Wireless sensor networks represent an emerging technology that has great potential for widespread applications. These networks consist of a large number of simple nodes with limited power sources and functionality, but they offer far greater utility than the sum of those individual nodes. To maintain highly dynamic, ad-hoc networks with resource- constrained nodes, however, trade-offs among power consumption, robustness and responsiveness must be made.
The most common way to reduce the power consumption of sensor network nodes is to minimize the time spent in transmitting, listening and processing. But low-duty-cycle operation inevitably leads to slow network response, and eventually may result in network congestion or even instability. This article will analyze the design trade-offs between node duty cycle and system-level performance, pinpointing opportunities to improve power efficiency while maintaining network robustness and responsiveness.
Many factors influence the power consumption of wireless communication devices. In a sensor network node, the radio consumes more power than other components in the node. A narrowband radio with 0-dBm output typically draws about 15 milliamps in transmit (TX) mode at 3 volts dc and 5 to 15 mA in receive (RX) mode. Spread-spectrum radios tend to draw more current in RX than in TX.
Such current consumption is excessive to a sensor node supported only by a coin cell battery with no more than 220 mA-hr of capacity. To save battery power, the radio must be turned off as much as possible. But in an ad-hoc, multihop wireless sensor network, sensor nodes cooperate to maintain connectivity and forward traffic on behalf of one another, so turning off the radio for an extended period would certainly increase the likelihood of a node's missing and dropping packets, resulting in numerous retransmissions or even disconnection.
To save battery power without the expense of dropped packets and loss of connectivity, network nodes should turn off and power up the radio in a coordinated manner. To this end, several design issues across the media-access control (MAC), network and application layers in the protocol stack should be addressed.
For wireless sensor networks, one popular approach to saving power based on MAC-layer protocol design is to turn off the radio to prevent overhearing. For example, when Node A hears a packet being sent by Node B to Node C, Node A turns itself off so that it doesn't waste energy receiving packets not destined for it. The duration for a node to remain powered off should depend on the estimated length of the packet being transmitted. This approach is well-suited for MAC protocols that employ request-to-send/clear-to-send (RTS/CTS) control packets that contain recipient node ID and the data packet length, so that after hearing the RTS/CTS exchange, neighboring nodes can decide whether and for how long to turn off their radios.
For MAC protocols that don't employ RTS/CTS, however, it is difficult for a node to determine when to turn off and for how long. Another approach to saving power in the MAC layer is to schedule the radio to sleep and listen periodically; the shorter the listen duty cycle, the more power saved. This technique requires nodes to exchange sleep and listen schedules with neighboring nodes. To enable nodes to communicate based on each other's sleep-and-listen schedules, neighboring nodes must keep their clocks synchronized to a certain degree of accuracy. The clock drift rate of today's hardware is typically on the order of 10-6, or approximately 1 second of error in 11 days.
Question of responsiveness
Although this drift rate is small compared with typical listen-and-sleep cycles of a node, time synchronization among neighboring nodes is nonetheless required from time to time in order to prevent schedule mismatch. Substantial power can effectively be saved by adopting a long sleep cycle for every node, but a low duty cycle directly leads to slow network response and long acquisition time for a new node to join the network.
One possible remedy to resolving the responsiveness issue is to schedule listen- and-sleep cycles on a very short time scale. For instance, with a 50 percent listen duty cycle, each node may power up to listen for 1 second and sleep for 1 second; using a time scale of milliseconds, a node powers up for 1 millisecond and sleeps for 1 millisecond, resulting in the same 50 percent duty cycle, but with a 1,000x shorter wait for a node to wake up.
The limiting factor to reducing time scale is the increase in power consumption due to the extra overhead and the higher processing speed of the microcontroller. As the time scale gets shorter, a tighter synchronization must be maintained among sensor nodes. That requires more frequent exchanges of control packet overhead for time synchronization, as well as a faster microcontroller to generate and capture higher-rate signals.
Examples also abound for saving power based on the design of the network-layer protocol. In many cases, the network protocol design is coupled to the MAC-layer design for power saving. For instance, in a reactive ad-hoc routing algorithm such as the well-known Ad-hoc, On-demand Distance Vectoring (AODV), a router request packet is sent whenever a new route is needed. The network protocol can coordinate the timing of transmitting the route request with the periodical sleep-and-listen schedules of neighboring nodes, so that new routes can be established without numerous retrials, saving power and avoiding delay.
The network-layer protocol can also use the neighbor list maintained by each routing node to estimate node density in the neighborhood, so as to adjust its own periodic sleep-and-listen schedule as defined in the MAC layer. When node density is high, more nodes in the neighborhood can share the burden of routing traffic, and therefore each node on average can sleep longer without impacting connectivity and latency.
A clear understanding of application-level requirements can help drive networking strategies for power-efficient operations. In numerous real-world applications, not all sensor nodes are limited to operate solely on batteries. For instance, in building automation applications, a node connected to a variable-air- volume (VAV) box, which regulates ventilation airflow, can tap into the VAV box's line power for supply. A routing strategy that takes advantage of the heterogeneous distribution of power resources among sensor nodes proves to be the most effective in optimizing overall power efficiency without compromising network robustness and response speed.
Resource constraint nodes can operate with a substantially lower duty cycle than those having abundant power resources. These low-power nodes operate as endpoints in the network, which allows them to sleep most of the time and wake up periodically or be awakened by triggering events.
For low-data-rate applications in which sample data is taken perhaps once a minute, an endpoint can easily run on a 220-mA-hr coin cell battery for more than five years. High-power nodes serve as routers and remain in listening mode for a longer period of time to capture and route data from endpoints and other routers. Since endpoints only need to communicate with routers, they can organize themselves around routers into a star topology.
And to fulfill the essential function of routing network traffic, routers form a mesh topology to ensure robust and responsive network connectivity.
This hybrid star/mesh topology holds great promise for solving sensor-networking problems in a variety of areas.
Sokwoo Rhee (email@example.com) is CTO and Sheng Liu (firstname.lastname@example.org) is vice president of research at Millennial Net Inc. (Cambridge, Mass.).