This paper presents PRESTO, a two-tier sensor architecture that comprises sensor proxies at the higher tier, each controlling tens of remote sensors at the lower tier. PRESTO1 proxies and sensors interact and cooperate for acquiring data and processing queries [4]. PRESTO strives to achieve energy efficiency and low query latency by exploiting resource-rich proxies, while respecting constraints at resource-poor sensors. Like TinyDB, PRESTO puts intelligence at the edge proxies while keeping the sensors inside the network simple. A key difference though is that PRESTO endows sensors with the ability to asynchronously push data to proxies rather than solely relying on pulls. Our design of PRESTO has led to the following contributions.
Model-driven Push: Central to PRESTO is the use of a feedback-based model-driven push approach to support queries in an energy-efficient, accurate and low-latency manner. PRESTO proxies construct a model that captures correlations in the data observed at each sensor. The remote sensors check the sensed data against this model and push data only when the observed data deviates from the values predicted by the model, thereby capturing anomalous trends. Such a model-driven push approach reduces communication overhead by only pushing deviations from the observed trends, while guaranteeing that unusual patterns in the data are never missed. An important requirement of our model is that it should be very inexpensive to check at resource-poor sensors, even though it can be expensive to construct at the resource-rich proxies. PRESTO employs seasonal ARIMA-based time series models to satisfy this asymmetric requirement.
Support for archival queries: Whereas PRESTO supports queries on current data using model-driven push, it also supports queries on historical data using a novel combination of prediction, interpolation, and local archival. By associating confidence intervals with the model predictions and caching values predicted by the model in the past, a PRESTO proxy can directly respond to such queries using cached data so long as it meets query error tolerances. Further, PRESTO employs interpolation methods to progressively refine past estimates whenever new data is fetched from the sensors. PRESTO sensors also log all observations on relatively inexpensive flash storage; the proxy can fetch data from sensor archives to handle queries whose precision requirements can not be met using the local cache. Thus, PRESTO exploits the proxy cache to handle archival queries locally whenever possible and resorts to communication with the remote sensors only when absolutely necessary.
Adaptation to Data and Query Dynamics: Long-term changes in data trends are handled by periodically refining the parameters of the model at the proxy, which improves prediction accuracy and reduces the number of pushes. Changes in query precision requirements are handled by varying the threshold used at a sensor to trigger a push. If newer queries require higher precision (accuracy), then the threshold is reduced to ensure that small deviations from the model are reported to the proxy, enabling it to respond to queries with higher precision. Overall, PRESTO proxies attempt to balance the cost of pushes and the cost of pulls for each sensor.
We have implemented PRESTO using a Stargate proxy and Telos Mote sensors. We demonstrate the benefits of PRESTO using an extensive experimental evaluation. Our results show that PRESTO can scale up to one hundred Motes per proxy. When used in a temperature monitoring application, PRESTO imposes an energy requirements that is one to two orders of magnitude less than existing techniques that advocate on-demand, proactive, or model-driven pulls. At the same time, the average latency for queries is within six seconds for a 1% duty-cycled five hop sensor network, which is an order of magnitude less than a system that forwards all queries to remote sensor nodes, while not significantly more than a system where all queries are answered at the proxy.
The rest of this paper is structured as follows. Section 2 provides an overview of PRESTO. Sections 3 and 4 describe the design of the PRESTO proxy and sensors, respectively, while Section 5 presents the adaptation mechanisms in PRESTO. Sections 6 and 7 present our implementation and our experimental evaluation. Finally, Sections 8 and 9 discuss related work and our conclusions.