A good data management architecture should achieve energy-efficiency
and low-latency performance even in large scale networks. Our first
set of scalability experiments test PRESTO at different system scales
on five days of data collected from the James Reserve
deployment. Queries arrive at the proxy as as a Poisson process at the
rate of one query/minute per sensor. The confidence interval of
queries is chosen from a normal distribution, whose expectation is
equal to the push threshold,
.
Figure 4 shows the query latency and query drop rate at system sizes ranging from 40 to 120. For system sizes of less than 100, the average latency is always below five seconds and has little variation. When the system size reaches 120, the average latency increases five-fold to 30 seconds. This is because the radio transceiver on the proxy gets congested and the queue overflows.
The effect of duty-cycling on latency is seen in Figure 4, which shows that the maximum latency increases with system scale. The maximum latency corresponds to the worst case of PRESTO when a sequence of query misses occur and result in pulls from sensors. This results in queuing of queries at the proxy, and hence greater latency. An in-network querying mechanism such as Directed Diffusion [11] that forwards every query into the network would incur even greater latency than the worst case in PRESTO since every query would result in a pull. These experiments demonstrate the benefits of model-driven pushes for user queries. By the use of caching and models, PRESTO results in low average-case latency by providing quick responses at the proxy for a majority of queries. We note that the use of a tiered architecture makes it easy to expand system scale to many hundreds of nodes by adding more PRESTO proxies.