A Case for a Sensor Network Architecture - Semantic Scholar

A Case for a Sensor Network Architecture Prabal Dutta? , Cheng Tien Ee∗, Rodrigo Fonseca? , Jonathan Hui? , Philip Levis? , Joseph Polastre? , Gillman Tolle? , Jerry Zhao†, David Culler? , Scott Shenker? † , Ion Stoica? 1.

INTRODUCTION

chitecture would identify the essential components and their conceptual relationships so that it would become possible to compose components in a manner that transcends particular generations of technology, allows innovation, and promotes interoperability.

Wireless Sensor Networks have the potential for tremendous societal benefit by enabling new science, better engineering, improved productivity, and enhanced security. Research in this area has progressed dramatically in the past decade. The hardware, particularly radio technology, is improving rapidly, leading to cheaper, 2. THE NATURE OF AN ARCHITECTURE faster, smaller, and longer-lasting nodes. Many One goal of an architecture is to allow comsystems challenges, such robust multihop routponents developed for one system to be used ing, effective power management, precise time in many other systems, especially those from synchronization, and efficient in-network query other developers. However, the need to build processing, have been tackled and several comcomposable components is hardly unique to senplete applications, where all these components sornets; it is a standard challenge in building have been integrated into a coherent system, large software systems. The challenge in senhave been deployed and demonstrated, includsor networks is that applications are meaninging some at relatively large scale [9, ?]. fully distributed over a large number of relaBut the situation in sensornets, while promistively constrained nodes that are embedded in ing, is also problematic. The literature presents physical space. The application dictates the a veritable alphabet soup of protocols and subsensor modalities and sample rates, the realsystems that make widely differing assumptions time processing and storage of this data, and about the other parts of the system and how the information exchange protocols among colthe parts should interact. The extent to which lections of nodes. Thus, the traditional division these components can be combined to build usof application, operating system, and network able systems is quite limited. In order to prohas been opened to re-examination [4] and isduce running systems, various research groups sues of scheduling, power-control, and informahave produced “vertically integrated” designs in tion flow cut across the boundaries. This situwhich their own set of components are specifiation easily leads to monolithic solutions or to cally designed to work together, but are unable to interoperate with components from other groups. subsystem components with arbitrary interface assumptions. We need to re-establish a meanThis greatly reduces the synergy between reingful separation of concerns. search efforts, and has impeded progress in the An architecture is a set of principles that guide field. where functionality should be implemented along It is the central tenet of this paper that the primary factor currently limiting progress in sen- with a set of interfaces, functional components, protocols, and physical hardware that (roughly) sornets is not any specific technical challenge follows those guidelines. For instance, the Inter(though many remain, and deserve much furnet architecture demonstrated powerfully how ther study) but is instead the lack of an overa properly chosen set of guiding principles can all sensor network architecture. Such an arshape the evolution of a complex system over ∗ vast changes in technology, scale, and usage [1]. CS Division, UC Berkeley, Berkeley, CA 94720. † ICSI, 1947 Center Street, Berkeley, CA 94704 The philosophy of designing for heterogeneity,

change and uncertainty was a radical shift from classical systems design, which more traditionally seeks a near optimal assembly of near optimal parts. Faced with integrating several existing networks with widely varying characteristics, the end-to-end principle and the focus on interoperability led to a design that has successfully coped with unprecedented scale, incorporated a vast array of new underlying technologies, and supported a dizzying variety of applications. However, it was not free of costs; the use of rigid layering sacrificed efficiency in various regards in return for increased interoperability. The power of this architecture is revealed not so much in the elegance or efficiency of its individual components, but in the overall ability to encompass tremendous growth in scale and in diversity as usage and technology rapidly evolved. This is our goal for developing an architecture for sensornets. We must be extremely mindful of any loss of efficiency for particular tasks as we seek to greatly enhance the interoperability between components and ability to advance.

3.

SENSOR-NET PROTOCOL, THE NARROW WAIST

A complete sensornet architecture will need to address a family of specific issues, such as discovery, topology management, naming, routing and so on, but the over-riding question is whether there is “narrow waist” — a unifying abstraction that permits a wide variety of uses above and a range of implementations below. At what level should it occur and what should it express? By requiring all network technologies to support IP, and all applications to run on top of IP, the Internet could accommodate, even encourage, a vast degree of heterogeneity and diversity in both applications and underly-

ing technologies.1 We have an analogous goal for sensornets; in both the application and device arenas we are in the midst of extremely rapid developments. Sensornets will only flourish if we can identify a narrow waist in the architecture that will allow device and protocol developments to proceed apace, while permitting significant optimization. We claim that such a narrow waist can exist, let’s call it Sensor-net Protocol (SP), and that it should be a best-effort single-hop broadcast with a rich enough interface to allow multiple network layer components above to optimize for a range of potential link layers below in a hardware-independent fashion.2 It is not only the resource limitations of sensornet that cause the SP to be closer in nature to the data link layer than the network layer at which IP resides. Applications differ dramatically in their communication patterns and are intimately tied to the associated network protocols. They generally do not require and often do not benefit from a common, universally routable addressing scheme. Instead, a single, simple interface needs to be provided to efficiently implement a range of routing protocols, independent of the underlying link layer, that facilitate in-network processing and collective communication, in addition to point-to-point transport. Moving the point of universal abstraction downward presents new issues that we do not typically concern ourselves about in the Internet architecture. It also requires a careful design of the layers above SP to provide a reasonably general platform on which to build various sensor network applications efficiently. If SP is to be a unifying abstraction with a common format and semantics across many physical layers, how function1

It is straightforward to incorporate sensornets as edge networks of the Internet with gateway nodes providing a bridge. IPv6 addressing makes this considerably easier. In the vast majority of cases, the gateway will also serve as a proxy, so TCP connections would rarely terminate at the actual sensor node. In the proxy case, it is also natural for collections of nodes to appear as a virtual Internet host. The most challenging question is the architecture within the sensornet. This is more than just another subnet, because distributed applications are spread over the many nodes in a manner dependent on its physical embedment. 2 While IEEE 802.2 provides a link interface to a variety of links, higher layers must know what physical layer is below and access the specific physical layer through a standardized set of calls.

In-Network Storage

Address-Free Protocols Suppression Discovery

Time Coordination

Custody Transfer Power Management

System Management

Sensor-Net Application

Security

ality divides across the packet boundary is a key question. We believe that the interface needs to be more expressive than current data link interfaces for higher level control, yet have decomposable functionality below for greater flexibility. To support the network protocols found in the sensornet literature, the underlying mechanisms over which a caller of the send operation should be able to exert control include generation of link level acknowledgments, performance of initial and collision-avoidance backoffs, post-media arbitration timestamping, degree of FEC, retransmission and power management (cf. [11]). In addition to providing such control points, SP should project costs (e.g., energy, delay, storage, bandwidth consumption) up to higher layers so protocols can optimize their behavior and how they exercise the available control. For example, in addition to the traditional question of whether fragmentation occurs below or above the basic communication interface, when designing SP we must consider where protocol exchanges to reduce hidden terminals and improve fairness, such as RTS/CTS, belong. Exposing control of the underlying mechanisms allows middle layers to provide the additional functionality when it is needed at performance that is competitive (or even exceeding) the general solution [11]. Simple defaults allow easy composition and use in basic operation. Allowing multiple network protocols to share control over an underlying link layer raises concern as to how these protocols work together and cooperate. This is just the kind of investigation that the existence of SP would promote. We suggest that this question is tractable and very interesting in sensornets because they typically host a small number of widely distributed applications. Such control is problematic in the Internet because the infrastructure is shared by arbitrary applications anywhere in the world. The classic operating systems boundary must address an arbitrary mix of complex applications. The application specific nature of sensornets is more conducive to cross-layer and crossapplication customization. In addition to the address-free broadcast, SP also has a directed send operation. Returning to the issue of dispatch to physical layers, the exact

Triggers

Name-Based Protocols

Predicates

Caching

Estimation

Graphs

Naming

Sensor-Net Protocol Data Link

Physical Architecture

Media Access Sensing

Timestamping

Energy Storage

Coding

Carrier Sense

Assembly Transmit

ACK Receive

Figure 1: Sensor Network Functional Layer Decomposition form this addressing scheme takes directed communication is an open research question. In the opposite direction of data flow, reception should provide an upward path for additional metadata and physical information, such as time of arrival, signal strength, and source. Additionally, reception demultiplexes packets across multiple network services and protocols.

4.

FUNCTIONAL LAYER DECOMPOSITION

Defining SP as the unifying abstraction leads a different network layer decomposition and r Having presented the design issues and constraints of the data-link and physical layers, we explore the ramifications of SP being the unifying abstraction, proposed an alternative layer decomposition. In addition to layers, this decomposition includes a set of services, such as time coordination, which cut across layers. We discuss two abstractions, the multihop (layer 3) protocol layers and cross-layer power management, where sensor networks diverge enough from their Internet analogues that they deserve particular attention when designing SP.

4.1

A Proposed Architecture

Figure 1 shows a possible layer decomposition of a sensor network architecture. SP is the unifying abstraction that bridges multihop and application protocols to the underlying data link and physical layers. Three basic abstractions lies above SP: address-free protocols, name-based protocols, and in-network storage. As sensor network research and deployments mature, more abstractions will surely emerge: these three are merely some which the sensor network community has established as critical components of

many applications. In addition to the network layering, a sensor network architecture has a set of services which cut across those layers. One common example is time coordination: media access can require time coordination for TDMA or TDMAlike approaches [14], multihop protocols can require time coordination to coordinate transmission times at the network layer [?], applications can require time coordination to schedule their data collection [?]. Correspondingly, the information necessary for time coordination cannot be encapsulated in a single network layer.

An important class of network-layer services provides multi-hop communication based on directly addressable nodes. An appropriate choice of routing protocol and associated addressing structure can enable more efficient communication. Protocols belonging to this class include data collection, dissemination, aggregation, and point-to-point routing. Key architectural issues that arise in designing these protocols include (1) what naming scheme to use, (2) how a node performs packet forwarding and processing relative to the naming scheme (data path), and (3) how does a node discovers and maintains routing and processing state (control path). When SP is the unifying network abstraction, the simplest network operation is a local wireless broadcast. Dissemination protocols and algorithms such as simple floods or Trickle [8] use only this primitive, depending on local connectivity as an implicit naming scheme. This class of protocols is address-free: although they may include names to refer to data items – such as a sequence number – they do not use the concept of a node identifier. As transmit energy is a large concern, address-free protocols can elide destinations addresses when the data link layer allows it. 3 Dissemination can be address-free because it is a network-wide operation. In constrast, directed protocols such as geographic [6] or logical coordinate routing [10, 2] have a specified destination. More abstract and flexible nam-

ing schemes such as directed diffusion use data identifiers [5]. Global network names, however, are powerful enough to support content-based storage within the sensor network, which requires arbitrary point-to-point routing with low stretch [12]. Protocols such as converge-cast/collection routing, where nodes send data up a tree to a sink, are an interesting combination of the two schemes. On one hand, to a user of the protocol, it can be address- and name-free: the send is implicitly up the tree. The protocol is usually implemented, however, with next hop addressing to specify the path to the root. Unlike an IP network, where there is a single network addressing scheme, different multihop protocols built on SP network can have different naming schemes to specify a destination. 4 A name can have either local or global scope, and can be used to identify an individual node, a set of nodes, or a communication structure such as a tree. From the point of view of a node, a name represents a handler that can be used to process an incoming packet and decide on its next hop. A fundamental requirement for building namebased protocols is the formation and maintenance of a connectivity graph, represented by routing tables in each of the nodes. At any point in time, each node can receive transmissions from a set of neighboring nodes with a non-zero packet success probability. Individual nodes perform link estimation to judge connection quality, using passive or active techniques, and build a table of well-defined links that act as an approximation to the underlying physical communication graph. As wireless links are often asymmetric [3, 15], a good incoming link from a node does not necessarily mean there is a good outgoing link to that node. Therefore, to form a bidirectionally-consistent mesh for the whole network, connectivity information must be exchanged between neighboring nodes. This connectivity exchange requires each node to be able to uniquely identify (address) its neighbors. Note that in this case addresses need only be unique within the neighborhood of a node, but do not need to be unique in the entire network. Unlike Internet routing, sensor network rout-

3 For example, 802.15.4 radios such as the CC2420 have several addressing modes: 0-byte, 2-byte, or 8-byte.

4 We use the terms name and address as defined by RFC 1498 [?].

4.2

Multi-Hop Protocols: Address-free and Name-based

ing does not generally follow the Internet’s endto-end principle [?]. In addition to packet forwarding, a node along a path can inspect received data and make local decisions with it, possibly transforming the data before forwarding it, or suppressing it. This in-network processing can greatly reduce communication while keeping higher-level semantic requirements. For example, when collecting a MAX query over a network (which returns the maximum sensor reading), nodes need only forward the highest reading they receive, and can suppress if they hear higher readings.

4.3

Power Management

Power management is fundamental in sensor networks and particularly challenging to abstract into a clean architectural concept. Most power aware networking has been dealt with at a single point in the stack in isolation. Some power aware MACs attempt to invisible turn off the radio [13]. Buffering transmissions can make power management appear as latency. Scheduled routing controls radio activity from a multihop level, assuming that no other communication services are involved. Most published work only deals with powering the radio, not the actual processor beneath these power-aware protocols: dropping from active to full sleep reduces power draw a hundred-fold, while shutting off the radio only cuts it in half. Long term sensor network deployments have approached this problem using vertically integrated power management algorithms. Although this has allowed lengthy deployments, the techniques and algorithms are suited to a particular application or deployment, and not reusable. In the Great Duck Island deployment, each TinyOS component include a standard power interface, allow to act and convey power transitions to other portions of the stack. The application level protocol goes into virtual sleep for long periods of time, the physical layer wakes periodically to sample for preambles, and the multihop routing layer between has an event-driven schedule. TinyDB adopts a related approach, but shuts down the whole stack between query processing epochs [7]. Ultimately, power management requires a holistic approach, so potentially multiple protocols

can establish a cooperative power schedule. Correspondingly, it represents an abstraction that cuts across many network layers. A network mesh table can, by watching local traffic, estimate whether neighboring nodes are awake or not. Once when nodes transmits a long preamble that wakes up a local neighborhood of nodes, those nodes can exchange normal length packets before returning to sleep. Assuming that nodes take turns (e.g., probablistically) transmitting this long wakeup packet, its cost can be amortized over the many following transmissions.

5.

CONCLUSION

6.

REFERENCES

[1] D. D. Clark. The design philosophy of the DARPA internet protocols. In SIGCOMM, pages 106–114, Stanford, CA, Aug. 1988. ACM. [2] R. Fonesca, D. Culler, S. Ratnasamy, S. Shenker, and I. Stoica. Beacon vector routing: Scalable point-to-point routing in wireless sensornets. In submission. [3] D. Ganesan, B. Krishnamachari, A. Woo, D. Culler, D. Estrin, and S. Wicker. An empirical study of epidemic algorithms in large scale multihop wireless networks. UCLA Computer Science Technical Report UCLA/CSD-TR 02-0013, 2002. [4] J. Hill, R. Szewczyk, A. Woo, S. Hollar, D. E. Culler, and K. S. J. Pister. System Architecture Directions for Networked Sensors. In Architectural Support for Programming Languages and Operating Systems, pages 93–104, 2000. TinyOS is available at http://webs.cs.berkeley.edu. [5] C. Intanagonwiwat, R. Govindan, and D. Estrin. Directed diffusion: a scalable and robust communication paradigm for sensor networks. In Proceedings of the International Conference on Mobile Computing and Networking, Aug. 2000. [6] B. Karp and H. T. Kung. GPSR: greedy perimeter stateless routing for wireless networks. In International Conference on Mobile Computing and Networking (MobiCom 2000), pages 243–254, Boston, MA, USA, 2000. [7] P. Levis, S. Madden, D. Gay, J. Polastre, R. Szewczyk, A. Woo, E. Brewer, and D. Culler. The emergence of networking abstractions and techniques in tinyos. In First USENIX/ACM Symposium on Network Systems Design and Implementation (NSDI), 2004. [8] P. Levis, N. Patel, D. Culler, and S. Shenker. Trickle: A self-regulating algorithm for code maintenance and propagation in wireless sensor networks. In First USENIX/ACM Symposium on Network Systems Design and Implementation (NSDI), 2004. [9] A. Mainwaring, J. Polastre, R. Szewczyk, D. Culler, and J. Anderson. Wireless Sensor Networks for Habitat Monitoring. In Proceedings of the ACM International Workshop on Wireless Sensor Networks and Applications, Sept. 2002. [10] J. Newsome and D. Song. Gem: graph embedding for routing and data-centric storage in sensor networks without geographic information. In Proceedings of the first international conference on Embedded networked sensor systems, pages 76–88. ACM Press, 2003.

[11] J. Polastre, J. Hill, and D. Culler. Versatile low power media access for wireless sensor networks. In submission. [12] S. Ratnasamy, B. Karp, L. Yin, F. Yu, D. Estrin, R. Govindan, and S. Shenker. Ght: a geographic hash table for data-centric storage. In Proceedings of the first ACM international workshop on Wireless sensor networks and applications, pages 78–87. ACM Press, 2002. [13] S. Singh and C. S. Raghavendra. Pamas: power aware multi-access protocol with signalling for ad hoc networks. SIGCOMM Comput. Commun. Rev., 28(3):5–26, 1998. [14] W. Ye, J. Heidemann, and D. Estrin. An energy-efficient mac protocol for wireless sensor networks. In In Proceedings of the 21st International Annual Joint Conference of the IEEE Computer and Communications Societies (INFOCOM 2002), New York, NY, June 2002. [15] J. Zhao and R. Govindan. Understanding packet delivery performance in dense wireless sensor networks. In Proceedings of the First International Conference on Embedded Network Sensor Systems, 2003.