MPI-I-2000-1-002

A refined heuristic for computing schedules for gossiping in the telephone model is presented. The heuristic is fast: for a network with n nodes and m edges, requiring R rounds for gossiping, the running time is O(R n log(n) m) for all tested classes of graphs. This moderate time consumption allows to compute gossiping schedules for networks with more than 10,000 PUs and 100,000 connections. The heuristic is good: in practice the computed schedules never exceed the optimum by more than a few rounds. The heuristic is versatile: it can also be used for broadcasting and more general information dispersion patterns. It can handle both the unit-cost and the linear-cost model. Actually, the heuristic is so good, that for CCC, shuffle-exchange, butterfly de Bruijn, star and pancake networks the constructed gossiping schedules are better than the best theoretically derived ones. For example, for gossiping on a shuffle-exchange network with 2^{13} PUs, the former upper bound was 49 rounds, while our heuristic finds a schedule requiring 31 rounds. Also for broadcasting the heuristic improves on many formerly known results. A second heuristic, works even better for CCC, butterfly, star and pancake networks. For example, with this heuristic we found that gossiping on a pancake network with 7! PUs can be performed in 15 rounds, 2 fewer than achieved by the best theoretical construction. This second heuristic is less versatile than the first, but by refined search techniques it can tackle even larger problems, the main limitation being the storage capacity. Another advantage is that the constructed schedules can be represented concisely.