On Modeling Wireless Sensor Networks - Semantic Scholar
On Modeling Wireless Sensor Networks Denis Graˇcanin Mohamed Eltoweissy Virginia Tech Department of Computer Science Falls Church, VA 22043 {gracanin,toweissy}@vt.edu Abstract Most of the current research in wireless sensor networks (WSN, for short) is constraint driven and focuses on optimizing the use of limited resources (for example, power) at each sensor. While such constraints are important, there is a need for more general performance metrics describing the effectiveness of WSNs. There is also a need for a unified model that would enable comparison of different types of WSNs. We propose a new service-centric model that focuses on services provided by a WSN and their corresponding performance metrics. A WSN is modeled at different levels of abstraction. For each level, a set of services and a set of metrics are defined. A mapping between metrics at different levels relates high-level, mission-oriented metrics to low-level capability-oriented metrics. The proposed model consists of mission, network, region, sensor, and capability layers. Within each layer, four planes are identified, namely, communications, management, application, and generation learning. The proposed model provides a flexible, open framework for expressing and evaluating capabilities, functionalities, management, behavior, and evolution of a WSN. In addition, the proposed model provides a holistic approach to comparing WSNs and to measuring their effectiveness. The generation learning plane is unique in that it serves to extend the longevity of the network and to enhance the network effectiveness over time.
1. Introduction Wireless sensor networks consist of large numbers of sensors that act cooperatively to provide “usable chunks of predigested information rather than a confusing wash of number” [6]. A WSN provides refined information, i.e. it processes the raw data collected by individual sensors before presenting it to the user. This amounts to providing a service or a collection of services based on sensor capabilities and on the underlying communications infrastructure.
Stephan Olariu Ashraf Wadaa Old Dominion University Department of Computer Science Norfolk, VA 23529 {olariu,wadaa}@cs.odu.edu
Due to their vast array of applications, WSNs have been viewed from different perspectives and standpoints [10]. Given the sensor capabilities, the next step is to have an operating system running on a sensor node. Such system has very restricted available resources but still has to provide required functionality and support for application development [4]. Some of the examples include TinyOS [1] and EYES [2] projects. The emphasis on (refined) data leads to a view of WSNs as a distributed database (a data-centric view). An example is TinyDB [3], a query processing system for extracting information from a network of TinyOS sensors. TinyDB provides a simple query language to specify required data. Given the query, TinyDB collects that data from sensor nodes, filters it, aggregates it together, and routes it out. Interestingly, the distributed nature of WSNs and the corresponding service-centric view have many similarities with web services. Indeed, WSNs and web services are on “opposite” sides of the network applications spectrum; however, it is still feasible to provide WSN services within the framework of web services. The pervasive nature of WSNs, combined with the service-centric view, represent an excellent real-world example of pervasive computing where web services could be used to seamlessly access WSN services. This is, in fact, what motivated our service-centric view of WSNs. WSN performance traditionally has been based on power consumption [8]. While power is certainly one of the most important and limiting factors, there are other performance issues. In the service-centric view, a Quality of Service (QoS) and mission oriented performance evaluation may take precedence over power-based performance measures. Security issues and service availability are other important aspects [12]. The main contribution of this work is to provide a general service-centric WSN model that includes both the datacentric and the power-centric views as special cases. Such a unified model allows for comparison of different types of WSNs. It also provides a flexible, open framework for ex-
Proceedings of the 18th International Parallel and Distributed Processing Symposium (IPDPS’04)
0-7695-2132-0/04/$17.00 (C) 2004 IEEE
Inter-region communication
pressing and evaluating capabilities, functionalities, management, behavior, and evolution of a sensor network. The remainder of the paper is organized as follows. Section 2 describes the proposed WSN model. Section 3 discusses performance metrics based on the proposed model and their relationship to the power-centric and the datacentric models. A case study is presented in Section 4. Finally, Section 5 offers concluding remarks and outlines directions future research.
Intra-region communication sensor
mission network region sensor capability
2 Wireless sensor network model The most generic model for a WSN is based on the data gathering and communication capabilities of sensors. If each sensor has a unique address, then some available layered models such as OSI or DARPA [9] can be used, with some modifications, to model the WSN. However, uniqueness of addresses may not be feasible or even required. In that case, a model can be developed based on the following assumptions: • Initially, all the sensors have identical capabilities. • The sensors are anonymous: they lack unique identifiers (e.g. addresses). • Several sensors can create a region (group): anonymity of a sensor in a sensor network dictates the creation of regions. • Each sensor belongs to exactly one region: the identity of this region is the only identifier available to the sensor. • A region has an address (coordinates) that uniquely identifies the region; no two regions can have the same address. • Communication among regions is based on addresses. • Sensor synchronization is short-lived and group-based, where a group is loosely defined as the collection of sensors that collaborate to achieve a given task. A five layer WSN model is proposed based on these assumptions. Figure 1 shows a generic WSN deployment and how individual layers of the model map to the underlying WSN components. The sensors are deployed uniformly at random over the area of interest; once deployed, they self-organize into a network that is expected to work unattended. The network is divided into addressable regions. Each region contains a set of sensor nodes. An example of such an organization can be provided using a base station, or a sink, that serves as a center of a polar coordinate system. The distance between a sensor and the sink is determined based on the (quantified)
Figure 1. Outline of the proposed model base-station signal level, as measured by the sensor node. The (quantified) angle between a sensor and the sink (relative to a reference direction) is determined by directional transmission from the sink. As a result, the area covered with sensors is divided into regions. Each region is uniquely identified by its distance from the sink (an integer) and its angle (an integer). All sensors within the region have the same values for distance and angle [11]. The mission determines the overall functioning of the WSN. It describes a high-level goal of the WSN, i.e. what data is of interest and what types of services are needed. A set of services is provided in support of a given mission. A service includes data collection and processing from large areas of the WSN. Since individual sensors are identified only by their region, service-related activities within a region are considered to be atomic. The service can be decomposed into a set of services, each limited to a single region and involving all the active sensors in the region. The sender, requesting such service, may be in the same region or outside the region. Requesting, performing and replying to the service requires communication among sensors. Each sensor has a set of sensory/control capabilities, described by attributes with a specified range of values and a specified resolution. Importantly, a given mission only requires a subset of capabilities, referred to as the sensor configuration. Thus, the configuration of individual sensors is determined by the mission of the WSN. A change in the mission often requires a change in the sensor configuration. The relationship between a mission and a corresponding sensor configuration can be explored to define Quality of Service (QoS) for a WSN. The QoS at the mission level may be qualitative, or even fuzzy. However, QoS for the lower layers can be much better defined using quantitative measures [5]. A mapping between the mission level QoS and the capability level QoS, created as a composition of four mappings between adjacent layers, provides performance metrics and measures for all layers. Such mapping
Proceedings of the 18th International Parallel and Distributed Processing Symposium (IPDPS’04)
0-7695-2132-0/04/$17.00 (C) 2004 IEEE
must work in both directions, from mission to capability and from capability to mission. A more detailed description of each layer of the proposed model provides a foundation for discussion about performance metrics. Within each layer, four different sets of functionalities are available: application, communication, management, and generation learning. These sets do not constitute vertical segments in the model, instead they group together similar functions. This grouping can be used as a reference point for mappings between layers.
Remote user (mission control)
Field user (pervasive computing)
Figure 3. Using sensor network services
3 Performance metrics and mapping The mission of a WSN is defined by required services, duration, space coverage, and other service parameters like spatial, temporal, and value data resolution, etc. These requirements translate into capabilities of individual sensor nodes and specifications for the underlying communications infrastructure.
Generation learning Management Communications Application Mission
3.1 Layers
Network Region Sensor Capability
Figure 2. WSN layered model
The application set contains all the functions necessary for raw data collection and processing in support of WSN services. The communications set supports messaging and data exchange functions that maintain network connectivity. The management set includes housekeeping functions that handle service access, authorization and security in general, as well as re-configuring the sensor network based on changes in mission parameters. The generation learning set handles transfer of data and knowledge from the current to the next generation of sensors, providing a natural way to extend the lifespan of the WSN well beyond the lifespan of individual sensors. It also provides for enhanced network effectiveness over time. Figure 2 illustrates the proposed layered structure. There can be several WSN services provided at the same time, both for remote users (e.g. mission control) and field users (e.g. pervasive computing). Remote users access the services through the base station while field users may communicate directly with individual sensors, provided they have the same region coordinates, as determined by the base station (Figure 3). These services can be provided as regular web services with location-dependent capabilities. The location of a field user (the region coordinates) can be used to customize provided services.
A capability can be defined as an unforgeable data structure for a specific resource, specifying exactly the access rights that the holder of the capability has with respect to that resource. Distributed operating systems, like Amoeba [9], use this concept in connection with security and messaging. Each resource is represented by a numerical value within a defined range. The simplest example is power. The value of power is between zero and the maximum power value Pmax . The power capability can be on or off (standby). Figure 4 shows four possible scenarios. Scenario a) illustrates a situation where the sensor is constantly on until the power is spent at time t1 . Scenario b) illustrates a situation where the sensor is on until time t0 . Between times t0 , and t1 the sensor is in standby mode and the rate of power consumption is reduced. At time t1 the sensor is on again until its power is spent at time t2 . Scenario c) illustrates a situation where the sensor is constantly on standby until its power is spent at time t3 . A graph for any other scenario will be somewhere between the two extremes represented by scenarios a) and c) (the shaded area in the scenario c). Scenario d) illustrates a situation where a new sensor is replacing an existing sensor (generation learning). In this case the transition from the old to the new sensor (time period between tg and t3 ) introduces some potential ambiguity, especially if the old and the new sensors overlap in time, i.e. work in parallel. The semantics of the generation learning determines the nature of the transition process. This simple example can be generalized to include a number of resources (including power) that can be on or off. The set of values for these resources define an N -dimensional resource space. A subset of this space defines the capability range, a set of all feasible uses of the sensor.
Proceedings of the 18th International Parallel and Distributed Processing Symposium (IPDPS’04)
0-7695-2132-0/04/$17.00 (C) 2004 IEEE
The example featured in Figure 4 demonstrates how power-centric models can be included into the proposed service-centric model by focusing on the power as the single capability and then defining the mission solely in terms of power consumption. In general, a mission can be defined in terms of a subspace of the N -dimensional resource space, the mission range. Ideally, the mission range is a subset of the resource range. In general, only a part of the mission range will be included in the resource range. The percentage of the mission range covered by the resource range represents a simple QoS measure. The construction of the mission range is the result of a composite mapping based on mappings for each of the remaining four layers. Power Pmax
Power Pmax
a)
0
t1
Power Pmax
0
Time
t0
t2
Time
c)
t0
t1
t2
Time
t3
Power Pmax
0
t1
d)
t0
t1 tg t2
t3
• Turn a sensor device on and off: in case of power that means the sensor is on or standby, • Read a value from a device: periodically or on demand, • Write a value to a device: periodically or on demand, • Store a value (together with time-stamp) in memory: depends on memory size, • Perform computation on values (and time-stamps) in an array: min, max, average, sum, count, etc., • Receive and send messages.
b)
0
lifetime of a sensor network thus further de-emphasizing power-based optimization. A sensor must support the following functionality:
Time
Figure 4. Power as a simple example of capability.
In case of a time-constrained mission, the maximum value for time is specified and creates a finite mission range (the resource range is finite due to a finite power supply). A QoS measure stipulates, given a time interval, the percentage of the mission range covered, during the time interval, by the corresponding slice of the resource range. In a general case, the QoS provides the amount of time during which the WSN can provide the desired percentage of coverage. In general, the relationship between resources is a complex one and captures the physical properties of sensors. In a simplest case, the constraints are linear and the resource range is a convex polyhedron in an N -dimensional space. In that case, linear or non-linear programming techniques can be used to determine a global optimum for the mission objective, as seen by an individual sensor. The presence of many sensors with unknown coordinates and distribution over regions makes this much more difficult. That also provides a way to move further away from the power-based optimization and to consider a coverage density, probabilities (random sensor distribution) and region longevity. The generation learning capabilities may significantly extend the
There are three types of messages that a sensor can handle: • Service request/response that originates outside the WSN: a protocol is required (e.g. a bully algorithm) to determine which sensor(s) will perform the service. • Service request/response that originates from a sensor processing the service request and is addressed to a set of regions. A protocol is required (e.g. a bully algorithm) to determine which sensor will perform the operation. • Service processing: region-based communications are used to exchange data. A region represents a collection of sensor nodes with identical coordinates. As a consequence, region-level services are based on the sensor-level services.
3.2 Functionality sets One of the conceptual differences between the proposed service-centric model and, for example, the OSI model is that layering is done based on the structural hierarchy, going from an individual sensing capability within a sensor node, region, network to the complete mission. The functional components used as layers in the OSI model are spread across each layer in the proposed model. As a consequence, there is much more flexibility in the way communication and management functions can be implemented. From the user’s point of view, accessing a sensor network involves issuing a mission service request, either from a remote site or co-located with the sensor network. In either case, the service is accepted using the application functionality set. Using the communication set, a sensor node is elected that coordinates service processing. The management set provides necessary authentication and authorization features,
Proceedings of the 18th International Parallel and Distributed Processing Symposium (IPDPS’04)
0-7695-2132-0/04/$17.00 (C) 2004 IEEE
as well as coordination of multiple service requests at the same time. If the service requires an extensive time period to be processed (beyond the lifespan of an individual sensor) a generation learning set is involved. This approach makes the networking aspects of the system transparent to the user. Data-centric models are subsumed by the proposed model within application and communication functionality sets (with some management features). For example, Directed Diffusion (DD) is a data-centric protocol where nodes are addressed by the type of data they sense [7]. In the proposed model the DD is represented through the regionbased addressing and services that request specific data. Generation learning is motivated by, and a direct consequence of, the limited lifespan of sensor nodes. Sensors, regions, and the whole WSN evolve, accumulating knowledge learned and transmitted from generation to generation. Obviously, the transition from one generation to the next should be effected within a time period significantly smaller than the lifespan of individual sensors. The WSN is then in two distinct states: steady state and transition. From an information theoretic point of view, the WSN can be described by entropy as follows: at the beginning of the steady state, entropy is at its lowest level. Over time, the entropy grows and once it exceeds certain threshold, a generation change starts. A part of information is transferred to next generation and the cycle starts again. Referring to Figure 5, a consequence of generation learning is that the starting entropy of a cycle grows over time, thus providing a measure for learning capabilities of the WSN.
Entropy
4 Case study We illustrate some of the aspects of the proposed servicecentric model for WSN by a simple case study. Assume that the WSN is deployed to measure ambient temperature. Each sensor has a temperature sensing capability specified by its temperature range, resolution, and sampling frequency. Individual sensors can measure temperatures in the 0–100F range with a resolution of 0.1F. Temperature measurements are taken every minute. Each sensor can store in its local memory the last 100 readings. A service request at the mission level is represented in a high-level or fuzzy way. Consequently, a fuzzy based approach is a viable option. In general, for each of the capabilities at the sensor level, a fuzzy variable is defined with a set of corresponding linguistic values and modifiers. An example of a service request is: WAS SOMEWHERE THE TEMPERATURE VERY HIGH LATELY? In this example, TEMPERATURE is a fuzzy variable, HIGH is a linguistic value and VERY is a linguistic modifier. Time is also considered a fuzzy variable. The difference is that range may be unlimited so most of the linguistic values are defined relative to the current time, like LATELY in this case. The fuzzy values are defined in terms of membership functions. A membership function has value between 0 and 1 over an interval of crisp values. Most of the mission requests are general in nature and should be broadcast to all regions. In this case a fuzzy service request is mapped to an SQL-like query. A defuzzification takes place so HIGH TEMPERATURE becomes > 90 and LATELY becomes > 10 minutes. The service request becomes FIND REGION WHERE TEMPERATURE > 90 AND TIME > 10
0
tg
2tg
3tg
4tg
Time
Figure 5. Entropy change in generation learning
The entropy change over time is also affected by the management functionality at the sensor level. A sensor can function in slave mode so it always responds to a request. The active mode allows the sensor to selectively respond and thus to some extent control the rate of entropy increase.
Such requests may be broadcast by the base station or, alternatively, a sensor can receive the request and can forward it to the base station which will then broadcast it. The de-fuzzification process is performed outside the sensor network. However, if the sensors have fuzzy processing capabilities, such processing can be embedded into the network. The network service request will be received in all regions. One of the sensors in the region becomes the managing node for that service. All other sensor nodes in the region automatically answer the second part of the network service query. At the region level, all nodes are processing the region service request: TEMPERATURE > 90 AND TIME > 10
Proceedings of the 18th International Parallel and Distributed Processing Symposium (IPDPS’04)
0-7695-2132-0/04/$17.00 (C) 2004 IEEE
At the sensor level, each sensor identifies temperature capability and its value 90 and the time value of 10. At the capability level, the current value and all values in the memory with a timestamp difference of 10 minutes less from the requests timestamp are checked. If all of them are larger than 90 (capability level), then this sensor determines that the request can be served (sensor level) and sends a message (region level). The coordinating node, waits for such messages within a given time period (timeout). If a message is received, it sends a message indicating region coordinates to the base station (network level). The base station waits for such messages within a given time period (timeout). It collects all region coordinates and then sends a (possibly empty) list of regions. This list of regions is a response to the original mission-level service request.
5 Conclusions and open problems
The proposed WSN model provides a flexible, open framework in which various specific WSNs can be represented and modeled. The key benefit is a clear separation between the high-level mission of the WSN (mission services) and the low-level hardware specific capabilities of an individual sensor node (capability services). A mapping between mission and capability layers, created as a composition of mappings between intermediate layers, provides a formalism for a service-centric description and evaluation of the sensor network. A power-based evaluation is then represented as a service-based evaluation focusing on a single capability, namely power.
References [1] Berkeley university TinyOS web page, [accessed January 18, 2004]. http://today.cs.berkeley.edu/tos. [2] EYES project web page, [accessed January 18, 2004]. http://eyes.eu.org. [3] TinyDB project web page, [accessed January 18, 2004]. http://telegraph.cs.berkeley.edu/tinydb/. [4] S. Dulman and P. Havinga. Operating system fundamentals for the EYES distributed sensor network. In Proceedings of the PROGRESS Workshop, Oct. 2002. [5] D. Graˇcanin, Y. Zhou, and L. DaSilva. Quality of service for networked virtual environments. IEEE Communications Magazine, 42(4), Apr. 2004. To appear. [6] G. T. Huang. Casting the wireless sensor net. Technology Review, 106(6):50–56, July/Aug. 2003. [7] C. Intanagonwiwat, R. Govindan, D. Estrin, J. Heidemann, and F. Silva. Directed diffusion for wireless sensor networking. IEEE/ACM Transactions on Networking, 11(1):2–16, Feb. 2003. [8] R. Muraleedharan and L. A. Osadciw. Balancing the performance of a sensor network using an ant system. In 37th Annual Conference on Information Sciences and Systems, 12–14 Mar. 2003. [9] A. S. Tanenbaum and M. van Steen. Distributed Systems: Principles and Paradigms. Prentice Hall, Upper Saddle River, New Jersey 07458, 2002. [10] S. Tilak, N. B. Abu-Ghazaleh, and W. Heinzelman. A taxonomy of wireless micro-sensor network models. SIGMOBILE Mobile Computing and Communications Review, 6(2):28– 36, Apr. 2002. [11] A. Wadaa, S. Olariu, L. Wilson, M. Eltoweissy, K. Jones, and P. Sundaram. Training a sensor network. Journal of Mobile Networks and Applications, 2004. To appear. [12] A. D. Wood and J. A. Stankovic. Denial of service in sensor networks. IEEE Computer, 35(10):54–62, Oct. 2002.
Ongoing research focuses on refining the proposed model and on the development of a simulator. Generation learning, combined with the support for mission change, allow for changes in available services. The effects of such changes can be studied using a formalism developed for each layer and the corresponding mapping between neighboring layers. The simulator will have a plug-in based architecture to allow dynamic configuration of the individual layers and functionality sets within a layer. The simulator will provide a virtual experimental test-bed with enhanced, virtual-reality based, visualization and presentation capabilities.
Acknowledgment: This work was supported, in part, by a grant from the Commonwealth of Virginia Technology Research Fund (SE 2001-01) through the Commonwealth Information Security Center.
Proceedings of the 18th International Parallel and Distributed Processing Symposium (IPDPS’04)
0-7695-2132-0/04/$17.00 (C) 2004 IEEE
1. Introduction Wireless sensor networks consist of large numbers of sensors that act cooperatively to provide “usable chunks of predigested information rather than a confusing wash of number” [6]. A WSN provides refined information, i.e. it processes the raw data collected by individual sensors before presenting it to the user. This amounts to providing a service or a collection of services based on sensor capabilities and on the underlying communications infrastructure.
Stephan Olariu Ashraf Wadaa Old Dominion University Department of Computer Science Norfolk, VA 23529 {olariu,wadaa}@cs.odu.edu
Due to their vast array of applications, WSNs have been viewed from different perspectives and standpoints [10]. Given the sensor capabilities, the next step is to have an operating system running on a sensor node. Such system has very restricted available resources but still has to provide required functionality and support for application development [4]. Some of the examples include TinyOS [1] and EYES [2] projects. The emphasis on (refined) data leads to a view of WSNs as a distributed database (a data-centric view). An example is TinyDB [3], a query processing system for extracting information from a network of TinyOS sensors. TinyDB provides a simple query language to specify required data. Given the query, TinyDB collects that data from sensor nodes, filters it, aggregates it together, and routes it out. Interestingly, the distributed nature of WSNs and the corresponding service-centric view have many similarities with web services. Indeed, WSNs and web services are on “opposite” sides of the network applications spectrum; however, it is still feasible to provide WSN services within the framework of web services. The pervasive nature of WSNs, combined with the service-centric view, represent an excellent real-world example of pervasive computing where web services could be used to seamlessly access WSN services. This is, in fact, what motivated our service-centric view of WSNs. WSN performance traditionally has been based on power consumption [8]. While power is certainly one of the most important and limiting factors, there are other performance issues. In the service-centric view, a Quality of Service (QoS) and mission oriented performance evaluation may take precedence over power-based performance measures. Security issues and service availability are other important aspects [12]. The main contribution of this work is to provide a general service-centric WSN model that includes both the datacentric and the power-centric views as special cases. Such a unified model allows for comparison of different types of WSNs. It also provides a flexible, open framework for ex-
Proceedings of the 18th International Parallel and Distributed Processing Symposium (IPDPS’04)
0-7695-2132-0/04/$17.00 (C) 2004 IEEE
Inter-region communication
pressing and evaluating capabilities, functionalities, management, behavior, and evolution of a sensor network. The remainder of the paper is organized as follows. Section 2 describes the proposed WSN model. Section 3 discusses performance metrics based on the proposed model and their relationship to the power-centric and the datacentric models. A case study is presented in Section 4. Finally, Section 5 offers concluding remarks and outlines directions future research.
Intra-region communication sensor
mission network region sensor capability
2 Wireless sensor network model The most generic model for a WSN is based on the data gathering and communication capabilities of sensors. If each sensor has a unique address, then some available layered models such as OSI or DARPA [9] can be used, with some modifications, to model the WSN. However, uniqueness of addresses may not be feasible or even required. In that case, a model can be developed based on the following assumptions: • Initially, all the sensors have identical capabilities. • The sensors are anonymous: they lack unique identifiers (e.g. addresses). • Several sensors can create a region (group): anonymity of a sensor in a sensor network dictates the creation of regions. • Each sensor belongs to exactly one region: the identity of this region is the only identifier available to the sensor. • A region has an address (coordinates) that uniquely identifies the region; no two regions can have the same address. • Communication among regions is based on addresses. • Sensor synchronization is short-lived and group-based, where a group is loosely defined as the collection of sensors that collaborate to achieve a given task. A five layer WSN model is proposed based on these assumptions. Figure 1 shows a generic WSN deployment and how individual layers of the model map to the underlying WSN components. The sensors are deployed uniformly at random over the area of interest; once deployed, they self-organize into a network that is expected to work unattended. The network is divided into addressable regions. Each region contains a set of sensor nodes. An example of such an organization can be provided using a base station, or a sink, that serves as a center of a polar coordinate system. The distance between a sensor and the sink is determined based on the (quantified)
Figure 1. Outline of the proposed model base-station signal level, as measured by the sensor node. The (quantified) angle between a sensor and the sink (relative to a reference direction) is determined by directional transmission from the sink. As a result, the area covered with sensors is divided into regions. Each region is uniquely identified by its distance from the sink (an integer) and its angle (an integer). All sensors within the region have the same values for distance and angle [11]. The mission determines the overall functioning of the WSN. It describes a high-level goal of the WSN, i.e. what data is of interest and what types of services are needed. A set of services is provided in support of a given mission. A service includes data collection and processing from large areas of the WSN. Since individual sensors are identified only by their region, service-related activities within a region are considered to be atomic. The service can be decomposed into a set of services, each limited to a single region and involving all the active sensors in the region. The sender, requesting such service, may be in the same region or outside the region. Requesting, performing and replying to the service requires communication among sensors. Each sensor has a set of sensory/control capabilities, described by attributes with a specified range of values and a specified resolution. Importantly, a given mission only requires a subset of capabilities, referred to as the sensor configuration. Thus, the configuration of individual sensors is determined by the mission of the WSN. A change in the mission often requires a change in the sensor configuration. The relationship between a mission and a corresponding sensor configuration can be explored to define Quality of Service (QoS) for a WSN. The QoS at the mission level may be qualitative, or even fuzzy. However, QoS for the lower layers can be much better defined using quantitative measures [5]. A mapping between the mission level QoS and the capability level QoS, created as a composition of four mappings between adjacent layers, provides performance metrics and measures for all layers. Such mapping
Proceedings of the 18th International Parallel and Distributed Processing Symposium (IPDPS’04)
0-7695-2132-0/04/$17.00 (C) 2004 IEEE
must work in both directions, from mission to capability and from capability to mission. A more detailed description of each layer of the proposed model provides a foundation for discussion about performance metrics. Within each layer, four different sets of functionalities are available: application, communication, management, and generation learning. These sets do not constitute vertical segments in the model, instead they group together similar functions. This grouping can be used as a reference point for mappings between layers.
Remote user (mission control)
Field user (pervasive computing)
Figure 3. Using sensor network services
3 Performance metrics and mapping The mission of a WSN is defined by required services, duration, space coverage, and other service parameters like spatial, temporal, and value data resolution, etc. These requirements translate into capabilities of individual sensor nodes and specifications for the underlying communications infrastructure.
Generation learning Management Communications Application Mission
3.1 Layers
Network Region Sensor Capability
Figure 2. WSN layered model
The application set contains all the functions necessary for raw data collection and processing in support of WSN services. The communications set supports messaging and data exchange functions that maintain network connectivity. The management set includes housekeeping functions that handle service access, authorization and security in general, as well as re-configuring the sensor network based on changes in mission parameters. The generation learning set handles transfer of data and knowledge from the current to the next generation of sensors, providing a natural way to extend the lifespan of the WSN well beyond the lifespan of individual sensors. It also provides for enhanced network effectiveness over time. Figure 2 illustrates the proposed layered structure. There can be several WSN services provided at the same time, both for remote users (e.g. mission control) and field users (e.g. pervasive computing). Remote users access the services through the base station while field users may communicate directly with individual sensors, provided they have the same region coordinates, as determined by the base station (Figure 3). These services can be provided as regular web services with location-dependent capabilities. The location of a field user (the region coordinates) can be used to customize provided services.
A capability can be defined as an unforgeable data structure for a specific resource, specifying exactly the access rights that the holder of the capability has with respect to that resource. Distributed operating systems, like Amoeba [9], use this concept in connection with security and messaging. Each resource is represented by a numerical value within a defined range. The simplest example is power. The value of power is between zero and the maximum power value Pmax . The power capability can be on or off (standby). Figure 4 shows four possible scenarios. Scenario a) illustrates a situation where the sensor is constantly on until the power is spent at time t1 . Scenario b) illustrates a situation where the sensor is on until time t0 . Between times t0 , and t1 the sensor is in standby mode and the rate of power consumption is reduced. At time t1 the sensor is on again until its power is spent at time t2 . Scenario c) illustrates a situation where the sensor is constantly on standby until its power is spent at time t3 . A graph for any other scenario will be somewhere between the two extremes represented by scenarios a) and c) (the shaded area in the scenario c). Scenario d) illustrates a situation where a new sensor is replacing an existing sensor (generation learning). In this case the transition from the old to the new sensor (time period between tg and t3 ) introduces some potential ambiguity, especially if the old and the new sensors overlap in time, i.e. work in parallel. The semantics of the generation learning determines the nature of the transition process. This simple example can be generalized to include a number of resources (including power) that can be on or off. The set of values for these resources define an N -dimensional resource space. A subset of this space defines the capability range, a set of all feasible uses of the sensor.
Proceedings of the 18th International Parallel and Distributed Processing Symposium (IPDPS’04)
0-7695-2132-0/04/$17.00 (C) 2004 IEEE
The example featured in Figure 4 demonstrates how power-centric models can be included into the proposed service-centric model by focusing on the power as the single capability and then defining the mission solely in terms of power consumption. In general, a mission can be defined in terms of a subspace of the N -dimensional resource space, the mission range. Ideally, the mission range is a subset of the resource range. In general, only a part of the mission range will be included in the resource range. The percentage of the mission range covered by the resource range represents a simple QoS measure. The construction of the mission range is the result of a composite mapping based on mappings for each of the remaining four layers. Power Pmax
Power Pmax
a)
0
t1
Power Pmax
0
Time
t0
t2
Time
c)
t0
t1
t2
Time
t3
Power Pmax
0
t1
d)
t0
t1 tg t2
t3
• Turn a sensor device on and off: in case of power that means the sensor is on or standby, • Read a value from a device: periodically or on demand, • Write a value to a device: periodically or on demand, • Store a value (together with time-stamp) in memory: depends on memory size, • Perform computation on values (and time-stamps) in an array: min, max, average, sum, count, etc., • Receive and send messages.
b)
0
lifetime of a sensor network thus further de-emphasizing power-based optimization. A sensor must support the following functionality:
Time
Figure 4. Power as a simple example of capability.
In case of a time-constrained mission, the maximum value for time is specified and creates a finite mission range (the resource range is finite due to a finite power supply). A QoS measure stipulates, given a time interval, the percentage of the mission range covered, during the time interval, by the corresponding slice of the resource range. In a general case, the QoS provides the amount of time during which the WSN can provide the desired percentage of coverage. In general, the relationship between resources is a complex one and captures the physical properties of sensors. In a simplest case, the constraints are linear and the resource range is a convex polyhedron in an N -dimensional space. In that case, linear or non-linear programming techniques can be used to determine a global optimum for the mission objective, as seen by an individual sensor. The presence of many sensors with unknown coordinates and distribution over regions makes this much more difficult. That also provides a way to move further away from the power-based optimization and to consider a coverage density, probabilities (random sensor distribution) and region longevity. The generation learning capabilities may significantly extend the
There are three types of messages that a sensor can handle: • Service request/response that originates outside the WSN: a protocol is required (e.g. a bully algorithm) to determine which sensor(s) will perform the service. • Service request/response that originates from a sensor processing the service request and is addressed to a set of regions. A protocol is required (e.g. a bully algorithm) to determine which sensor will perform the operation. • Service processing: region-based communications are used to exchange data. A region represents a collection of sensor nodes with identical coordinates. As a consequence, region-level services are based on the sensor-level services.
3.2 Functionality sets One of the conceptual differences between the proposed service-centric model and, for example, the OSI model is that layering is done based on the structural hierarchy, going from an individual sensing capability within a sensor node, region, network to the complete mission. The functional components used as layers in the OSI model are spread across each layer in the proposed model. As a consequence, there is much more flexibility in the way communication and management functions can be implemented. From the user’s point of view, accessing a sensor network involves issuing a mission service request, either from a remote site or co-located with the sensor network. In either case, the service is accepted using the application functionality set. Using the communication set, a sensor node is elected that coordinates service processing. The management set provides necessary authentication and authorization features,
Proceedings of the 18th International Parallel and Distributed Processing Symposium (IPDPS’04)
0-7695-2132-0/04/$17.00 (C) 2004 IEEE
as well as coordination of multiple service requests at the same time. If the service requires an extensive time period to be processed (beyond the lifespan of an individual sensor) a generation learning set is involved. This approach makes the networking aspects of the system transparent to the user. Data-centric models are subsumed by the proposed model within application and communication functionality sets (with some management features). For example, Directed Diffusion (DD) is a data-centric protocol where nodes are addressed by the type of data they sense [7]. In the proposed model the DD is represented through the regionbased addressing and services that request specific data. Generation learning is motivated by, and a direct consequence of, the limited lifespan of sensor nodes. Sensors, regions, and the whole WSN evolve, accumulating knowledge learned and transmitted from generation to generation. Obviously, the transition from one generation to the next should be effected within a time period significantly smaller than the lifespan of individual sensors. The WSN is then in two distinct states: steady state and transition. From an information theoretic point of view, the WSN can be described by entropy as follows: at the beginning of the steady state, entropy is at its lowest level. Over time, the entropy grows and once it exceeds certain threshold, a generation change starts. A part of information is transferred to next generation and the cycle starts again. Referring to Figure 5, a consequence of generation learning is that the starting entropy of a cycle grows over time, thus providing a measure for learning capabilities of the WSN.
Entropy
4 Case study We illustrate some of the aspects of the proposed servicecentric model for WSN by a simple case study. Assume that the WSN is deployed to measure ambient temperature. Each sensor has a temperature sensing capability specified by its temperature range, resolution, and sampling frequency. Individual sensors can measure temperatures in the 0–100F range with a resolution of 0.1F. Temperature measurements are taken every minute. Each sensor can store in its local memory the last 100 readings. A service request at the mission level is represented in a high-level or fuzzy way. Consequently, a fuzzy based approach is a viable option. In general, for each of the capabilities at the sensor level, a fuzzy variable is defined with a set of corresponding linguistic values and modifiers. An example of a service request is: WAS SOMEWHERE THE TEMPERATURE VERY HIGH LATELY? In this example, TEMPERATURE is a fuzzy variable, HIGH is a linguistic value and VERY is a linguistic modifier. Time is also considered a fuzzy variable. The difference is that range may be unlimited so most of the linguistic values are defined relative to the current time, like LATELY in this case. The fuzzy values are defined in terms of membership functions. A membership function has value between 0 and 1 over an interval of crisp values. Most of the mission requests are general in nature and should be broadcast to all regions. In this case a fuzzy service request is mapped to an SQL-like query. A defuzzification takes place so HIGH TEMPERATURE becomes > 90 and LATELY becomes > 10 minutes. The service request becomes FIND REGION WHERE TEMPERATURE > 90 AND TIME > 10
0
tg
2tg
3tg
4tg
Time
Figure 5. Entropy change in generation learning
The entropy change over time is also affected by the management functionality at the sensor level. A sensor can function in slave mode so it always responds to a request. The active mode allows the sensor to selectively respond and thus to some extent control the rate of entropy increase.
Such requests may be broadcast by the base station or, alternatively, a sensor can receive the request and can forward it to the base station which will then broadcast it. The de-fuzzification process is performed outside the sensor network. However, if the sensors have fuzzy processing capabilities, such processing can be embedded into the network. The network service request will be received in all regions. One of the sensors in the region becomes the managing node for that service. All other sensor nodes in the region automatically answer the second part of the network service query. At the region level, all nodes are processing the region service request: TEMPERATURE > 90 AND TIME > 10
Proceedings of the 18th International Parallel and Distributed Processing Symposium (IPDPS’04)
0-7695-2132-0/04/$17.00 (C) 2004 IEEE
At the sensor level, each sensor identifies temperature capability and its value 90 and the time value of 10. At the capability level, the current value and all values in the memory with a timestamp difference of 10 minutes less from the requests timestamp are checked. If all of them are larger than 90 (capability level), then this sensor determines that the request can be served (sensor level) and sends a message (region level). The coordinating node, waits for such messages within a given time period (timeout). If a message is received, it sends a message indicating region coordinates to the base station (network level). The base station waits for such messages within a given time period (timeout). It collects all region coordinates and then sends a (possibly empty) list of regions. This list of regions is a response to the original mission-level service request.
5 Conclusions and open problems
The proposed WSN model provides a flexible, open framework in which various specific WSNs can be represented and modeled. The key benefit is a clear separation between the high-level mission of the WSN (mission services) and the low-level hardware specific capabilities of an individual sensor node (capability services). A mapping between mission and capability layers, created as a composition of mappings between intermediate layers, provides a formalism for a service-centric description and evaluation of the sensor network. A power-based evaluation is then represented as a service-based evaluation focusing on a single capability, namely power.
References [1] Berkeley university TinyOS web page, [accessed January 18, 2004]. http://today.cs.berkeley.edu/tos. [2] EYES project web page, [accessed January 18, 2004]. http://eyes.eu.org. [3] TinyDB project web page, [accessed January 18, 2004]. http://telegraph.cs.berkeley.edu/tinydb/. [4] S. Dulman and P. Havinga. Operating system fundamentals for the EYES distributed sensor network. In Proceedings of the PROGRESS Workshop, Oct. 2002. [5] D. Graˇcanin, Y. Zhou, and L. DaSilva. Quality of service for networked virtual environments. IEEE Communications Magazine, 42(4), Apr. 2004. To appear. [6] G. T. Huang. Casting the wireless sensor net. Technology Review, 106(6):50–56, July/Aug. 2003. [7] C. Intanagonwiwat, R. Govindan, D. Estrin, J. Heidemann, and F. Silva. Directed diffusion for wireless sensor networking. IEEE/ACM Transactions on Networking, 11(1):2–16, Feb. 2003. [8] R. Muraleedharan and L. A. Osadciw. Balancing the performance of a sensor network using an ant system. In 37th Annual Conference on Information Sciences and Systems, 12–14 Mar. 2003. [9] A. S. Tanenbaum and M. van Steen. Distributed Systems: Principles and Paradigms. Prentice Hall, Upper Saddle River, New Jersey 07458, 2002. [10] S. Tilak, N. B. Abu-Ghazaleh, and W. Heinzelman. A taxonomy of wireless micro-sensor network models. SIGMOBILE Mobile Computing and Communications Review, 6(2):28– 36, Apr. 2002. [11] A. Wadaa, S. Olariu, L. Wilson, M. Eltoweissy, K. Jones, and P. Sundaram. Training a sensor network. Journal of Mobile Networks and Applications, 2004. To appear. [12] A. D. Wood and J. A. Stankovic. Denial of service in sensor networks. IEEE Computer, 35(10):54–62, Oct. 2002.
Ongoing research focuses on refining the proposed model and on the development of a simulator. Generation learning, combined with the support for mission change, allow for changes in available services. The effects of such changes can be studied using a formalism developed for each layer and the corresponding mapping between neighboring layers. The simulator will have a plug-in based architecture to allow dynamic configuration of the individual layers and functionality sets within a layer. The simulator will provide a virtual experimental test-bed with enhanced, virtual-reality based, visualization and presentation capabilities.
Acknowledgment: This work was supported, in part, by a grant from the Commonwealth of Virginia Technology Research Fund (SE 2001-01) through the Commonwealth Information Security Center.
Proceedings of the 18th International Parallel and Distributed Processing Symposium (IPDPS’04)
0-7695-2132-0/04/$17.00 (C) 2004 IEEE
