Performance analysis of a CBIR system on shared-memory systems and heterogeneous clusters

Abstract

This paper presents a parallel implementation of a CBIR system which has been designed for several multiprocessor architectures. The paradigm that best fits the problem's needs is a farm based solution: a master process distributes the work load among the slave processes, and when they have finished, the master collects the partial results computed on each slave process. The multiprocessors used in the tests are a shared memory machine with eight processors - a Silicon Graphics Origin 2000 - and a cluster with 16 PC. In order to evaluate both architectures, the experimental results have been collected under several configurations, considering dynamic load balancing for the cluster based implementation. This paper analyzes the performance achieved by the different setups from the viewpoints of speed, price and scalability, presenting the conclusions that can be extracted from the results' comparison.