“If it’s difficult to transmit data reliably across a network of 50 nodes, what happens when you build a network of 500 or 5,000 sensors?” The question is posed by Lakshman Krishnamurthy, the Principal Investigator for Intel’s EcoSense research project. The project team is tackling a difficult challenge: how to network large numbers of inexpensive wireless sensor nodes while maintaining a high level of network performance.
Wireless sensor networks are formed by small nodes or “motes”— tiny, self-contained, battery-powered computers with radio links that enable the motes to self-organize into a network, communicate with each other and exchange data. As data “hops” from mote to mote across these networks (also referred to as “multihop” networks), information may be lost along the way.
“When data is sent from one node to the next in a multi-hop network, there’s a chance that a particular packet may be lost, and the odds grow worse as the size of the network increases,” say Krishnamurthy. “When a node sends a packet to a neighboring node, and the neighbor has to forward it, that takes energy. The bigger the network, the more nodes that must forward data, and the more energy that is consumed.” The end result: as the network grows, performance degrades.
The Potential Solution: Heterogeneous Networks
To address this performance problem, Intel researchers are exploring the concept of heterogeneous networks. The concept is simple: an 802.11 mesh network comprised of high-end nodes, such as Intel XScale® based nodes, is overlaid on a sensor network. The structure is analogous to a highway overlaid on a roadway system. Sensors can enter and exit the 802.11 highway at multiple interchanges (the XScale nodes) in order to bypass side roads (motes). Theoretically, this should enable faster trips across the network and result in improved performance. “The number of nodes the data has to pass through is much lower, so you should get more reliability and use less energy,” says Krishnamurthy.
Ad hoc sensor network with a high bandwidth 802.11 mesh overlay network based on Intel XScale® technology
PXA250 (Intel XScale® core) Sensor Gateway
Mark Yarvis is the Principal Investigator for Intel’s Heterogeneous Sensor Networking research project. His team is working with Krishnamurthy’s to explore various forms of heterogeneous networks and measure their impact on performance.
“We wanted to measure the impact of adding a few 802.11-enabled nodes, plugged into a power source,” says Yarvis. “Would this improve the network’s performance? Would it help the rest of the network to survive longer, by shifting the burden of forwarding packets to these nodes and away from the battery-powered motes?”
“We also wanted to explore other forms of heterogeneity,” Yarvis continues. “For example, how would performance change if we didn’t add any high-end nodes to the network but simply plugged a few of the inexpensive motes into the wall?”
Heterogeneous Sensor Network
Path across overlay network
Testing the Concept
With hypothesis in hand, the researchers set out to test the concept of using heterogeneous networks to improve performance. In one experiment, they scattered 50 motes throughout a large conference room. Onto this network, they overlaid an 802.11 mesh network comprised of four XScale nodes. Then they measured the performance of this heterogeneous network, with and without the XScale nodes plugged into wall outlets.
In evaluating performance, researchers used a program that records how often data is being received through the gateway nodes. In addition to recording data age, the program calculates the remaining lifetime of the network, based on data individual nodes transmit about the amount of energy they are consuming.
First, researchers measured network performance with the XScale nodes plugged in. “These links caused the sensor network to fold up onto itself,” says Yarvis. “Nodes that were actually far away suddenly appeared to be much closer together, from their point of view, so that’s the route they took. They didn’t know why this was a good route, just that it was good.” Krishnamurthy adds: “There’s nothing hard-coded that says a particular set of nodes must find a higher-end node to send data to. Every node tries to find a good route back to the sink node. It’s just that routing data through the high-end powered nodes improves performance.”
Next, researchers unplugged the XScale nodes and measured performance again. With the power off, the motes began routing data through a different path; the XScale nodes no longer appeared to be a superior route.
This experiment confirmed the researchers’ hypothesis that heterogeneous networks enhance overall performance. With the XScale nodes disabled, the average data age doubled, from approximately 10 seconds to 20-25 seconds, and the network lifetime
decreased by 20%.
802.11 overlay enabled
802.11 overlay disabled
This experiment, and others, confirmed another important objective of the research: to show that while 802.11 overlays improve performance, sensor networks can function without them. “We want the network to perform better with the presence of these higher-end capabilities, but we don’t want the network to become dependent on them,” says Yarvis. “We want to make the 802.11 mesh network as transparent as possible. Our research has confirmed that the network will still run, it will still self-organize after you unplug the high-end nodes. It just might not live as long, and the quality of the data might go down.”
Embedding Processing in the Network
Now that researchers have made strides in scaling sensor networks without jeopardizing performance, they are exploring the next research frontier: embedding processing in the network. With an 802.11 mesh network overlay, large amounts of data can move rapidly across a wireless sensor network. The drawback is that moving data consumes energy and other resources, including bandwidth. In some cases, there is no need for data to be continually streaming across the network. A more efficient, energy-conserving option is to embed local processing capabilities within selected network nodes.
A simple example illustrates the concept: If you wanted to measure the average temperature in a room, you could have sensors continuously transmit data, perhaps every second or two, and have a client application compute the average. A more efficient approach is to have one or more nodes in the network calculate a weighted average of temperature readings reported by individual sensors, and send that figure back to the client, reducing the amount of data that must traverse the network.
Another way to extract the results of processing done in the network is to exploit the services of TinyDB, a query processing system developed by Intel Research Berkeley**. "The ultimate goal of TinyDB is to allow people to query sensor networks without having to program them,” says Mark Yarvis. His team is working with researchers in the Berkeley lab to explore the potential for using TinyDB in heterogeneous sensor networks.
“Essentially, TinyDB imposes a database model on top of the sensor networks,” says Yarvis. “Once the model is in place, users can do sensing tasks based on simply posing queries, similar to querying a standard database.”
In addition to exploring the potential of TinyDB, researchers are investigating how to organize networks so that query processing is optimized. “What if some nodes are better suited to query processing than others are?” Yarvis asks. “Even in a network of completely homogenous hardware, some nodes may be underutilized. That makes them good places to cache data to handle complex group queries or similar tasks. We’re trying to determine how to organize the network so we can leverage those nodes.”
Enhancing the Nodes: Intel® Mote Technology
While Yarvis is exploring the potential for embedding processing in the network, Ralph Kling’s team is enhancing the motes that will do some of the processing. As Principal Investigator for the Intel Mote research project, he is leading a team to create the next generation of motes. The team is collaborating with Intel Research Berkeley, which has done pioneering work in this area and is well known for its 'Berkeley mote'.
“At the start of our research, we talked to the folks at Intel Research Berkeley, to several start-ups, and to people working in standards bodies, to learn more about the environment that motes operate in—the hardware and software they use, the applications they are targeted for, and the networking that takes place when the motes are placed in an actual application,” says Kling. “We learned that people wanted a number of enhancements to the current generation of motes.”
Based on these discussions as well as conversations with companies that plan to use motes in the future, Kling’s team developed a prototype next-generation Intel mote. This enhanced mote, which is half the size of the original Berkeley mote, provides increased CPU power, for tasks such as location detection and digital signal processing. Other enhancements include security features, more reliable radio links, and additional on-board memory. The Intel mote is a modular, stackable platform that can be customized for a range of applications.
Intel® Mote prototype (3x3cm)
With work on the first-generation Intel mote near completion, Kling’s team is looking ahead to further enhancements. One area of focus is low-power operation. “We are striving to develop ultra-low power design—at least two orders of magnitude below the operational power of traditional low-power platforms, such as XScale-based designs,” says Kling.
The team is also focusing on system level integration. “There are many technologies, such as radio components, CPU core, Flash memory, SRAM memory and the sensors themselves, that could be placed either on a chip or in the package,” says Kling. “We are exploring how to integrate these components in the best and most cost-efficient way.”
The third and final focus of the team’s current research is on hardware reconfiguration. “Since new radios, protocols and applications are emerging all the time, we cannot predict what the demands on computing power will be,” says Kling. “So we want to have the ability to reconfigure the mote hardware to make it flexible and adaptable, able to efficiently balance power and performance.”
Routing Protocols
In addition to building enhanced motes, researchers are striving to improve the underlying routing layer in order to make communication among the nodes more efficient. Communication is expensive in terms of battery power, so it’s important to have battery-powered nodes “sleep” most of the time, waking up only infrequently to communicate. “In a large network, you have to agree when nodes are going to ‘talk’ and when they are not, for two reasons,” says Kling. “You have to ensure that the nodes that want to communicate are awake at the same time. Ideally, the others will be asleep, so they won’t interfere with the conversation. That’s not a simple task, because two motes may talk to each other via intermediaries, which is the multihop concept.”
The teams led by Yarvis and Krishnamurthy are experimenting with routing protocols, focusing on the identification and utilization of heterogeneous capabilities, such as 802.11 links and TinyDB services. That research is in the early stages, but the researchers are making progress. “We currently have an implementation of TinyDB that works on top of an implementation of our routing protocol,” says Yarvis. “We’re poised to start running experiments using the two together, to understand in what cases this structure is beneficial and in what cases it’s not.”
Translating Research Into Real-World Applications
With the EcoSense project underway for two years now, Krishnamurthy’s team is focusing on building applications that deliver measurable business value. “We want to take all of the research we’ve done over the last two years and apply it to the real world,” says Krishnamurthy. The team is exploring two applications: preventive maintenance for equipment in Intel’s fabs, and sensor networks for theme parks. Both applications leverage the concept of heterogeneous sensor networks, and both solve important business problems in their domains.
“A key goal in exploring these applications is to evaluate the return on investment that a company could expect on the deployment of a sensor network,” says Krishnamurthy. He believes that these applications represent the next step in the evolution of sensor network applications, which thus far have focused on solving researchers’ problems rather than meeting business needs.
Preventive Maintenance
At Intel’s semiconductor fabs, thousands of sensors track vibrations coming from various pieces of equipment to determine if the machines are about to fail. “There’s an established science that enables us to determine the particular signature that a well-functioning machine should have,” says Krishnamurthy.
Currently, employees in the fab must manually gather the sensor data from each node—a costly and time-consuming process that is carried out periodically, on a schedule determined by the expected failure rate of the equipment. Networking the sensors would make the process far more efficient and cost-effective.
“Creating a wired network of these sensors would be very expensive,” Krishnamurthy notes. “Why not make use of the mote technology to build an application that acquires the data automatically? This would save a great deal of time and money, and it would allow you to acquire data more frequently.”
Krishnamurthy’s team is collaborating with vendors to develop such a preventive maintenance application. Prototypes have already been built, and the application is expected to be ready for testing within Intel fabs by the end of 2003.
Intel Fab Network Architecture
Wireless Theme Parks
Krishnamurthy’s team is also exploring the deployment of heterogeneous sensor networks in theme parks. Such networks could be used for multiple purposes. One potential use is monitoring the quality of water in tanks. Today such monitoring is done manually. A sensor network could make the process more accurate and efficient.
Another potential use of the network is to provide Internet access to park visitors. “Visitors could use the wireless network to reserve a space at a particular park attraction, or to learn more about an exhibit,” says Krishnamurthy. “If you’re at the park with your kids all day, you could use the Internet to stay in touch with friends and the office as well. And the kids could download and store information for a school paper. In general, it would provide a way to stay connected.”
The wireless network could improve park management as well. Sensors could track attendance at park exhibits and rides, and management could use the network to access office applications from various stations throughout the park
Intel is currently exploring this application with a major theme park. The potential return on investment will depend in part on the benefits visitors perceive. “If park visitors are willing to pay more for the additional services, the increased revenue might cover the cost of the infrastructure,” says Krishnamurthy. “If the infrastructure pays for itself, management could gather a variety of additional information at no extra cost. This would improve park management, which would benefits visitors as well.”
Theme Park Network Architecture
Reusing core 802.11 mesh
Preventive maintenance and wireless theme parks are the only applications Krishnamurthy’s group is pursuing for now, but the potential for others is enormous. “Anywhere there’s a need to connect large numbers of sensors, the concept of marrying sensor networks and 802.11 mesh networks can play a key role,” says Krishnamurthy. “We view our research as an enabling activity that will help vendors to build many useful real-world applications that leverage heterogeneous sensor nets. As we move forward, we will try to identify more applications that will make good business sense and deliver a solid return on investment. That’s what our research is driving towards.”
