Improved On Demand Clustering on Scale-Free Topologies

We define and analyze two directions to improve On Demand Clustering (ODC) in self-organizing networks. The first direction, Spyglass, uses excessive lookup in the match maker's two-hop neighborhood. The second direction, Shuffling ODC, applies link reconfiguration also in case of unsuccessful search, causing minor modifications in the adjacent topology without immediate gain. Both directions were evaluated with simulation using scale-free graphs as network abstraction. Spyglass results in significantly faster reaction times and larger clusters than the bask ODC, but with a significantly higher communication overhead, Shuffling ODC brings similar advantages, significantly larger clusters and less separated nodes, but without the increment in the message count. Shuffling ODC, due its emergent self shuffling effect, is able to reach more nodes than the other two, without breaking the rule of locality. We found 20-30% improvement in job processing and 35-50% gain in job sharing with our algorithms, compared to the baseline ODC.