, trying to take advantage of the availability and frequent infra- utilization of this computational resource. The parallelization is based on message-passing tool parallel virtual machine (PVM). Firstly SVEX and its computational scheme are described, detailing the structure of the first level (the pixel processor). Then different distribution paradigms are studied, selecting for its implementation the parallelism based on the data. Considering this alterative, the research addresses two fundamental problems: analysis of basic load-balancing schemes and obtaining a model for predicting parallelization behavior as new machines are added to the computational network. The results produced in a series of experiments permit the comparison of load-balancing schemes and the validation of the proposed model. The experiments include the processing of both static images and sequences.