Increasing Aggregation Convergecast Data Collection Frequency through Pipelining

Abstract

We consider the problem of increasing the data collection frequency of aggregation convergecast. Previous studies attempt to increase the data collection frequency by shortening the completion of a single data collection cycle. We aim at increasing the frequency at which data collection updates are collected by the use of pipelining and, consequently, increasing the overall data collection frequency and throughput. To achieve this, we overlap the propagation schedule of multiple data snapshots within the same overall schedule cycle, thus increasing parallelism through pipelining. Consequently, the effective data collection time of an individual snapshot may span over multiple, successive, schedule cycles. To this end, we modify the aggregation convergecast model, decoupling schedule length, and data collection delay, by relaxing its precedence constraints. Our solution for this new problem involves the unconventional approach of constructing the schedule before finalizing the exact form of the data aggregation tree, which, in turn, requires that the schedule construction phase guarantees that every node can reach the sink. We compare our results using snapshot pipelining against a previously proposed algorithm that also uses a form of pipelining, as well as against an algorithm that though lacking pipelining, exhibits the ability to produce very short schedules. The results confirm the potential to achieve a substantial throughput increase, at the cost of some increase in latency.