An Interactive,Source-Centric,Open Testbed for Developing and Profiling Wireless Sensor Systems

Abstract

The difficulty of developing wireless sensor systems is widely recognized. Problems associated with testing, debugging, and profiling are key contributing factors. While network simulators have proven useful, they are unable to capture the subtleties of underlying hardware, nor the dynamics of wireless signal propagation and interference; and physical experimentation remains a necessity. To this end, developers increasingly rely on shared deployments exposed for physical experimentation. Sensor network testbeds are under development across the world.

We present a complementary testbed architecture that derives its novelty from three characteristics. First, the system is interactive; users can profile source and network level components across a network in real time, as well as inject transient state faults and external network traffic. Second, the system is source-centric; it enables automated source code analysis, instrumentation, and compilation. Finally, the design is open; developers can extend the set of exposed inter faces as appropriate to particular projects without modifying the underlying middleware. We present the testbed design and implementation, a graphical user interface, a shell-based macro programming interface, example scenarios that illustrate their use, and discuss the testbed's application in the research and teaching activities at client institutions.

Keywords

Sensor Networks Testbeds Testing Experimentation Instrumentation Debugging Profiling

1. Introduction

Wireless sensor networks are key enablers of the ubiquitous computing vision. Deployments support ecological monitoring [66, 43, 31, 37], structural damage detection [8, 27], wildfire prediction and tracking [29, 17], disaster response [36], intruder detection and classification [4, 1], and myriad other scenarios. The lowest tiers of these networks are composed of “motes” [63], tiny computing devices that sense, process, and communicate environmental stimuli. With form factors that range from matchbox [12] to quarter-size [13], motes can be unobtrusively deployed at ultra-dense scales. Deployments have already surpassed the 1,000 node threshold [4].

Despite the increase in application opportunities, sensor systems remain difficult to develop. Applications typically include network-specific data aggregation protocols, signal processing algorithms, and power management services; large-scale parallelism and distribution characterize the common case. The resulting complexity is at tension with the limited resource capacity of the hosting devices, the gross unpredictability of the underlying channels, and the limited tools available to support system debugging and profiling. One consequence is the widely recognized difficulty of achieving predictable performance, and the consequent necessity of rigorous experimentation and evaluation.

Network simulators such as ns-2 [39], and more specialized sensor network simulators such as TOSSIM [35] and Avrora [55], have helped. Developers are able to prototype and simulate their systems to gain initial measures of correctness and performance. The tools have not, however, eliminated the need for physical experimentation. Existing simulators are unable to capture the complex dynamics of wireless signal propagation and interference [71, 54], nor are they able to capture hardware subtleties that aspect performance, such as preemption priority, clock rate variation, and load-induced behavioral anomalies. As a result, empirical evidence gathered through simulation often fails to translate to real deployments.

Yet physical experimentation presents its own challenges. Wireless network reprogramming is possible (e.g., using Deluge [32]), but the required programming time is significantly greater than if the constituent devices were programmed directly (but in parallel). More important is the void in infrastructure support for runtime debugging, profiling, and experimentation. Motes are “headless”; debugging and profiling messages are logged to a base station, or stored in EEPROM for later analysis. The requisite instrumentation logic must be developed and integrated a priori. In effect, the “interesting” portions of an application—those portions that might conceal defects—must be identified and instrumented in advance. As a result, developers do not benefit from the rapid maintenance, deployment, and debugging cycles used to improve programmer productivity and system reliability in other development domains.

Contributions

To address these limitations, we present the design and implementation of a wireless sensor network testbed tailored to support system debugging, profiling¹, and experimentation. The supporting middleware platform exposes multiple physical deployments, shared in effect, as virtual devices accessible to a distributed research community. The installation reported here includes 80 devices, expandable to an arbitrary size. The principal contribution of the architecture lies in the design of the supporting middleware platform, and the synergistic combination of features it provides. The platform is engineered to be interactive, source-centric, and open.

By interactive, we refer to a design that enables users to profile source- and network-level components in real-time, as well as inject network packets and state modifications. This improves the ability of developers to evaluate system performance, localize defects, and observe behaviors in the presence of anomalous network conditions and transient faults. By source-centric, we refer to a design that targets application source materials, as opposed to application images; source-centered features include automated analysis, instrumentation, and compilation services. These services improve programmer productivity by eliminating the need for manual integration of testbed management components. Equally important, the services increase the level of implementation detail available to software testers, and provide a foundation for software configuration testing. A range of application configurations can be evaluated without the expense of developing the individual variants.

Finally, by open, we refer to a design that enables developers to extend the set of interfaces used to access the testbed without modifying or even restarting the underlying middleware. Developers may choose to use the default graphical user interface, the default shell scripting interface, or a custom interface appropriate to a particular system or experimentation task.

We refer to the testbed architecture as the “NESTbed,” for “Network Embedded Sensor Testbed.” While much of the design is platform independent, the current implementation targets applications developed using nesC [24] and TinyOS [30, 59]. We take this platform to represent the state-of-the-art in sensor network platforms, and the evolving de facto standard.

Novelty

A number of testbeds have been discussed in the literature, including architectures tailored for 802.11 networks [72, 16, 34, 47, 44, 68, 7], and more recently, wireless sensor networks [21, 61, 28, 67, 33, 64]. While a detailed discussion of related work appears in Section 6, it is useful at this point, given the number of related systems, to preview the points of novelty in our work. In short, the novelty lies in the three distinguishing characteristics of the supporting middleware: interactive, source-centric, and open.

Interactive

Existing sensor testbeds are largely batch-based; they provide few features for real-time debugging and experimentation. They do not, for instance, provide integrated support for real-time network and source-level profiling (without a priori knowledge of the elements to be profiled), nor do they support source-level fault injection. This lack of execution accessibility reduces developer productivity, and limits possible evaluation scenarios, including those designed for assessing state related fault tolerance properties.

Source-centric

Existing sensor testbeds are image-based; developers are required to upload precompiled applications that include requisite testbed management components integrated by the developer. Consequently, existing designs preclude automated source-level analyses, instrumentation services, and build processes. This imposes a burden on developers responsible for testbed specific component integration, reduces the level of implementation detail exposed by the testbed, and limits the efficacy of software configuration testing to evaluate the program variants.

Open

Existing sensor testbeds are principally realized as web based systems that expose their services through standard forms pages. This approach prevents client institutions from extending testbed functionality without modifying the shared server installation. Remote developers are unable, for instance, to construct specialized testbed driver components, or interfaces tailored to particular experimentation tasks or applications.

The point of departure for our work is the observation that these characteristics are fundamental to a testbed engineered to support rapid debugging, profiling, and experimentation, the focus of the NESTbed system.

Paper Organization

The remainder of the paper is organized as follows. Section 2 presents an overview of the NESTbed system architecture, including the key hardware and software components underlying its design. Section 3 presents use-case scenarios that highlight key features and benefits of the system. The scenarios are presented in the context of the default graphical interface. The shell-based scripting interface is presented in Section 4, along with a representative use-case scenario. Section 5 discusses current applications of the system in the research and teaching activities at client institutions. Section 6 compares the NESTbed system to existing testbeds discussed in the literature. Finally, Section 7 concludes with a summary of contributions and pointers to future work.

2. System Architecture

The NESTbed system architecture is illustrated in Fig. 1. The architecture is composed of three layers:

(i)
physical network deployments,
(ii)
a centralized application and database server, and
(iii)
client interfaces for remote users, who may optionally connect one or more remote sensor subnets.

Figure 1.
NESTbed system architecture.

We briefly describe each of the architectural layers in the paragraphs that follow.

The system supports multiple physical deployments. Given the design goal of enabling interactive use, each deployment is dedicated to a single user at a given time. Multiple users may, however, access different deployments concurrently. Our prototype configuration includes one physical deployment consisting of 80 Tmote Sky [42] devices arranged in a dense grid. Small web cameras mounted overhead provide streaming video feeds that show the actuation state of the network. While the feeds are not strictly necessary for gathering this state, they appear to have an important psychological benefit. Users note that the video feeds provide an improved sense of presence; they support the view of the testbed as a locally attached virtual device.

The prototype deployment is shown in Fig. 2. Each mote is attached to the server through a USB connection. The grid measures 4ʹ × 8ʹ. Although our current facility is not large enough to house such a deployment, the addition of wireless USB extenders would enable connections in excess of 150ʹ, significantly increasing the potential deployment scale, especially if the connections were chained. No additional hardware or software changes would be required. It is natural to question whether the geographic scale of the deployment admits of interesting (and realistic) wireless topologies. We will return to this question with an affirmative response in Section 3.

Figure 2.
NESTbed prototype deployment.

Each Tmote Sky includes a 16-bit microcontroller clocked at 8Mhz, a 2.4GHz ZigBee radio, 48K of ROM, 10K of RAM, and 1M of off-chip EEPROM storage. Each device additionally includes integrated temperature, light, and humidity sensors, and can be configured to support a range of additional sensors. Each mote exposes a unique hardware identifier that enables the software running on the server to associate a physical grid position with the device, independent of how the operating system assigns port addresses. This allows the server to preserve mote addressability across reboots. We note that the total equipment cost for the prototype installation, including the application and database server, is less than Ȥ10,000, making the deployment economically feasible to replicate at other institutions.

The back-end server hosts a suite of APIs that enable remote clients to work with NESTbed projects on one or more network deployments. Each API is implemented as a collection of Java RMI objects² [52] referred to as “managers.” The RMI based design simplifies the construction of remote client interfaces, and exposes the testbed for programmatic control. We will consider two client interfaces, designed and implemented independently, in Section 3. As illustrated in Fig. 1, the API suite consists of six core components: (i)
the Configuration API,
(ii)
the Instrumentation and Compilation API,
(iii)
the Deployment API,
(vi)
the Profiling API,
(v)
the Power Control API, and
(iv)
the Gateway Control API.

The manager objects within each API, and the key resource dependencies among them, are illustrated in Fig. 3. The services provided by each API are summarized below.

Figure 3.
NESTbed server software architecture.

Configuration API

The Configuration API consists of managers that support the construction and maintenance of NESTbed projects and deployment configurations. A project includes a collection of nesC source files, meta-data about the files (e.g., program symbols, message structures), and a corresponding set of deployment configurations. Each deployment configuration describes a project installation, including the mappings between application images and physical devices, network- and source-level profiling to be performed, and radio power settings for each device. Projects and configurations are maintained in a persistent store to maintain experimental controls across runs³. Data adapter objects abstract the underlying storage technology (e.g., MySQL, PostgreSQL) to enable pluggable storage implementations.

Instrumentation and Compilation API

The Instrumentation and Compilation API consists of managers that support static analysis, instrumentation, and compilation of source files within a project. The analysis services include parsing functions, program symbol identification, and message structure identification. The analysis results are maintained in persistent storage to improve system response time across experiment configurations. The instrumentation services include support for integrating NESTbed management components required by the server library, and integrating alternative system and application level nesC components (e.g., alternative radio stack implementations). Component alternatives may be selected from the NESTbed library, or provided as part of a user project. The managers rely on source-level weavers to perform these functions.

Deployment API

The Deployment API is implemented by a manager that provides services for programming and configuring the shared network deployments based on a specified project and deployment configuration. The services include both the whole network and the individual mote programming functions, as well as error detection and reporting features. In the event of a device programming error, the client is notified of the failed installation, and may optionally choose to reprogram the device.

Profiling API

The Profiling API consists of managers that provide source and network-level profiling functions. The source-level functions enable remote clients to dynamically read and write program variables associated with the application image executing on a device. The network-level functions enable clients to subscribe to one or more message streams corresponding to the network messages received by a specified network subset. Recall that the program symbols and message structures associated with each application are maintained in persistent storage. This information may be queried by remote users to assist in configuring the activities performed by the Profiling Manager.

Power Control API

The Power Control API is implemented as a manager that provides services for toggling power to specified devices. The services are implemented using USB power control functions included as part of the USB 2.0 [11] standard. The API services enable the injection of transient and persistent node failures to support fault-tolerance experimentation. The services additionally support device recovery when nodes enter unresponsive states.

Gateway Control API

The Gateway Control API consists of managers that enable remote users to create and destroy network gateways. This allows remote clients to extend static NESTbed deployments with remote networks, system controllers, and applications. Developers can, for instance, inject live sensor data into a NESTbed deployment from an outdoor field experiment. The API manages a set of TinyOS SerialForwarder instances⁴ that serve as mote-to-TCP bridges, one for each device. Messages received by a mote over its wireless radio are forwarded through its USB connection, and retransmitted at an advertized port by the server. Messages received at a port are forwarded to the corresponding device through its USB connection, and retransmitted over its wireless radio.

The NESTbed system includes two default user interfaces. The first is a graphical user interface designed to enable “user-friendly” NESTbed access. The second is a shell-based interface that provides scripting services for complex experimentation tasks, or tasks involving a large degree of repetition. In the next sections, each interface is used to illustrate the testbed functionality summarized here. We emphasize, however, that the testbed services are exposed by the NESTbed server. Other researchers may choose to access this functionality through a custom interface. In the case of a deployment intended for continuous use, for example, researchers might choose to develop a scheduling system on top of the exposed services.
3. System Use-Cases

We now turn our attention to a series of use-cases that illustrate the features and benefits of the NESTbed system. We begin with a scenario involving the development of a multi-hop sensing application. The scenario assumes the use of the NESTbed graphical interface.

3.1. The NESTbed Manager

The interface is Java based, and supports “one-click” web deployment using Java Web Start. The initial NESTbed Manager window is shown in Fig. 4. The first display segment lists the physical network deployments available for use. In the prototype installation, the “Ultra-dense Network” is the only available deployment. This information is configured statically by the system administrator. The second and third segments display the associated NESTbed projects and deployment configurations populated by system users. Project and configuration management functions (e.g., for adding projects, cloning configurations) are realized by the Configuration API exposed by the server.

Figure 4.

NESTbed manager.

The scenario begins with the selection of a physical network deployment, and the creation of an associated NESTbed project and deployment configuration. As shown in the figure, the user has created a project named “Multi-Hop Sensing Application” associated with the “Ultra-dense Network,” and an empty configuration within the project, named “Default Configuration.” The Deployment Configuration Manager is then used to configure the system installation and associated profiling settings for a particular debugging or experimentation task.

3.2. The Deployment Configuration Manager

The Deployment Configuration Manager is shown in Fig. 5. The left panel displays the nesC programs associated with the active configuration, and for each program, a list of the constituent nesC modules, module variables, and messaging structures. The panel is populated automatically as programs are uploaded by the user. The right panel displays the physical topology of the mote network; the information is configured statically by the system administrator. The display for each mote includes the associated network identifier, an indication of the application image to be installed, and the radio power level that should be set upon installation. The bottom panel displays a list of module variables and messaging structures selected for profiling. As we will see, these selections control the available profiling actions when the deployment configuration is used to program the network. The NESTbed Instrumentation and Compilation API is used to generate the displayed program data.

Figure 5.

Deployment configuration manager.

The user scenario continues with the selection of the application source directories to be archived and uploaded to the NESTbed server. After making this selection, the user is presented with the Component Rewiring dialog shown in Fig. 6. The dialog enables the selection of alternative operating system and application level component implementations⁵. A user might, for example, upload a single source directory multiple times, selecting alternative radio stack and network routing implementations in each case⁶. This would eliminate the development effort normally required to construct the individual program variants. In the scenario captured in Fig. 5, the user has uploaded the MultiHopSensing application, and chosen to use the ReliableComm component in place of the default radio stack (GenericComm). When the application was uploaded, the files were automatically parsed, instrumented, and compiled for use. The status window shown in Fig. 7 was displayed during this process. Upon successful completion, the left panel was updated with the name of the uploaded program, and its associated program symbols. MultiHopSensing defines three messaging structures and a range of program modules. The SensingM module, for instance, includes two program variables, msg and pending, as shown in the figure.

Figure 6.

Component rewiring dialog.

Figure 7.

Compilation status window.

The next step is to configure the applications to be installed on the network. This involves dragging programs from the left panel to the node(s) on which the programs should be installed. Unconfigured devices will be disabled when the configuration is activated on the network. In the ongoing scenario captured in Fig. 5, every second device has been configured; unconfigured devices are identified by hashmarks. The value shown in the bottom right corner of each mote icon indicates the radio power level to be set when the device is programmed. In this scenario, the radio power level has been reduced across the network to account for the density of the deployment. The reduction limits the effective range of each device, creating opportunities for more interesting network topologies. The goal is to generate topologies more consistent with a geographically distributed deployment. We will return to this idea later in the section.

The final configuration step involves selecting the runtime profiling information to be made available when the configuration is activated. The user can select two types of elements from the left panel. Module variables selected for profiling can be inspected and modified during program execution. Messaging structures corresponding to the types of packets transmitted over the radio and/or USB port can also be selected. This enables the user to inspect the contents of messages transmitted via USB during program execution. To enable inspection of the wireless network, the user can include simple USB forwarding logic within their programs. Alternatively, they may choose to install a radio-to-USB forwarding application on one or more of the unused motes. Several such applications are included as part of the TinyOS distribution (e.g., TOSBase) [59].

In the scenario shown in Fig. 5, the user has selected two variables for profiling, each declared within the RoutingM module. RoutingM implements a variation of the TinyOS Beaconing Protocol [30] to maintain a shortest-path spanning tree rooted at node 0. The tree is used to route sensor data to the root node, which in turn forwards the data through its USB port for upper tier processing. The first selected variable, distance, stores the hosting node's distance from the root, measured in hops. The second, parent, stores the identifier of the node's parent in the tree. Although not visible in the figure, the user has also selected the UartMsg structure, which corresponds to the messages received and forwarded by the root node. The user is now ready to activate the configuration using the NESTbed Network Monitor.

3.3. The Network Monitor

The interface of the Network Monitor is similar to the Deployment Configuration Manager; we omit an additional screenshot. The interface enables users to install, debug, and profile applications based on the active deployment configuration. Single device and whole network programming are supported. In the latter case, motes are programmed in parallel to reduce installation time. The former option, used less frequently, supports scenarios that require a particular installation order. For example, when evaluating fault tolerance characteristics, it may be useful to introduce corrupted nodes gradually.

Visual feedback is provided during installation to indicate success or failure. Nodes are shown within a flashing green box during installation; solid green and solid red indicate success and failure, respectively. Installation failures are often remedied by reprogramming the failed devices. It is possible, however, for a mote to enter a hardware state in which the device cannot be reprogrammed. In such a case, the device can be power-cycled to re-enable programmability. The Deployment API is used to implement the installation and status reporting features exposed through the client interface. The Power Control API is used to implement the power-cycling feature.

Upon successful installation, users may view video feeds of the network to inspect its actuation state. Sample images are shown in Figs. 8 and 9; each captures approximately one-half of the network (with some overlap). The actuation state consists of the LED states of the individual devices — a useful debugging tool for signaling phase transitions, error conditions, and other significant events. In general, the actuation state may also include the states of external devices under network control (e.g., lights, physical switches, motors). The goal of the feeds is to provide convenient real-time access to this information. Again, the feeds are not strictly necessary; the actuation state can be inspected through corresponding state variables. Users have noted, however, that the video feeds provide a sense of presence. They support the abstraction of the testbed as a virtual device, and provide users with additional confidence in their perception of the network state.

Figure 8.

NESTbed video feed #1.

Figure 9.

NESTbed video feed #2.

The user scenario continues after the installation of the active configuration. Each configured mote is executing an instance of MultiHopSensing. The user is now interested in determining whether a stable spanning tree has formed. In addition to the video feeds, the Network Monitor supports real-time debugging and profiling of individual nodes; both source and network-level inspection is supported. After selecting the device of interest, the user is presented with the Mote Detail window shown in Fig. 10.

Figure 10.

Mote Detail.

The window summarizes information about the selected device, including its network identifier, physical location, and hardware characteristics. The bottom panel shows the variables and message structures previously selected for profiling. When a variable is highlighted, the system retrieves its current value and updates the display. In the figure, the user has queried the values of parent and distance. Since a child of node 0 is one hop from the root, the values appear correct for the selected device. The user may similarly choose to update a program variable. The user might choose, for instance, to inject a transient state fault to force the selection of a new parent, or set an invalid distance from the root to determine whether the system can recover from state corruption. The static information displayed within the Mote Detail window is retrieved using the Configuration API. The variable profiling features are implemented using the Profiling API.

The NESTbed server relies on the Nucleus toolkit [56] to read and write program variables. The toolkit provides a lightweight query system for TinyOS that exposes nesC variables as attributes. Exposed attributes can be read and written at runtime using the Nucleus Java library. To enable these features, the NESTbed compilation system includes Nucleus [56] support in every uploaded application.

The user scenario continues with the task of determining whether the appropriate data packets are being received and forwarded by the root node. To achieve this, the user selects the type of message to be intercepted from the Message Profiling tab shown in Fig. 10. In this case, UartMsg is selected, and the Message Monitoring window shown in Fig. 11 is displayed. The window is generated dynamically based on the fields contained within the selected structure. In this case, the window includes a field to identify the source mote, and a variety of sensor readings defined within UartMsg. The fields are updated in real-time based on the messages transmitted over the root node's USB port. Received messages may also be logged to the client's local machine for later analysis. Multiple logging sessions associated with different devices may be active simultaneously.

Figure 11.

Message monitoring.

It is useful to consider the implementation of this feature. When an application is uploaded to the NESTbed server, its messaging structures are identified by the Instrumentation and Compilation API. A corresponding Java class is generated for each structure⁷. These message classes provide methods for parsing raw packet data and populating class fields. When a structure is selected for monitoring, the NESTbed server is notified via the Profiling API. The server in turn begins to inspect data received through the USB port of the relevant device. When a message of the appropriate type is received, the server constructs an instance of the corresponding Java data class, and transmits the object to the client application. When received, the Java Reflection API [51] is used to inspect the object, and to create a Message Monitoring window of the type shown in Fig. 11. Logged messages are recorded in Java's serialized object format, simplifying the construction of external analysis tools.

3.4. Topology Control

The spatial scale of the NESTbed deployment raises questions concerning its use in evaluating applications intended for geographically distributed environments. Office space is a factor; the current deployment measures approximately 4ʹ × 8ʹ. If space were available, however, the existing hardware components could be spaced to create a deployment in excess of 90ʹ × 30ʹ with no additional purchases or revisions. USB extenders could be used to distribute the nodes even further, potentially in excess of several hundred feet in both dimensions.

Still, the achievable scale is not without limits; there are target environments that outstrip the spatial capacity of any existing testbed. To address this limitation, the NESTbed system enables users to control the radio power level of each device. The key observation is that network link quality varies predictably as a function of transmission power and distance. When a user desires a deployment environment beyond the spatial capacity of the physical network, radio power can be reduced to achieve link quality consistent with the desired distribution. The tradeoff between distance and power, and its use in emulating target environments is also noted in [21].

To enable users to select the power level appropriate for a desired emulation scenario, we have conducted extensive field experiments to characterize the precise relationship between link quality, transmission distance, and transmission power for the Tmote Sky platform. The complete results, included as a reference sheet in the NESTbed distribution, are presented in [62]. A representative sample is shown in Table 1. Each row corresponds to a transmission distance; each column corresponds to a radio power level on the Tmote Sky. The value of each cell denotes the expected base (interference-free) packet reception rate (PRR) at the corresponding distance and power combination. To emulate an environment, the user can adjust the radio power level to achieve a PRR consistent with the desired deployment scenario. As we will see in Section 5, even in a confined deployment, this approach results in interesting wireless topologies.

It is useful to note that some scenarios can not be faithfully emulated using this approach; the desired packet reception rate may be too low to achieve in a dense deployment, even at the lowest radio power level. In such a case, a user may substitute UniformLossyComm, included as part of the NESTbed system, in place of the default radio stack on one or more devices. (See the Component Rewiring dialog shown in Fig. 6.) The alternative radio stack discards packets with a specified uniform probability, allowing users to emulate low-quality links consistent with large-scale spatial distribution.

Table 1
Topology Control Reference: Impact of Distance and Power on PRR

Distance (ft.) Power (CC2420)

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31

1 100 100 100 99 99 100 100 99 100 99 100 100 89 100 100 99

3 0 100 100 100 99 100 99 100 99 100 100 99 100 100 100 99

5 0 100 100 100 100 100 100 100 100 100 100 100 99 100 100 100

7 0 100 100 100 100 94 99 100 99 100 100 95 99 94 100 100

9 0 100 99 100 99 99 99 100 99 100 100 100 99 99 100 99

11 0 87 100 99 99 99 100 99 100 100 100 100 100 100 100 100

13 0 36 100 88 100 100 99 93 100 100 99 99 99 100 100 100

15 0 0 100 100 100 100 100 89 99 100 100 100 100 99 99 100

17 0 0 92 100 100 100 100 91 100 100 100 100 100 100 100 99

19 0 0 36 99 100 100 93 99 99 100 100 100 100 100 100 100

21 0 0 35 99 100 100 100 100 100 99 100 100 100 100 99 99

23 0 0 0 22 98 99 89 100 99 100 100 99 99 97 100 99

25 0 0 2 95 100 99 100 100 100 99 100 100 99 100 100 86

27 0 0 0 0 23 87 98 99 99 100 100 100 100 100 99 100

29 0 0 0 0 43 89 99 99 99 99 100 100 100 100 84 99

31 0 0 0 0 0 0 0 4 38 39 54 83 96 89 86 93

33 0 0 0 0 0 7 77 79 79 97 98 99 99 95 99 99

35 0 0 0 0 0 2 28 84 94 100 100 100 99 100 100 100

37 0 0 0 0 0 0 2 21 49 54 76 93 97 91 80 76

39 0 0 0 0 0 12 56 88 93 99 98 99 99 99 99 98

41 0 0 0 0 0 0 3 23 38 68 80 93 81 91 91 99

43 na 0 0 0 0 0 15 65 80 91 94 99 98 99 100 98

45 0 0 0 0 0 0 35 61 66 99 99 100 99 100 94 95

47 0 0 0 0 0 0 1 6 61 88 99 99 99 99 91 96

49 na 0 na 0 0 0 1 3 1 1 3 14 40 45 na 79

51 0 0 0 0 0 0 0 9 2 1 75 36 90 91 61 93

53 0 0 0 0 0 1 39 84 98 99 99 100 99 100 99 100

55 0 0 0 0 0 0 0 0 21 60 77 96 99 99 99 98

57 0 0 0 0 0 0 0 2 11 22 10 11 53 54 56 52

59 0 0 0 0 0 0 0 1 4 12 na 39 28 35 20 10

61 0 0 0 0 0 0 2 0 na 15 27 5 na na na na

63 0 0 0 0 0 0 8 38 54 49 56 84 96 79 98 99

65 0 0 0 0 0 0 60 94 98 89 99 99 99 99 99 99

Distance (ft.)	Power (CC2420)
1	100	100	100	99	99	100	100	99	100	99	100	100	89	100	100	99
3	0	100	100	100	99	100	99	100	99	100	100	99	100	100	100	99
5	0	100	100	100	100	100	100	100	100	100	100	100	99	100	100	100
7	0	100	100	100	100	94	99	100	99	100	100	95	99	94	100	100
9	0	100	99	100	99	99	99	100	99	100	100	100	99	99	100	99
11	0	87	100	99	99	99	100	99	100	100	100	100	100	100	100	100
13	0	36	100	88	100	100	99	93	100	100	99	99	99	100	100	100
15	0	0	100	100	100	100	100	89	99	100	100	100	100	99	99	100
17	0	0	92	100	100	100	100	91	100	100	100	100	100	100	100	99
19	0	0	36	99	100	100	93	99	99	100	100	100	100	100	100	100
21	0	0	35	99	100	100	100	100	100	99	100	100	100	100	99	99
23	0	0	0	22	98	99	89	100	99	100	100	99	99	97	100	99
25	0	0	2	95	100	99	100	100	100	99	100	100	99	100	100	86
27	0	0	0	0	23	87	98	99	99	100	100	100	100	100	99	100
29	0	0	0	0	43	89	99	99	99	99	100	100	100	100	84	99
31	0	0	0	0	0	0	0	4	38	39	54	83	96	89	86	93
33	0	0	0	0	0	7	77	79	79	97	98	99	99	95	99	99
35	0	0	0	0	0	2	28	84	94	100	100	100	99	100	100	100
37	0	0	0	0	0	0	2	21	49	54	76	93	97	91	80	76
39	0	0	0	0	0	12	56	88	93	99	98	99	99	99	99	98
41	0	0	0	0	0	0	3	23	38	68	80	93	81	91	91	99
43	na	0	0	0	0	0	15	65	80	91	94	99	98	99	100	98
45	0	0	0	0	0	0	35	61	66	99	99	100	99	100	94	95
47	0	0	0	0	0	0	1	6	61	88	99	99	99	99	91	96
49	na	0	na	0	0	0	1	3	1	1	3	14	40	45	na	79
51	0	0	0	0	0	0	0	9	2	1	75	36	90	91	61	93
53	0	0	0	0	0	1	39	84	98	99	99	100	99	100	99	100
55	0	0	0	0	0	0	0	0	21	60	77	96	99	99	99	98
57	0	0	0	0	0	0	0	2	11	22	10	11	53	54	56	52
59	0	0	0	0	0	0	0	1	4	12	na	39	28	35	20	10
61	0	0	0	0	0	0	2	0	na	15	27	5	na	na	na	na
63	0	0	0	0	0	0	8	38	54	49	56	84	96	79	98	99
65	0	0	0	0	0	0	60	94	98	89	99	99	99	99	99	99

3.5 Remote Extensions

For some experimentation tasks, a fixed indoor deployment may be insufficient; users may wish to add remote subnets, system controllers, and applications. In testing a hierarchical system, for example, it may be useful to attach a tier of computationally rich sensor nodes that interact with a lower tier NESTbed deployment. A user might also wish to connect a remote field deployment to inject live (or pre-recorded) sensor data into a NESTbed experiment, as in [21]. External applications that analyze network performance or sensor stream data may also be required.

To enable these extensions, we adopt a variation of the serial forwarding approach described in [67]. When a user wishes to connect a remote service, a network gateway is associated with one or more NESTbed devices through the NESTbed Network Monitor. Each mote is assigned an advertized TCP port; the gateway creation function establishes a mote to TCP bridge on this port. More precisely, the Serial Forwarder Control API constructs a set of SerialForwarder instances on the NESTbed server. Each instance relays messages received through the USB port of its associated device to the corresponding TCP port, and vice versa. Remote applications connect to these ports to interact with NESTbed devices. To attach a remote subnet, a client side SerialForwarder instance is also required; a TCP-to-TCP relay forwards packets between client and server side SerialForwarder instances. Several such relay applications are freely available.

3.6 Experimental Repeatability

The NESTbed system enables users to quickly save, modify, and clone deployment configurations. These features assist in improving experimental repeatability and control. Users can quickly retrieve and redeploy previous experiments without any risk of modifying non-environmental experimental parameters. This is especially useful in trying to replicate experimental results among NESTbed users. The ability to clone projects and configurations is also useful, improving control of variation across experiments. A user can clone a deployment configuration, make a single change, and quickly redeploy the new experiment without any risk of modifying other non-environmental parameters.

4. NESTShell Interface

The graphical user interface can be cumbersome for complex multiphase experimentation tasks, and tasks involving a high degree of repetition. The open design of the NESTbed system allows end-users to provide supplementary interfaces to address these scenarios. One such interface is NESTShell, a shell-based scripting interface. In addition to supporting automation of complex and repetitive tasks, the scripting language provides constructs for interacting with client side tools, enabling users to extend the interface as appropriate to particular scenarios. Here we summarize the key features of the NESTShell interface; a complete description is presented in [15].

The interface parallels that of a typical operating system shell, providing a file system abstraction that models the hierarchical structure of network deployments, projects, deployment configurations, programs, and profiling data. Each directory defines a working context and a set of UNIX style commands. For example, when the working directory is the message profiling directory for a particular device, the shell provides commands for recording message streams from the device. Directories may also include files—read using the standard UNIX cat command—that provide information about the working context. A given program directory, for instance, includes a file that specifies information about the symbols contained within the program.

The structure of the NESTShell file system is shown in Fig. 12. The root of the file system contains subdirectories corresponding to the available physical deployments, which in turn include project subdirectories created and managed by end-users. Project directories include deployment configuration subdirectories, user managed through five further subdirectories:

(i)
Programs,
(ii)
SymbolProfiling,
(iii)
MessageProfiling,
(iv)
Motes, and
(v)
NetworkMonitor.

Figure 12.
NESTShell file system.

We describe each in turn.

Programs

The Programs directory contains subdirectories corresponding to the applications uploaded by end-users, and provides program upload and removal commands. When a program is uploaded, the new corresponding directory is populated with subdirectories that store information about the program symbols and message structures defined by the uploaded application (e.g., names, types). The working context provides commands for selecting these elements for profiling.

SymbolProfiling

, MessageProfiling. The SymbolProfiling and MessageProfiling directories contain files that represent the symbols and message structures selected for profiling. The working context includes commands to deselect profiling elements.

Motes

The Motes directory provides a context for configuring the images to install on each device when the configuration is deployed. Files within the directory correspond to network nodes, and specify hardware characteristics (e.g., platform type, RAM size, ROM size) and configuration status information. The latter includes the name of the application to install on the device, and the radio power level to be set when the application is deployed. The working context provides commands for configuring nodes, as well as an extended ls command that provides network topology and configuration status information.

NetworkMonitor

The NetworkMonitor directory includes a subdirectory for each mote, and provides a context for controlling the network deployment. The context provides commands for installing application images (and setting radio power levels) based on the mappings defined in the Motes directory, toggling power to individual nodes, performing soft resets, and creating and destroying network gateways. Each mote directory includes a file detailing the characteristics of the device, including its activity state. Possible states include unknown, installing, programmed, failed, and gateway. The ls command is extended to include this state information.

Two profiling directories, ProfilingSymbols and ProfilingMessages, are contained within each mote directory. The ProfilingSymbols directory contains files corresponding to the program symbols selected for profiling. The content of each file corresponds to the most recent recorded value for the symbol. The working context provides commands for reading and writing to these files, with consequent effects on the associated nodes. The ProfilingMessages directory is analogous; it provides commands for accessing log files corresponding to message streams. Each log entry includes a line separated list of the values contained within the message fields using a simple field = value format.

The shell interface additionally supports environment variables and standard UNIX style commands for setting (set), removing (unset), and enumerating (env) them. Variable values are accessed using the familiar “$_{name_}” syntax. The special variable status stores the exit status of the last executed command. This is useful for introducing recovery actions in experimentation scripts.

Finally, the interface provides conditional evaluation and iteration primitives. The iferror construct conditionally executes a block of code in the event that the status variable is non-zero. The foreach construct enables iteration over a list of items, where a user-defined variable takes on the next value from the list during each iteration.
4.1. Example Script

To illustrate the use of the NESTShell interface and the experimentation scenarios it enables, we consider a simple example. Listing 1 includes a portion of a script used to collect profiling information from a modified version of SurgeTelos, a spanning-tree-based, multi-hop sensing application included as part of the TinyOS distribution. The script begins by defining variables MOTES and LEVELS, to control the target nodes and power levels of interest, respectively (lines 1–2). Next, a deployment configuration is created for each power level, and the SurgeTelos source materials are uploaded to the server (elided).

Next is the main loop over each power level of interest (lines 6–33). Seven variables are selected for profiling in each deployment configuration. The first 5 are selected from module MultiHopLQI: rawRSSI, rawLQI, gbCurrentParent, gbCurrentHopCount, and gbCurrentLinkEst (lines 8–10). These variables correspond to the RSSI of the parent link, the LQI of the parent link, the parent's network address, the distance to the root node (in hops), and an internal link quality measure defined by the application, respectively. Finally, ledsOn, declared within the LedsC module, and low32, declared within SimpleTimeM, are selected (elided). These variables correspond to the current state of the LEDs and the low order bytes of the mote's clock, respectively.

Listing 1.

SurgeTelos Experimentation Script (partial).

After configuring each mote to run SurgeTelos at the current power level (elided), the application is installed on each node in the network (lines 16–18). The subsequent wait command blocks until all pending installation requests are complete (line 19)⁸. The shell command is used to invoke the UNIX sleep program to allow the network to execute for 1 minute before profiling begins (line 20).

The profiling portion of the script consists of two nested loops. The first loop iterates over the target nodes (lines 21–32); the second iterates over the variables selected for profiling (lines 26–29). The effect is to query the current value of each profiled variable on each device in the network. The query command used in the innermost loop body updates the respective NESTShell profiling file, and displays the retrieved value. The echo command is used to display each mote's address to allow users to associate profiling results with particular nodes (line 22). This information might, for example, be used to analyze the impact of node placement on regional link quality (since each address implies the spatial coordinates of the associated device).

Note that the experimentation script queries the value of the SimpleTimeM.low32 variable twice for each mote, once before querying the other variables of interest, and once after (lines 24, 30). This is a common NESTShell experimentation idiom, enabling users to bound the time at which recorded variable values held. If a time synchronization component is included in the installed application image, this approach can be extended to impose a partial ordering on distributed system events. Interval bounds can be tightened by interleaving additional time queries between the queries for each of the other variables.

It is interesting to note that if the above experiment were performed using the graphical user interface presented in Section 3, it would have required over 400 manual actions for a single run.

Multi-phase experiments can be even more complex. Phase transitions might be triggered by wall clock time, network state conditions, or remote IP packets. Triggering these transitions reliably using the graphical interface would present significant obstacles. However, the graphical interface is appropriate to other scenarios, especially those involving a low degree of repetition, novice users, or classroom use. The open design of the NESTbed system enables users to select the most appropriate interface for the task at hand and to extend the system with additional interfaces if necessary.

5. Applications

To gain a sense of the system's benefits, it is useful to consider some of the contexts in which it has been applied. The system is in active use by researchers and students from Clemson University and Cleveland State University. The main physical deployment is installed on the Clemson campus, and is accessed by users from both institutions through the interfaces described in the preceding sections. A secondary deployment is installed on the Cleveland State campus. Here we summarize a few key usage scenarios that illustrate the types of research and teaching activities enabled by these deployments. We focus on the use of the NESTbed system and the types of results it can be used to collect, rather than on the technical details of the contexts themselves.

5.1. Reliable Communication

Recall that a user may substitute an alternative radio stack implementation as part of the deployment configuration process. Section 3 highlighted ReliableComm, an alternative implementation designed to improve the reliability of wireless network links. Before its inclusion as part of the NESTbed distribution, the implementation was developed, debugged, and evaluated using the NESTbed system.

When testing ReliableComm using the NESTbed system, we found that some nodes exhibited unexpected behaviors under high load. First, and most obviously, the faulty nodes exhibited unexpected LED states. Second, by observing the values of program variables at key system execution points, we were able to determine that certain program invariants had been violated. The runtime observations made possible by the NESTbed design ultimately led us to discover a synchronization error in one of the interrupt driven state machines. Based on our past experiences debugging similar errors, we believe the NESTbed design allowed us to correct the error in a small fraction of the time it would otherwise have taken. The runtime network and the program level visibility the design affords is an important debugging aid, especially when debugging problems that are difficult to replicate. Concurrency and memory-related errors, for example, are notoriously difficult to reproduce — and these are exactly the types of errors repeatedly encountered when developing embedded sensor systems.

To evaluate the performance of ReliableComm, we used the NESTbed interface to install a test program on each mote. The program instructs the host device to transmit packets to each of its neighbors at a specified rate for a specified duration. Each mote also records the number of messages received from each of its neighbors. By dividing the number of messages received on each link by the number of messages transmitted on the link, we are able to calculate the PRR on the link.

To control the experiment, we developed a Java application to communicate with a designated leader node through its USB port. To enable connectivity from a remote location, we used the NESTbed interface to construct a network gateway, and then connected the Java application to this gateway. Upon termination of the experiment, we collected the results through the same gateway. The application was tested using both GenericComm and ReliableComm to compare the relative performance of the radio stack implementations. The NESTbed rewiring interface was used to perform this configuration step, eliminating the need for additional programming. A portion of the performance results collected using the NESTbed system are shown in Figs. 13, 14, and 15.

Figure 13.

Link Quality (GenericComm).

Figure 14.

Link Quality (ReliableComm).

Figure 15.

Aggregate network link quality.

Figure 13 presents a graphical representation of link quality, as measured by the test application deployed with GenericComm. Figure 14 is analogous and corresponds to ReliableComm. Each column represents a transmitting node, and each row represents a receiving node; the shading of the cell at their intersection represents the PRR of the link. The cells are shaded on a uniform scale from black to white, with black denoting a PRR of 0%, and white denoting a PRR of 100%. The aggregate impact on link quality is illustrated by the histogram shown in Fig. 15. There are 10 link categories considered; vertical bars represent the number of links in each category. By examining the results captured in these figures—results made possible by the NESTbed system— it becomes immediately clear that ReliableComm significantly improves the reliability of mid quality links without negatively impacting high quality links. Packets destined for low quality links are silently discarded to reduce network congestion.

The implementation of ReliableComm is interesting in its own right, and serves as a useful radio stack alternative for NESTbed users. However, we summarize our development and evaluation experiences here only to emphasize the types of remote evaluation studies made possible by the NESTbed design.

5.2. Student Experimentation

In addition to its use as a research instrument, the NESTbed system is a valuable teaching tool. At Clemson the system is used as part of a graduate course in embedded sensor network design. The course covers both algorithmic issues and software engineering principles as they relate to the development of large-scale embedded deployments.

Given the emphasis on large networks, scalability is a major theme of the course. In the first offering, before the NESTbed system was available for use, scalability issues were difficult to motivate, and even more difficult to analyze. Each student had access to only a small number of motes (i.e., 5–10) on which to perform their assignments. As a result, they were unable to gain experience addressing realistic congestion problems, hardware load limitations, synchronization defects, and other difficulties magnified in large-scale networks. Assignment solutions lacking required scalability properties might appear correct since they could not be tested at scale. Further, with only a few LEDs (and perhaps a few pre-programmed debugging messages) to expose the inner workings of an algorithm implementation, it was difficult for students to evaluate implementation correctness and performance.

In the second offering of the course, students were additionally given access to the NESTbed system. Consider as an example, the second major assignment, for which students were required to demonstrate the correctness of their solutions on a minimum of 35 nodes. The assignment required the development of a multi-hop sensing application that closely parallels the scenario presented in Section 3. The most challenging aspect of the assignment was the development of a self-stabilizing routing tree that tolerates multiple node failures.

The NESTbed system made it easy for students to quickly install their applications on a large number of nodes, and to configure the radio power level of each device to ensure the construction of interesting routing topologies. During each demonstration, the system was used to query the values of program variables on a subset of the network. More specifically, the routing tree component was examined to determine the parent identifier of each node, as well as the distance of each node from the root of the tree. By inspecting the variables that store this information at various points in the network, it was easy to construct quick maps of the routing tree. One such tree, constructed using information gleaned through profiling on a slightly larger network, is shown in Fig. 16.

Figure 16.

Routing tree (pre-fault).

Each circle in the figure represents a network node; nodes selected for profiling are shown in boldface. The identifier of each node is shown at the top of its corresponding circle. The pair of numbers at the bottom of each circle represent the values retrieved during profiling — the identifier of the node's parent and its distance from the root, respectively. The arrows between each node depict the routing links implied by these values. (Note that even in this spatially constrained deployment, the NESTbed power management features enable interesting network topologies.) The rapid development of such a map, made possible by the NESTbed system, provides immediate student feedback concerning the correctness of their routing implementation.

After demonstrating the initial formation of the routing tree, students were required to demonstrate tolerance to node failure. To achieve this, power was cut to a subset of the network nodes, causing immediate fail stop faults. The query system was again used to construct a map of the new routing tree. Figure 17 shows the routing map constructed immediately following a fail stop fault involving node 4, marked with an “X” in the figure. Arrows with solid heads represent routing links unaffected by the fault. Arrows with hollow heads represent affected links. Again, this type of map provides immediate feedback that would otherwise be time-consuming and error prone to collect.

Figure 17.

Routing tree (post-fault).

6. Related Work

The difficulty of testing, debugging, and profiling wireless sensor systems is well recognized. A number of tool-based solutions have been proposed to address these difficulties. We survey some of the most relevant here.

Network Simulators

Several platform-independent wireless network simulators have been discussed in the literature [39, 70, 5]; platform-specific sensor network simulators have also emerged [35, 55, 50]. These tools have proven effective in providing initial measures of correctness and performance. The point of departure for our work, however, was the observation that they have not supplanted the need for physical experimentation. Simulators offer limited fidelity with respect to modeling wireless signal propagation and interference [71, 54], as well as in capturing the behavioral subtleties of underlying hardware platforms. Consequently, our focus has been on infrastructure support for physical experimentation, debugging, and profiling.

Hybrid Simulators

Hybrid approaches that combine aspects of physical experimentation and network simulation have also been proposed. The key idea is to identify system aspects that cannot be faithfully simulated, and factor them out to the physical world. In the context of wireless system simulation, this typically involves deferring communication to physical devices while simulating other system aspects (e.g., application execution, network traffic). Hybrid testbeds have been used to evaluate both wired and wireless ethernet networks [72, 16, 68]. Similar approaches have been used in the context of wireless sensor networks. The EmStar development platform [25], for example, targets microservers, Linux based sensor nodes with computational resources equivalent to a PDA device. The platform can also target Linux based desktops and servers. Applications developed using EmStar can be simulated using EmSim, a simulator that allows physical radios to be used in place of simulated network channels. EmTOS [26], an extension of EmStar, allows applications developed using nesC and TinyOS to be simulated by a microserver (or desktop/server). As a result, the architecture supports hybrid simulation, enabling designers to experiment with alternative realizations of physical network interfaces. The SeNeTs framework [6] provides similar features, but offers support for large scale simulation through the use of distributed processing. While hybrid frameworks have yielded important research results, they have not addressed the fidelity issues associated with simulating mote hardware. They cannot, for example, be used to gather precise results concerning the effects of hardware interrupts, load-induced execution anomalies, or other complex phenomenon that cannot be faithfully simulated. Moreover, results obtained using a particular network interface and supporting software stack rarely apply to other interfaces and network drivers. By contrast, our work supports pure physical experimentation using standard hardware components, offering the highest degree of experimental fidelity.

Network Testbeds

We are not the first to describe a network testbed designed to support pure physical experimentation. Several 802.11 efforts have been discussed in the literature [34, 47, 44, 7], and more recently, testbeds focused on wireless sensor systems have emerged. Our work aligns most closely with the latter category. We consider some of the most important testbed efforts representative of the current state-of-the-art.

Harvard's MoteLab testbed [67, 65] was one of the first sensor network testbeds discussed in the literature. The physical network includes 190 Tmote Sky [42] devices. Each mote is attached to an ethernet based gateway device [41], allowing the network to be reprogrammed from a centralized server. The server exposes a web interface that allows users to upload executable application images, and to configure the deployment of those images on the physical network. The system also allows users to upload Java classes that can be used to log USB data. The NESTbed approach to creating network gateways for injecting packets from a remote location is based on a similar feature available in MoteLab. In contrast to the NESTbed system, however, MoteLab is batch based rather than interactive; submitted jobs are queued for later execution. As a result, the system does not support real-time source or network level profiling⁹, nor does it support the injection of transient state faults. The design is also image-centric, requiring users to generate application images, as well as to construct the Java classes that parse application data transmitted over the USB port. In addition to the productivity benefits provided by the NESTbed system, its source-centric design introduces opportunities for automated source-level analysis and instrumentation. Finally, the MoteLab server appears to be closed; it does not seem to expose an API for programmatic control, restricting users to a single web interface for all experimentation tasks. There is no equivalent, for instance, of the NESTShell scripting interface, nor an apparent mechanism to add such an interface.

More recently, Ohio State deployed the Kansei testbed [21, 2, 45]. The supporting physical network is one of the largest to date, with over 400 devices. The testbed supports experimentation over multi-tiered networks; the deployment includes Extreme Scale [18], Tmote Sky [42], and Trio motes [19]. The basic hardware architecture is similar to that of MoteLab. Motes are attached to ethernet based gateway devices [14], and are programmed through a centralized server. A key point of novelty in Kansei is its focus on sensing experiments. Sensor nodes are housed in stationary, portable, and mobile arrays. Portable and mobile arrays are used to collect field data for ex post facto analysis. Alternatively, the arrays can be used to inject data into the (larger) stationary deployment using a forwarding system similar to that of MoteLab and NESTbed. Unlike these systems, however, Kansei includes support for sensor stream scaling through parametric, probabilistic, and sample based data generation. Perhaps most interesting is the third option which generates large scale sensor streams by replaying pre-recorded data (usually from a smaller portable array) with temporal and spatial shifts. The software architecture also provides some support for job coordination.

Kansei is well-suited to batch style experimentation, especially when the experiments are focused on high fidelity sensing of parameters that cannot be captured in a laboratory context. The NESTbed system offers a complementary design; it is engineered to support interactive use, with a focus on software experimentation. Like MoteLab, Kansei does not provide real-time profiling or fault injection support. It also provides limited support for logging network traffic [46]. Further, Kansei is image-centric, precluding source-level analysis and instrumentation. One consequence seems to be that developers are required to integrate specialized Kansei components before compiling and uploading their application images. Finally, the degree of controllability provided to external applications by the Kansei API is unclear. The PHP based design seems to suggest a closed system, precluding the addition of interface extensions such as the NESTShell scripting interface provided by the NESTbed system.

MoteLab and Kansei exemplify testbed development projects underway at research institutions around the world (e.g., [61, 28, 33, 60, 10, 9, 64])¹⁰. While these testbeds share similarities with our work, they are principally batch based, image-centric, and closed. By contrast, the NESTbed design is interactive, source-centric, and open.

Other Tools

In addition to testbed infrastructures, other related development and testing tools have also recently been described in the literature. Most relevant to our work are tools designed to improve runtime observability and controllability of network software. The Deployment Support Network (DSN) [20] is a key example. The approach is to connect a secondary device, the “DSN node,” to each target mote. The DSN nodes form a reliable out-of-band backbone for controlling and observing a target mote network. The current implementation uses Bluetooth based BTnodes [22] for the out-of-band backbone. These devices are managed through a centralized server that exposes an RPC interface to remote processes. The interface is used to deploy new application images, transmit and receive messages from target devices, and monitor the status of DSN nodes and their targets. In- and out-of-network buffering strategies are used to provide reliable communication. The key benefit of the toolkit is to eliminate the need for wired mote connections. In effect, the toolkit serves as a replacement for USB connections in Tmote based (and other sensor) deployments¹¹. It may be worth noting that the DSN approach has been used to construct a heterogeneous network testbed consisting of 66 nodes at ETH Zurich [23]. The testbed has features (and consequent limitations) analogous to those of the MoteLab testbed.

Recall from Section 3.3 that the NESTbed system uses Nucleus [56] to provide variable inspection and modification facilities. Marionette [69] is a significant extension of Nucleus to support RPC based interactive development and debugging. The system provides a Python interface for exploring static program structures (e.g., modules, type declarations), reading and modifying the program state at runtime, and invoking nesC commands. Like the NESTShell interface, Marionette enables developers to script debugging and profiling activities. It is not, however, tailored for testbed experimentation; it lacks services for managing projects and deployment configurations, reprogramming devices, constructing network gateways, and others. Further, mote interactions are handled in-band, limiting throughput and reliability. Marionette's integration with a popular object based scripting language, however, is a point of advantage over the NESTShell interface. Finally, it may be worth noting that the NESTbed graphical interface bears some similarity to existing integrated development environments for nesC and TinyOS. In particular, it shares design characteristics with the various Eclipse plugins for TinyOS [49, 48, 57], as well as TOSDev [40]. By contrast to the NESTbed system, however, these tools are focused on providing syntactic assistance (e.g., syntax highlighting, code completion), and managing source distributions and component dependencies. They provide support for programming a single device; they do not provide testbed related features.

7. Conclusion

The goal of our work is to improve the ability of developers to test, debug, and profile wireless sensor systems. To this end, we presented the design and implementation of the NESTbed system, a network testbed specialized for sensor system development. The design exposes an expandable set of hardware resources to a distributed research community as, in effect, a shared virtual device. The supporting software infrastructure is engineered to enable rapid configuration, deployment, and analysis on a large-scale mote installation. The novelty of the system lies in its interactive, source-centric, and open design. We presented the NESTbed architecture, the NESTbed server API, two client interfaces, and system use-case scenarios involving the development of a multi-hop sensing application. We additionally summarized current applications of the system in the teaching and research activities of client institutions.

We are planning three NESTbed extensions as part of our future work. First, we are developing a distributed predicate detection system. The enhancement will enable users to monitor the status of distributed state conditions relevant to system correctness and performance. The enhancement will be especially valuable in studying the behavior of self-stabilizing protocols. Second, we are developing a distributed control flow visualization enhancement. The enhancement will enable NESTbed users to trace the control flow of system components across network nodes. The enhancement will be of particular value in reasoning about concurrent system behaviors. Finally, we plan to integrate support for alternative programming languages and programming language extensions. The focus will be on languages and extensions that compile to nesC and TinyOS (e.g., nesC RPC extensions [38], the DESAL programming language [3]).

One goal of this paper is to assist in bootstrapping a community interested in using our implementation and in replicating our efforts at other institutions. The software tools described in this paper, as well as instructions for their installation and use, are available for download from the project website (http://dsrg.cs.clemson.edu/nestbed/). We encourage other researchers to evaluate our work, and request their suggestions for future enhancements.

Footnotes

1

We use the term “profiling” to refer to the process of collecting system events and state information during program execution (e.g., state changes, radio events, voltage data). The collected data may be analyzed in real time, or stored for ex post facto analysis.

2

Our current API implementation uses the Java Remote Method Protocol (JRMP) for interprocess communication, requiring that client interfaces be implemented in Java. With minor modifications, the Internet Inter-Orb Protocol (IIOP) could be used [] to enable CORBA compatibility, eliminating this restriction.

3

Note, however, that some experimental controls can not be preserved in software. External network interference, for instance, may vary from one experimental run to another.

4

The SerialForwarder libraries are packaged as part of the TinyOS distribution [].

5

The NESTbed distribution includes two alternative radio stack implementations; these are the only operating system alternatives available for selection through the graphical interface. The system is, however, extensible to an arbitrary number and type of alternative services.

6

The Instrumentation and Compilation API assumes interface compatibility between selected components and user provided alternatives. Syntactic errors introduced during the instrumentation process due to interface violations in user provided components will be reported at compile time. Semantic errors can of course not be checked.

7

The Message Interface Generator included as part of the TinyOS distribution is used to create this class [].

8

We have omitted standard error checking and recovery logic for the sake of presentation

9

This discounts the possibility of forwarding raw packet data through a gateway for remote inspection.

10

It may be useful to note that MoteLab, Kansei, the Deployment Support Network, and other testbeds include integrated health monitoring services. The basic approach is to periodically poll each device to determine whether it is in a programmable state. Unresponsive nodes are avoided by manual and automatic allocation strategies. Since the NESTbed system is intended for interactive use, users are notified of device problems at the point of installation (as indicated by programming failures). Unresponsive nodes can be power-cycled through the NESTbed interface. Hence, while useful in batch-based systems, the benefit of periodic health monitoring is unclear in the context of the NESTbed design.

11

We intend to investigate the use of DSN in future NESTbed enhancements.

Acknowledgements

The work presented in this article was funded in part by the National Science Foundation (CNS-0520222) and the South Carolina Space Grant Consortium. The authors gratefully acknowledge these agencies for their support. The authors also wish to thank Sally K. Wahba for her work in characterizing the relationship between packet reception rate, transmission distance, and transmission power (Section 3.4, ). Finally, the authors wish to thank the anonymous reviewers for their detailed comments and suggestions, which were of great value in improving this article.

References

Arora

Dutta

Bapat

Kulathumani

Zhang

Naik

Mittal

Cao

Demirbas

Gouda

Choi

Herman

Kulkarni

Arumugam

Nesterenko

Vora

Miyashita

“A line in the sand: A wireless sensor network for target detection, classification, and tracking,”

Computer Networks vol. 46, no. 5, pp. 605–634, 2004.

Arora

Ertin

Ramnath

Nesterenko

Leal

, “Kansei: A high fidelity sensing testbed. IEEE Internet Computing,” vol. 10, no. 2, 35–47, 2006.

Arora

Gouda

Hallstrom

Herman

Leal

Sridhar

, “A statebased language for sensoractuator networks,” in The International Workshop on Wireless Sensor Network Architecture, p. 6. IEEE Computer Society Press, April 2007.

Arora

Ramnath

Ertin

Sinha

Bapat

Naik

Kulathumani

Zhang

Cao

Sridharan

Kumar

Seddon

Anderson

Herman

Trivedi

Zhang

Gouda

Choi

Nesterenko

Shah

Kulkarni

Aramugam

Wang

Culler

Dutta

Sharp

Tolle

Grimmer

Ferriera

Parker

, “Exscal: Elements of an extreme scale wireless sensor network,” in The 11th IEEE International Conference on Embedded and Real-time Computing Systems and Applications, pp. 102–108. IEEE Computer Society Press, August 2005.

Barr

Haas

Renesse

“Scalable wireless ad hoc network simulation,” in Wu

, ed., Handbook on Theoretical and Algorithmic Aspects of Sensor, Ad Hoc Wireless, and Peer to Peer Networks, pp. 297–311. CRC Press, 2005.

Blumenthal

Reichenbach

Golatowski

Timmermann

, “Controlling wireless sensor networks using SeNeTs and EnviSense,” in The 3rd IEEE International Conference on Industrial Informatics, pp. 262–267. IEEE Computer Society, August 2005.

Chambers

, “The grid roofnet: a rooftop ad hoc wireless network,” Master's thesis, Massachusetts Institute of Technology, May 2002.

Chintalapudi

Paek

Kothari

Rangwala

Caffrey

Govindan

Johnson

Masri

, “Monitoring civil structures with a wireless sensor network,” IEEE Internet Computing vol. 10, no. 2, pp. 26–34, 2006.

Chun

, “Mirage User Manual,” https://mirage.berkeley.intelresearch.net/usermanual/single-page.html, 2004.

10.

Chun

Buonadonna

AuYoung

Parkes

Shneidman

Snoeren

Vahdat

, “Mirage: A microeconomic resource allocation system for sensornet testbeds,” in The 2nd IEEE Workshop on Embedded Networked Sensors, p. 10. IEEE Computer Society, May 2005.

11.

Compaq Computer Corporation, Hewlett Packard Company, Intel Corporation, Lucent Technologies Inc., Microsoft Corporation, NEC Corporation, and Koninklijke Philips Electronics N.V., “Universal serial bus specification (revision 2.0),” http://www.usb.org/developers/docs/usb2005122006.zip, 2000.

12.

Crossbow Technology Incorporated, “Mica2 datasheet,” http://www.xbow.com/Products/Productpdffiles/Wirelesspdf/6020004206BMICA2.pdf, 2003.

13.

Crossbow Technology Incorporated, “Mica2Dot datasheet,” http://www.xbow.com/Products/Productpdffiles/Wirelesspdf/6020004304CMICA2DOT.pdf, 2003.

14.

Crossbow Technology Incorporated, “Stargate datasheet,” http://www.xbow.com/Products/Productpdffiles/Wirelesspdf/Stargate Datasheet.pdf, 2003.

15.

Dalton

Hallstrom

, “A file system abstraction and shell interface for a wireless sensor network testbed,” in The 3rd International Conference on Testbeds and Research Infrastructures for the Development of Networks and Communities. IEEE Computer Society, May 2007.

16.

Raniwala

Sharma

Chiueh

, “MiNT: A miniaturized network testbed for mobile wireless research.” in The 24th Annual Joint Conference of the IEEE Computer and Communications Societies, pp. 2731–2742. IEEE Computer Society, March 2005.

17.

Doolin

Sitar

, “Wireless sensors for wildfire monitoring,” in SPIE Symposium on Smart Structures and Materials / NDE 2005, pp. 477–484. SPIE Press, March 2005.

18.

Dutta

Grimmer

Arora

Bibyk

Culler

, “Design of a wireless sensor network platform for detecting rare, random, and ephemeral events,” in The 4th International Conference on Information Processing in Sensor Networks, pp. 497–502. IEEE Computer Society, April 2005.

19.

Dutta

Hui

Jeong

Kim

Sharp

Taneja

Tolle

Whitehouse

Culler

, “Trio: Enabling sustainable and scalable outdoor wireless sensor network deployments,” in Proceedings of the 5th International Conference on Information Processing in Sensor Networks, pp. 407–415. ACM Press, April 2006.

20.

Dyer

Beutel

Thiele

Kalt

Oehen

Martin

Blum

, “Deployment support network a toolkit for the development of WSNs,” in The 4th European Conference on Wireless Sensor Networks, pp. 195–211. Springer Verlag, January 2007.

21.

Ertin

Arora

Ramnath

Nesterenko

Naik

Bapat

Kulathumani

Sridharan

Zhang

Cao

, “Kansei: A testbed for sensing at scale,” in The 5th International Conference on Information Processing in Sensor Networks, pp. 399–406. ACM Press, April 2006.

22.

ETH Zurich, “BTnode rev. 3 hardware reference,” http://www.btnode.ethz.ch/Documentation/BTnodeRev3HardwareReference, 2007.

23.

ETH Zurich, “Deployment support network tool,” http://tecpcbtnode.ethz.ch/dsngui/index.php?act=GetDSN, 2007.

24.

Gay

Levis

von Behren

Welsh

Brewer

Culler

, “The nesC language: A holistic approach to networked embedded systems,” in ACM SIGPLAN 2003 Conference on Programming Language Design and Implementation, pp. 1–11. ACM Press, June 2003.

25.

Girod

Elson

Cerpa

Stathopoulos

Ramanathan

Estrin

, “EmStar: a software environment for developing and deploying wireless sensor networks,” in The 2004 USENIX Technical Conference, USENIX Association, June–July 2004.

26.

Girod

Stathopoulos

Ramanathan

Elson

Estrin

Osterweil

Schoellhammer

, “A system for simulation, emulation, and deployment of heterogeneous sensor networks,” in The 2nd International Conference on Embedded Networked Sensor Systems, pp. 201–213. ACM Press, November 2004.

27.

Glaser

, “Some real world applications of wireless sensor nodes,” in SPIE Symposium on Smart Structures & Materials / NDE 2004, pp. 344–355. SPIE Press, March 2004.

28.

Handziski

Köpke

Willig

Wolisz

, “TWIST: a scalable and reconfigurable testbed for wireless indoor experiments with sensor networks,” in Proceedings of the 2nd International Workshop on Multi hop Ad Hoc Networks: From Theory to Reality, pp. 63–70. ACM Press, January 2006.

29.

Hartung

Holbrook

Han

Seielstad

, “FireWxNet: a multitiered portable wireless system for monitoring weather conditions in wildland fire environments.” in The 4th International Conference on Mobile Systems, Applications, and Services, ACM Press, June 2006. (to appear).

30.

Hill

Szewczyk

Woo

Hollar

Culler

Pister

, “System architecture directions for networked sensors,” in The 9th International Conference on Architectural Support for Programming Languages and Operating Systems, vol.34, pp. 93–104. ACM Press, November 2000.

31.

Tran

Bulusu

Chou

Jha

Taylor

, “The design and evaluation of a hybrid sensor network for canetoad monitoring,” in The 4th International Symposium on Information Processing in Sensor Networks, pp. 503–508. IEEE Computer Society Press, April 2005.

32.

Hui

Culler

, “The dynamic behavior of a data dissemination protocol for network programming at scale,” in The 2nd International Conference on Embedded Networked Sensor Systems, pp. 81–94. ACM Press, November 2004.

33.

Johnson

Stack

Fish

Flickinger

Ricci

Lepreau

, “TrueMobile: A mobile robotic wireless and sensor network testbed,” in The 25th Annual Joint Conference of the IEEE Computer and Communications Societies, IEEE Computer Society, April 2006.

34.

Karygiannis

Antonakakis

, “mLab: a mobile ad hoc network testbed,” in The 1st Workshop on Security, Privacy, and Trust in Pervasive and Ubiquitous Computing, pp. 88–97. Diavlos S.A., July 2005.

35.

Levis

Lee

Welsh

Culler

, “TOSSIM: Accurate and scalable simulation of entire TinyOS applications,” in The 1st ACM Conference on Embedded Networked Sensor Systems, pp. 126–137. ACM Press, November 2003.

36.

Lorincz

Malan

Jones

T. Fulford

Nawoj

Clavel

Shnayder

Mainland

Welsh

Moulton

, “Sensor networks for emergency response: Challenges and opportunities,” IEEE Pervasive Computing vol. 3, no. 4, pp. 16–23, 2004.

37.

Mainwaring

Polastre

Szewczyk

Culler

Anderson

, “Wireless sensor networks for habitat monitoring,” in The 1st ACM International Workshop on Wireless Sensor Networks and Applications, pp. 88–97. ACM Press, September 2002.

38.

May

Dunning

Dowding

Hallstrom

, “An RPC design for wireless sensor networks.” Journal of Pervasive Computing and Communications, vol.1, no. 1, 2007. (to appear).

39.

McCanne

Floyd

Network simulator ns 2. http://www.isi.edu/nsnam/ns/, 1997.

40.

McCartney

Sridhar

, “TOSDev.,” http://selab.csuohio.edu/dsnrg/tosdev/index.html, 2006.

41.

Moteiv Corporation, “Tmote Connect datasheet,” http://www.moteiv.com/products/docs/tmoteconnectdatasheet.pdf, 2006.

42.

Moteiv Corporation, “Tmote Sky datasheet,” http://www.moteiv.com/products/docs/tmotesky datasheet.pdf, 2006.

43.

Musăloiu

Terzis

E. A.

Szlavecz

Szalay

Cogan

Gray

, “Life under your feet: A wireless soil ecology sensor network,” in The 3rd Workshop on Embedded Networked Sensors, pp. 51–55, May 2006.

44.

Nordström

Gunningberg

Lundgren

, “A testbed and methodology for experimental evaluation of wireless mobile ad hoc networks,” in The Proceedings of the 1st International Conference on Testbeds and Research Infrastructures for the Development of Networks and Communities, pp. 100–109. IEEE Computer Society, February 2005.

45.

Ohio State University, “Kansei sensor testbed,” http://ceti.cse.ohiostate.edu/kansei/, 2006.

46.

Ohio State University, “Using Kansei — basics,” http://exscal.nullcode.org/kansei/help.php, 2006.

47.

Raychaudhuri

Seskar

Ott

Ganu

Ramachandran

Kremo

Siracusa

Liu

Singh

, “Overview of the orbit radio grid testbed for evaluation of next generation wireless network protocols,” in Wireless Communications and Networking Conference vol. 3, pp. 1664–1669. IEEE Computer Society Press, March 2005.

48.

Sallai

Balogh

Dora

, “TinyDT – TinyOS plugin for the Eclipse platform,” http://www.tinydt.net/, 2005.

49.

Schuler

Burri

, “TinyOS plugin for Eclipse,” http://www.dcg.ethz.ch/~rschuler/, 2006.

50.

Simon

Volgyesi

Maroti

Ledeczi

, “Simulation based optimization of communication protocols for large-scale wireless sensor networks,” in The 2003 IEEE Aerospace Conference vol. 3, pp. 1339–1346. IEEE Computer Society, March 2003.

51.

Sun Microsystems, “java.lang.reflect (Java 2 platform SE 5.0),” http://java.sun.com/j2se/1.5.0/docs/api/java/lang/reflect/packagesummary.html, 2004.

52.

Sun Microsystems, “Java(TM) remote method invocation (Java RMI),” http://java.sun.com/j2se/1.5.0/docs/guide/rmi/index.html, 2004.

53.

Sun Microsystems, “Java(TM) RMI over HOP documentation,” http://java.sun.com/j2se/1.5.0/docs/guide/rmiiiop/index.html, 2004.

54.

Takai

Martin

Bagrodia

, “Effects of wireless physical layer modeling in mobile ad hoc networks,” in The 2nd ACM International Symposium on Mobile Ad Hoc Networking & Computing, pp. 87–94. ACM Press, October 2001.

55.

Titzer

Lee

Palsberg

, “Avrora: Scalable sensor network simulation with precise timing in The 4th International Symposium on Information Processing in Sensor Networks, pp. 477–482. IEEE Computer Society Press, April 2005.

56.

Tolle

, “Nucleus network management,” http://www.cs.berkeley.edu/∼get/nucleus/nucleusmanual.pdf, 2005.

57.

Tynan

, “TinyOS Eclipse plugin,” http://tide.ucd.ie/, 2006.

58.

UC Berkeley, “MIG — message interface generator for nesC,” http://www.tinyos.net/tinyosl.x/doc/nesc/mig.html, 2003.

59.

UC Berkeley, “TinyOS community forum: an open source OS for the networked sensor regime,” http://www.tinyos.net/, 2004.

60.

UC Berkeley, “Soda hall wireless sensor network testbeds,” http://www.millennium.berkeley.edu/sensornets/, 2006.

61.

University of Southern California, “Tutornet: A tiered wireless sensor network testbed,” http://enl.usc.edu/projects/testbed/, 2006.

62.

Wahba

Hallstrom

, “An empirical evaluation of embedded link quality,” in The International Conference on Sensor Technologies and Applications, page (to appear). IEEE Computer Society Press, pp. 430–435, October 2007.

63.

Warneke

Last

Liebowitz

Pister

, “Smart dust: Communicating with a cubic millimeter computer,” Computer, vol.34, no. 1, 44–51, 2001.

64.

Welsh

Fish

Frantz

, “GNOMES: a testbed for low power heterogeneous wireless sensor networks,” in IEEE International Symposium on Circuits and Systems, pp. 836–839. IEEE Computer Society Press, May 2003.

65.

Welsh

Allen

G. Werner

, “MoteLab: Harvard network sensor testbed,” http://motelab.eecs.harvard.edu/index.php, 2006.

66.

WernerAllen

Lorincz

Welsh

Marcillo

Johnson

Ruiz

Lees

, “Deploying a wireless sensor network on an active volcano,” IEEE Internet Computing vol. 10, no. 2, pp. 18–25, 2006.

67.

Allen

G. Werner

Swieskowski

Welsh

, “MoteLab: a wireless sensor network testbed,” in The 4th International Conference on Information Processing in Sensor Networks, pp. 483–488. IEEE Computer Society Press, April 2005.

68.

White

Lepreau

Stoller

Ricci

Guruprasad

Newbold

Hibler

Barb

Joglekar

, “An integrated experimental environment for distributed systems and networks,” in The 5th Symposium on Operating Systems Design and Implementation, pp. 255–270. ACM Press, December 2002.

69.

Whitehouse

Tolle

Taneja

Sharp

Kim

Jeong

Hui

Dutta

Culler

, “Marionette: Using RPC for interactive development and debugging of wireless embedded networks,” in The 5th International Conference on Information Processing in Sensor Networks, pp. 416–423. ACM Press, April 2006.

70.

Zeng

Bagrodia

Gerla

, “GloMoSim: A library for parallel simulation of large-scale wireless networks,” in The 12th Workshop on Parallel and Distributed Simulation, pp. 154–161. IEEE Computer Society, May 1998.

71.

Zhou

Krishnamurthy

Stankovic

, “Impact of radio irregularity on wireless sensor networks,” in The 2nd International Conference on Mobile Systems, Applications, and Services, pp. 125–138. ACM Press, June 2004.

72.

Zhou

Bagrodia

, “TWINE: a hybrid emulation testbed for wireless networks,” in The 25th IEEE Conference on Computer Communications, IEEE Computer Society Press, April 2006.