SCALABLE PARALLEL IMPLEMENTATIONS OF LIST RANKING ON FINE-GRAINED MACHINES

We present analytical and experimental results for fine-grained list r anking algorithms. We compare the scalability of two representative al gorithms on random lists, then address the question of how the localit y properties of image edge lists can be used to improve the performanc e of this highly data-dependent operation. Starting with Wyllie's algo rithm and Anderson and Miller's randomized algorithm as bases, we use the spatial locality of edge links to derive scalable algorithms desig ned to exploit the characteristics of image edges. Tested on actual an d synthetic edge data, this approach achieves significant speedup on t he MasPar MP-I and MP-2, compared to the standard list ranking algorit hms. The modified algorithms exhibit good scalability and are robust a cross a wide variety of image types. We also show that load balancing on fine grained machines performs well only for large problem to machi ne size ratios.