An Analytical Approach to Optimizing Parallel Image Registration/Retrieval

Image-content queries or image registration algorithms typically have very high computational requirements. In this paper, we address the problem of minimizing the total execution time of data-parallel image-content query algorithms on heterogeneous platforms. The model we use to capture the inner workings of these algorithms is comprehensive enough to incorporate not only the communication overheads, both distribution and result collection, but also the presence of local data caches that could exist as a result of previous queries. The problem is solved under all possible computation and communication configurations, including single and multiple-port communications and block or stream-type tasks. Our analysis, either, yields closed-form solutions to the partitioning problem, or, formulates the problem in a fashion that allows the use of linear programming tools toward this end. The latter are used for solving the multi-installment data distribution approaches, that tend to utilize the computational resources more efficiently. Additionally, a heuristic algorithm is presented, for producing a close-to-optimum sequence of load distribution/result collection operations for single-port communications. Based on our analytical results, a thorough simulation and experimental study is performed, yielding substantial design guidelines for implementation strategies.