Low-latency Synchronization of Loosely-coupled Sensornet Republishing

Today many individual deployments of sensornets are successful, but they will have much greater impact when, rather than standing alone, they share data across deployments so each can build upon the others. We expect data to be shared over the Internet, and as the number of processing and reprocessing steps grows, timely data synchronization is increasingly important. Today, such sharing is often hard-coded or driven by fixed-interval polling. Fixedinterval polling can provide poor worst-case performance (mean latency approaching the data generation period), and best performance requires careful manual configuration of both poll period and phase. We instead propose Data Generation Tracking (DGT), a new family of adaptive polling algorithms that learn and predict good times to pull data to minimize both latency and unfruitful queries. Our approach avoids manual configuration and automatically adapts to outages and changes in data generation rate. To evaluate our work, we examine four sensornet deployments and develop a rough model of sensornet data generation. We then use this model and replay of real traces to evaluate DGT, finding that, depending on application, its median latency is only 10–30% of that of fixed-interval polling, with a configurable rate of network load that is the same or slightly higher.