Congratulations to George Exarchakos for successfully passing his PhD viva

Wednesday 25 March 2009

Heterogeneous distributed applications deployed on different networks may have variable network throughput requirements during their lifetime frequently swapping between underloaded and overloaded situations. This study proposes a model for on-demand capacity sharing among networks allowing their sizes to adapt based on their workload fluctuations. Its aim is to keep all nodes normally loaded and reduce the messaging cost of deployed applications and focuses on those with highly dynamic workload fluctuations, bursts of traffic and/or massive node failure rates.

Conceptual migration of capacity from one network to another may improve spare capacity utilization and network reliability even in high workload situations. Hoverlay is the proposed two-tier Unstructured P2P management architecture which realises capacity sharing between networks in three steps: publishing, discovery, and commissioning of capacity. While the first two refer to capacity publishing and discovery mechanisms the last one fetches and commissions it to the requesting network. This architecture facilitates the cooperation of heterogeneous networks each of which is represented by a single server on an interconnected server overlay.

The second contribution of this thesis is the proposed search mechanisms Stalkers deployed on Hoverlay to support the intermittent behaviour of service capacity. The idea behind is the use of fresh and/or incoming server-to-server links so that queries can trace resource migrations. There are three variations of this technique: k-Stalkers, FireStalkers and FloodStalkers aiming at improving number of messages, query latency and success rate.

A number of experiments were carried out and showed that Hoverlay increases node utilization and allows the underlying network to resize based on its workload. Node (capacity) migration between networks saves the overlay from multiple queries and increases query success rate reducing its average latency if compared to the competitive Flock of Condors system. A number of experiments to evaluate Stalkers mechanisms showed that all the three variations of Stalkers achieve their aim in environments of highly dynamic resources such as Hoverlay. They outperform Flooding and k-Walkers in success rate and latency keeping the number of messages relatively low. The thesis concludes with its findings and contributions as well as proposals for further work.