TAG: a Tiny AGgregation Service for Ad-Hoc Sensor Networks

This paper presents a system for mapping declarative SQL-like queries onto a sensor network and computing results efficiently. Because TAG aggregates data inside the network, it can reduce the messaging costs over the simple approach of just sending all the data to one receiver, which runs the query. The paper builds trees for routing and also uses caching of values to improve reliability and performance. It is evaluated through simulations but there is also a real implementation.

Although the system is neat and sounds like a good interface for querying sensor networks once they're up, I have the following questions / concerns:

1. Often we think of queries only after collecting some data, so we want to save all the data in a centralized repository anyway. In this case, is it worth trying to build a system like TAG rather than focusing on just a fast data collection service? As an added bonus, although collecting raw data costs more power than running TAG, once we have the data we can run any number of queries for free, whereas with TAG, every query would cost some more power and the total cost may be higher.
2. The paper uses a simulator for most of the results, which may lead to better performance than in reality due to the model in the simulation. Why wasn't the real implementation used for more measurements?