In ACM/IFIP/USENIX 13th International Conference on Middleware, 2012.
Acceptance rate: 19%.
Fine-grained filtering capabilities prevalent in content-based Publish/Subscribe (pub/sub) overlays lead to scenarios in which publications pass through brokers with no matching local subscribers. Processing of publications at these pure forwarding brokers amounts to inefficient use of resources and should ideally be avoided. This paper develops an approach that largely mitigates this problem by building and adaptively maintaining a highly connected overlay mesh superimposed atop a low connectivity primary overlay network. While the primary network provides basic end-to-end forwarding routes, the mesh structure provides a rich set of alternative forwarding choices which can be used to bypass pure forwarding brokers. This provides unique opportunities for load balancing and congestion avoidance. Through extensive experimental evaluation on the SciNet cluster and PlanetLab, we compare the performance of our approach with that of conventional pub/sub algorithms as baseline. Our results indicate that our approach improves publication delivery delay and lowers network traffic while incurring negligible computational and bandwidth overhead. Furthermore, compared to the baseline, we observed significant gains of up to 115% in terms of system throughput.