Partitioning of image processing tasks on heterogeneous computer systems

Many computer vision tasks can be decomposed into a set of subtasks which are by their nature heterogeneous. By partitioning such tasks onto different machines that communicate via high-speed links, each level or stage of processing can be executed simultaneously on the machine to which it is best suited. A fundamental problem with heterogeneous computing, however, is the difficulty of optimally partitioning an application program across the machines. In this paper, we address the problem of partitioning a chain or a tree-structured parallel or pipelined program over a two-processor heterogeneous system and show that it is possible to approximately solve this problem. The algorithm presented in this paper is based on a fully polynomial time approximation scheme