|
The Sensor Network as a Massively Parallel Distributed System Gul Agha
Computer science professor Gul Agha, whose Open Systems Laboratory developed the Actor model for developing and reasoning about distributed applications, continues to come up with novel ways to deal with all kinds of issues in parallel and distributed programs. They are often based on the Actor model. Actors are autonomous objects or software agents that move things from one computer to another while a program is running. (On the Internet, actors are known as softbots.) Now he is extending the notion of actors to the sensors that live in huge networks of sensors. Perhaps the Big Dig Tunnel disaster in July 2006 could have been avoided had a sensor network, like the one Agha envisions, been deployed throughout the tunnels. "Not only could a collapse potentially have been prevented," said civil engineering professor Bill Spencer, "but early detection would have facilitated preventive maintenance and kept the tunnel open to traffic." Agha and Spencer are collaborating on an NSF-sponsored research project to build a sensor network to monitor the structural health of civil infrastructure, such as buildings, bridges, and tunnels. They are applying parallel and distributed programming techniques to sensor networks, in which thousands of small computing devices, called smart sensors, collaborate with each other wirelessly. If the system works reliably, we can know, for example, if a structure is in danger of collapse, and we can know this in an easy, low-maintenance, and inexpensive way. The job of a sensor is to acquire information and transform it into electrical signals. Because a sensor element can be a resistor, capacitor, transistor, piezoelectric material (responsive to physical stress), photodiode (responsive to light), etc., sensors can be used to monitor a variety of properties. They can be embedded in the body of the structure, like a bridge or building, so that they can locally sense material properties. For example, a wave can be sent throughout the structure, and its distribution can be recorded at different points to assess damage. Sensors may be smart or dumb. Dumb sensors do no computing themselves but simply report their information to a base station. Because of the massive quantity of data that needs to be communicated in the case of dumb sensor, the use of such sensors is not scalable. Now we have smart sensors that contain microprocessors and wireless communication links, with data transmission based on radio frequency communication, the same thing used by cell phones and wireless Internet. In this case, the intelligent sensors are independent actors of a large-scale system. For such actors, time synchronization (clocks on sensors may show slightly different times), scheduling messages (too many sensors communicating with the base station at the same time will result in a bottleneck), faults, data correlation, battery life, and network issues are just some of the system's headaches. Current algorithms used to estimate damage based on dynamic structural characteristics assume that all data is centrally collected and processed in real-time, as is the case with dumb sensor systems. These algorithms cannot be implemented in the distributed computing environment employed by smart sensors. The sensor networks Agha and Spencer work with are smart. With many smart sensors, you have a massively parallel and distributed system. "There are ways to reason about large-scale things," reflected Agha. "Large-scale systems contain many nodes with many states, and the state of the network (the set of all possible states of the system) is the cross-product of the state of each node. If you have 100 nodes, and each has 3 states, then the entire thing would be 3100 , which would be impossible to evaluate." This is called the state explosion problem. The state explosion problem is yet another reason the traditional approach does not scale to systems that contain many sensors. Because sensor networks share characteristics with statistical physics, in which the states of molecules are calculated, Agha is examining the sensor network problem as a statistical abstraction. Viewed in this way, the state of the network can be calculated by the probability distribution of the state of each node-a vector of the probabilities. Agha is also exploiting the symmetry of the system, because most civil structures are symmetrical, by partitioning and reasoning about things based on configuration and the fact that a structure will exhibit different behaviors in different areas of the structure. Written by Judy Tolliver, July 28, 2006
|
|