Performance of Model-Driven Push

In this section, we validate our claim that intelligently exploiting both proxy and sensor resources offers greater energy benefit than placing intelligence only at the proxy or only at the sensor. We compare the performance of model-driven push used in PRESTO against two other data-acquisition algorithms. The first algorithm, model-driven pull, is representative of the class of techniques where intelligence is placed solely at the proxy. This algorithm is motivated by the approach proposed in BBQ [3]. In this algorithm, the proxy uses a model of sensor data to predict future data and estimate the confidence interval in the prediction. If the confidence interval exceeds a pre-defined threshold ($\delta$), the proxy will pull data from the sensor nodes, thus keeping the confidence interval bounded. The sensor node is simple in this case, and performs neither local storage nor model processing. While BBQ uses multi-variate Gaussians and dynamic Kalman Filters in their model-driven pull, our model-driven pull uses ARIMA predictions to ensure that the results capture the essential difference between the techniques and not the difference between the models used. The second algorithm that we compare against is a relatively naive value-driven push. Here, the sensor node pushes the data to the proxy when the difference between current data and last pushed data is larger than a threshold ($\delta$). The proxy assumes that the sensor value does not change until the next push from the sensor. In general, a pull requires two messages, a request from the proxy to the sensor and a response, whereas push requires only a single message from the sensor to the proxy.

We compared the three techniques using Matlab simulations that use real data traces from James Reserve. Each experiment uses 5 days worth of data and each data point is the average of 10 runs. Figure 3 compares these three techniques in terms of the number of messages transmitted and mean-square error of predictions. In communication cost, PRESTO out-performs both the other schemes irrespective of the choice of $\delta$. When $\delta$ is 100, the communication cost of PRESTO is half that of model-driven pull, and 25% that of value-driven push. At the same time, the mean square error in PRESTO is 30% that of model-driven pull, and 60% that of value driven push. As $\delta$ decreases, the communication cost increases for all three algorithms. However, the increase in communication cost for model-driven pull is higher than that for the other two algorithms. When $\delta$ is 50, value driven push begins to out perform model-driven pull. When $\delta$ reaches 10, the number of messages in model-driven pull is 20 times more than that of PRESTO, and 8 times more than that of value driven push. This is because in the case of model-driven pull, the proxy pulls samples from the sensor whenever the prediction error exceeds $\delta$. However, since the prediction error is often an overestimate and since each pull is twice as expensive as a push, this results in a larger number of pull messages compared to PRESTO and value-driven push. The accuracies of the three algorithms become close to each other when $\delta$ decreases. When $\delta$ is smaller than 40, model-driven pull has slightly lower mean square error than PRESTO but incurs 4 times the number of messages.

Summary: These performance numbers demonstrate that model-driven push combines the benefits of both proxy-centric as well as sensor-centric approaches. It is 2-20 times more energy-efficient and upto 3 times more accurate than proxy-centric model-driven pull. In addition, PRESTO is upto 4 times more energy-efficient than a sensor-centric value-driven push approach.