List Ranking on PC Clusters

We present two algorithms for the List Ranking Problem in the Coarse Grained Multicomputer model (CGM for short): if $p$ is the number of processors and $n$ the size of the list, then we give a deterministic one that achieves $O(\log p \log^* p)$ communication rounds and $O(n \log^* p)$ for the required communication cost and total computation time; and a randomized one that requires $O(\log p)$ communication rounds and $O(n)$ for the required communication cost and total computation time. We report on experimental studies of these algorithms on a PC cluster interconnected by a Myrinet network. As far as we know, it is the first portable code on this problem that runs on a cluster. With these experimental studies, we study the validity of the chosen CGM-model, and also show the possible gains and limits of such algorithms for PC clusters