Energy and QoS Aware Routing in Wireless Sensor Networks

Cluster Computing 8, 179–188, 2005C© 2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.

Energy and QoS Aware Routing in Wireless Sensor Networks

KEMAL AKKAYA * and MOHAMED YOUNISDepartment of Computer Science and Electrical Engineering, University of Maryland, Baltimore County, Baltimore, MD 21250

Abstract. Many new routing protocols have been proposed for wireless sensor networks in recent years. Almost all of the routing protocolsconsidered energy efficiency as the ultimate objective since energy is a very scarce resource for sensor nodes. However, the introductionimaging sensors has posed additional challenges. Transmission of imaging data requires both energy and QoS aware routing in order toensure efficient usage of the sensors and effective access to the gathered measurements. In this paper, we propose an energy-aware QoSrouting protocol for sensor networks which can also run efficiently with best-effort traffic. The protocol finds a least-cost, delay-constrainedpath for real-time data in terms of link cost that captures nodes’ energy reserve, transmission energy, error rate and other communicationparameters. Moreover, the throughput for non-real-time data is maximized by adjusting the service rate for both real-time and non-real-timedata at the sensor nodes. Such adjustment of service rate is done by using two different mechanisms. Simulation results have demonstratedthe effectiveness of our approach for different metrics with respect to the baseline approach where same link cost function is used withoutany service differentiation mechanism.

Keywords: sensor networks, QoS routing, energy-aware routing, real-time traffic

1. Introduction

Recent advances in micro-electro-mechanical systems(MEMS) and low power and highly integrated digital elec-tronics have led to the development of micro sensors[1,9,17,21,22,24,27]. Such sensors are generally equippedwith data processing and communication capabilities. Thesensing circuitry measures ambient conditions related to theenvironment surrounding the sensor and transforms them intoan electric signal. Processing such a signal reveals some prop-erties about objects located and/or events happening in thevicinity of the sensor. The sensor sends such sensed data, usu-ally via radio transmitter, to a command center either directlyor through a data concentration center (a gateway). The gate-way can perform fusion of the sensed data in order to filter outerroneous data and anomalies and to draw conclusions fromthe reported data over a period of time.

The continuous decrease in the size and cost of sensors hasmotivated intensive research in the past few years addressingthe potential of collaboration among sensors in data gatheringand processing via an ad hoc wireless network. Networkingunattended sensor nodes is expected to have significant im-pact on the efficiency of many military and civil applications,such as combat field surveillance, security and disaster man-agement. A network of sensors can be used to gather mete-orological variables such as temperature and pressure. Thesemeasurements can be used in preparing forecasts or detectingharsh natural phenomena. In disaster management situationssuch as earthquakes, sensor networks can be used to selec-tively map the affected regions directing emergency responseunits to survivors. In military situations, sensor networks can

* Corresponding author.E-mail: [email protected]

be used in surveillance missions and can be used to detectmoving targets, chemical gases, or presence of micro-agents.

However, sensor nodes are constrained in energy supplyand bandwidth. Such constraints combined with a typical de-ployment of large number of sensor nodes have necessitatedenergy-awareness at the layers of networking protocol stackincluding network layer. Routing of sensor data has beenone of the challenging areas in wireless sensor network re-search. Current research on routing in wireless sensor net-works mostly focused on protocols that are energy aware tomaximize the lifetime of the network, scalable for large num-ber of sensor nodes and tolerant to sensor damage and batteryexhaustion [9,13–15,25,27,30]. Since the data they deal withis not in large amounts and flow in low rates to the sink, theconcepts of latency, throughput and delay were not primaryconcerns in most of the published work on sensor networks.However, the introduction of imaging sensors has posed addi-tional challenges for routing in sensor networks. Transmissionof imaging data requires careful handling in order to ensurethat end-to-end delay is within acceptable range. Such per-formance metrics are usually referred to as quality of service(QoS) of the communication network. Therefore, collectingsensed imaging data requires both energy and QoS aware rout-ing in order to ensure efficient usage of the sensors and effec-tive access to the gathered measurements.

QoS protocols in sensor networks have several applica-tions including real time target tracking in battle environ-ments, emergent event triggering in monitoring applicationsetc. Consider the following scenario: In a battle environmentit is crucial to locate, detect and identify a target. In order toidentify a target, we should employ imaging sensors. Afterlocating and detecting the target without the need of imagingsensors, we can turn on those sensors to get for instance an im-age of the target periodically for sending to the base station or

180 AKKAYA AND YOUNIS

gateway. Since, it is a battle environment; this requires a real-time data exchange between sensors and controller in order totake the proper actions. However, we should deal with real-time data, which requires certain bandwidth with minimumpossible delay. In that case, a service differentiation mecha-nism is needed in order to guarantee the reliable delivery ofthe real-time data.

Energy-aware QoS routing in sensor networks will ensureguaranteed bandwidth (or delay) through the duration of aconnection as well as providing the use of the most energyefficient path. To the best of our knowledge, no previous re-search has addressed QoS routing in sensor networks. In thispaper, we present an energy-aware QoS routing mechanismfor wireless sensor networks. Our proposed protocol extendsthe routing approach in [30] and considers only end-to-enddelay. The protocol looks for a delay-constrained path withthe least possible cost. The cost function which captures re-maining and transmission energy and error rate, is defined foreach link. Alternative paths with bigger costs are tried un-til one, which meets the end-to-end delay requirement andmaximizes the throughput for best effort traffic is found. Ourprotocol does not introduce any extra overhead to the sensors.

In the balance of this section we describe the sensor net-work architecture that we consider and summarize the relatedwork. In Section 2, we analyze the complexity of the QoS rout-ing problem in sensor networks and describe our approach.Section 3 includes simulations and evaluations of the proto-col. Finally we conclude the paper in Section 4 and outlineour future research.

1.1. Sensor network architecture

A set of sensors is spread throughout an area of interest todetect and possibly track events/targets in this area. The sen-sors are battery-operated with diverse capabilities and typesand are empowered with limited data processing engines. Theavailability of imaging sensors is of particular interest due tothe quality of service constraints associated with data gener-ated by such sensors. The mission for these sensors is dynam-ically changing to serve the need of one or multiple commandnodes. Command nodes can be stationary or mobile. In a dis-aster management environment, coordination centers are typi-cal stationary command nodes, while paramedics, fire trucks,rescue vehicles and evacuation helicopters are examples ofmobile command nodes. A gateway node is a less energy-constrained node deployed in the physical proximity of sen-sors. The gateway is responsible for organizing the activitiesat sensor nodes to achieve a mission, fusing data collectedby sensor nodes, coordinating communication among sensornodes and interacting with command nodes. We are consider-ing both the gateway and sensor nodes as stationary. All thesensors are assumed to be within the communication range ofthe gateway node. The architecture is depicted in figure 1.

The sensor is assumed to be capable of operating in anactive mode or a low-power stand-by mode. The sensing andprocessing circuits can be powered on and off. In addition both

Figure 1. Three-tier sensor network architecture.

the radio transmitter and receiver can be independently turnedon and off and the transmission power can be programmed fora required range. It is also assumed that the sensor can act asa relay to forward data from another sensor. It is worth notingthat most of these capabilities are available on some of theadvanced sensors, e.g. the Acoustic Ballistic Module fromSenTech Inc. [6]. The gateway node is assumed to know itslocation, e.g. via the use of GPS.

The described system’s architecture raises many interestingissues such as mission-oriented sensor organization, networkmanagement, gateway to command node communication pro-tocol, support of QoS traffic generated by imaging sensors, etc.While many of these issues are studied in the context of wire-less networking research, the naturally resource constrainedsensor-based environment makes these technical issues un-traditional and challenging. For example energy efficiencyhas to be a core objective of the system design, a factor thathas not been considered for typical networks. In this paper,we only focus on the energy-aware and QoS routing of sen-sor data among the communicating nodes. While the gatewaywill take charge of sensor organization based on the missionand available energy in each sensor, we assume knowledgeof which sensors need to be active in signal processing, e.g.using the approaches presented in [1,3].

1.2. Related work

In traditional best-effort routing throughput and delay are themain concerns. There is no guarantee that a certain perfor-mance in throughput or delay will be ensured throughout theconnection. However, in some cases where real-time or mul-timedia data are involved in communication, some perfor-mance guarantees in certain metrics such as delay, bandwidthand delay jitter are needed. Such guarantees can be achievedby employing special mechanisms known as QoS routingprotocols.

While contemporary best-effort routing approaches ad-dress unconstrained traffic, QoS routing is usually performedthrough resource reservation in a connection-oriented com-munication in order to meet the QoS requirements for each

ENERGY AND QoS AWARE ROUTING IN WIRELESS SENSOR NETWORKS 181

individual connection. While many mechanisms have beenproposed for routing QoS constrained real-time multimediadata in wire-based networks [5,18,20,28,31], they cannot bedirectly applied to wireless networks due to inherent charac-teristics of wireless environments and limited resources, suchas bandwidth. Therefore, several new protocols have beenproposed for QoS routing in wireless ad-hoc networks takingthe dynamic nature (due to mobility of the nodes) of the net-work into account [4,19,23,26,32]. Some of these proposedprotocols consider the imprecise state information while de-termining the routes [4,23]. CEDAR is another QoS awareprotocol, which uses the idea of core nodes (dominating set)of the network while determining the paths [26]. Using routesfound through the network core, a QoS path can be easilyfound. However, if any node in the core is broken, it will costtoo much in terms of resource usage to reconstruct the core.Lin [19] and Zhu et al. [32] have proposed QoS routing proto-cols specifically designed for TDMA-based ad-hoc networks.Both protocols can build a QoS route from a source to des-tination with reserved bandwidth. The bandwidth calculationis done hop-by-hop in a distributed fashion.

Another protocol for wireless networks that includes somenotion of QoS in its routing decisions is the Sequential As-signment Routing (SAR) [27]. The SAR protocol creates treesrouted from one-hop neighbor of the sink by taking the QoSmetric, the energy resource on each path and the priority levelof each packet into consideration. By using created trees, mul-tiple paths from sink to sensors are formed. One of these pathsis selected according to the energy resources and achievableQoS on each path. In our approach, we not only select a pathfrom a list of candidate paths that meet the end-to-end delayrequirement, but maximize the throughput for best effort traf-fic as well. In addition, the SAR approach suffers the overheadof maintaining the node states at each sensor node and main-taining the multiple paths from each node to the sink. Ourprotocol does not require sensor’s involvement in route setup.

Most of the QoS routing algorithms discussed in this sec-tion are based on the mobility of the nodes and none of themconsider energy awareness along with the QoS parameters.Although they are well suited to mobile ad hoc networks,the emerging complexity from mobility in such routing algo-rithms will be an over-kill for the systems where nodes arenot mobile and have limited resources, such as bandwidth andenergy. On the other hand, routing protocols proposed specif-ically for wireless sensor networks are designed according tothe needs of sensor networks, none of them considers any QoSor service differentiation mechanism in order to handle chal-lenges posed by imaging sensors and real-time applications ofsensor networks. Our proposed approach tackles these chal-lenges into account so that both the system lifetime will bemaximized and QoS requirements are met.

2. Energy-aware QoS routing

Our aim is to find an optimal path to the gateway in terms ofenergy consumption and error rate while meeting the end-to-

end delay requirements. End-to-end delay requirements areassociated only with the real-time data. Note that, in this casewe have both real-time and non-real-time traffic coexisting inthe network, which makes the problem more complex. We notonly should find paths that meet the requirements for real-timetraffic, but need to maximize the throughput for non-real timetraffic as well. This is because most of the critical applicationssuch as battlefield surveillance have to receive for instanceacoustic data regularly in order not to miss targets. Thereforeit is important to prevent the real-time traffic from consumingthe bulk of network bandwidth and leave non-real-time datastarving and thus incurring large amount of delay.

The described QoS routing problem is very similar to typ-ical path constrained path optimization (PCPO) problems,which are proved to be NP-complete [10]. We are trying to findleast-cost path, which meets the end-to-end delay path con-straint. However, in our case there is an extra goal, which isbasically to maximize the throughput of non-real-time traffic.Our approach is based on associating a cost function for eachlink and used a K least cost path algorithm to find a set of can-didate routes. Such routes are checked against the end-to-endconstraints and the one that provides maximum throughput ispicked. Before explaining the details of proposed algorithm,we introduce the queuing model.

2.1. Queuing model

The queuing model is specifically designed for the case ofcoexistence of real-time and non-real-time traffic in each sen-sor node. The model we employ is inspired from class-basedqueuing [11]. We use different queues for the two differenttypes of traffic. Basically, we have real-time traffic and non-real-time (normal) traffic whose packets are labeled accord-ingly. On each node there is a classifier, which checks thetype of the incoming packet and sends it to the appropriatequeue. There is also a scheduler, which determines the orderof packets to be transmitted from the queues according to thebandwidth ratio “r” of each type of traffic on that link. Themodel is depicted in figure 2.

The bandwidth ratio r , is actually an initial value set bythe gateway and represents the amount of bandwidth to bededicated both to the real-time and non-real-time traffic on aparticular outgoing link. Moreover, both classes can borrowbandwidth from each other when one of the two types of trafficis non-existent or under the limit. As indicated in figure 3, thisr -value is also used to calculate the service rate of real-timeand non-real-time traffic on that particular node, with riµ and(1 − ri )µ being respectively the service rate for real-time andnon-real-time data on sensor node i .

Since the queuing delay depends on this r -value, we can-not calculate the end-to-end delay for a particular path with-out knowing the r -value. Therefore we should first find a listof candidate least-cost paths and then select one that meetsthe end-to-end delay requirement. Our approach is based ona two-step strategy incorporating both link-based costs andend-to-end constraints. First we calculate the candidate paths


Figure 2. Queuing model in a particular sensor node.

Figure 3. Bandwidth sharing and service rates for a sensor node.

without considering the end-to-end delay. What we do is sim-ply calculate costs for each particular link and then use anextended version of Dijkstra’s algorithm to find an ascendingset of least cost paths. Once we obtain these candidate paths,we further check them to identify those that meet our end-to-end QoS requirements by trying to find an optimal r -value thatwill also maximize the throughput for non-real-time traffic.

2.2. Calculation of link costs

We consider the factors for the cost function on each partic-ular link separately except the end-to-end delay requirement,which should be for the whole path (i.e. all the links on thatpath). We define the following cost function for a link betweennodes i and j :

cos ti j =2∑

k=0

CFk = c0 × (disti j )l + c1 × f (energy j )

+ c2 × f (ei j )

where,

� disti j is the distance between the nodes i and j ,� f (energy j ) is the function for finding current residual en-

ergy of node j ,� f (ei j ) is the function for finding the error rate on the link

between i and j .

Hence, it’s not part of the cost function. Cost factors are de-fined as follows:

� CF0 (Communication Cost) = c0 × (disti j )l , where c0 isa weighting constant and the parameter l depends on theenvironment, and typically equals to 2. This factor reflectsthe cost of the wireless transmission power, which is di-rectly proportional to the distance raised to some power l.The closer a node to the destination, the less its cost factorCF0 and more attractive it is for routing.

� CF1 (Energy Stock) = c1 × f (energy j ). This factor re-flects the remaining battery lifetime (i.e. energy usagerate), which favors nodes with more energy. The moreenergy the node contains, the better it is for routing.

� CF2 (Error rate) = c2 × f (ei j ) where f is a function ofdistance between nodes i and j and buffer size on nodej (i.e. disti j/buffer size j ). The links with high error ratewill increase the cost function, thus will be avoided.

2.3. Estimation of end-to-end delay for a path

In order to find a QoS path for sending real-time data to thegateway, end-to-end delay requirement should be met. Beforeexplaining the computation of the delay for a particular pathP , we introduce the notation below:

λRT Real-time data generation rate for imagingsensors

riµ Service rate for real-time data on sensor node i(1 − ri )µ Service rate for non-real-time data on sensor

node ipi The number of sensing neighbors (data

generators) of node i on path Pqi The number of relaying neighbors (data

forwarders) of node i on path Pλ

(i)RT Real-time data rate on sensor node i

Q(i)RT Queuing delay on a node i for real-time traffic

TE End-to-end queuing delay for a particular pathP (ignoring propagation delay)

Tend-end End-to-end delay for a particular path PTrequired End-to-end delay requirement for all pathsm The number of nodes on path PNodes The set of all the sensing nodes that are part of

path P

We assume that the propagation delay is negligible. We alsoassume that all the imaging sensors have the same real-timedata generation rate λRT. Total real-time data rate by pi nodeswill be piλRT and total real-time data rate by qi nodes willbe added recursively for each relaying only node. Then totalreal-time data load on a sensor node is:

λ(i)RT = piλRT +

qi∑j=1

p jλ( j)RT

The average waiting time including the service time in thequeue in M/M/1 model is stated as W = 1

µ−λwhere µ is the

link transmission rate or service rate and λ is the packet arrival


rate [16]. Hence, total queuing delay (including the servicetime), Q(i)

RT on a node i is:

Q(i)RT = 1

riµ − λ(i)RT

(1)

We make an approximation to simplify the end-to-end queu-ing delay by assuming the incoming traffic to real-time andnon-real-time queues are stochastically independent. Thus,the end-to-end queuing delay for a particular path is:

TE =∑

∀i∈Path

Q(i)RT =

∑∀i∈Path

1

riµ − λ(i)RT

=∑

∀i∈Path

1

riµ − piλRT − ∑qi

j=1 p jλ( j)RT

Since we ignore the propagation delay, total end-to-end delaywill be:

Tend-end =∑

∀i∈Path

1

riµ − piλRT − ∑qi

j=1 p jλ( j)RT

(2)

2.4. Single-r mechanism

While we generate a formula for calculating the end-to-enddelay for a particular path, finding the optimal r -values foreach link as far as the queuing delay is concerned, will bevery difficult optimization problem to solve. Moreover, thedistribution of these r -values to each node is not an easy taskbecause each value should be unicasted to the proper sen-sor node rather than broadcasting it to all the sensors, whichmight bring a lot of overhead. Therefore, we follow an ap-proach, which will eliminate the overhead and complexity ofthe problem. Basically, we define each r -value to be same oneach link so that the optimization problem will be simple andthis unique r -value can be easily broadcasted to all the sensorsby the gateway.

If we let all r -values be same for every link then the formulawill be stated as:

Tend-end =∑

∀i∈Path

1

rµ − piλRT − ∑qi

j=1 p jλ( j)RT

Then the problem is stated as an optimization problem asfollows:

Max

( ∑∀i∈Path

((1 − r )µ)

)

subject to: Tend-end ≤ Trequired and 0 ≤ r < 1.

In order to find r -value from the above inequality ofTend-end ≤ Trequired, for simplification we consider finding anr -value which will satisfy the last hop node’s delay since thelast node will be getting the actual longest queuing delay. As aconsequence, the other nodes before the last node will alreadybe satisfied with that r -value and will use the same value. Wedivide Trequired into m equal time slots, where m is the numberof nodes on a particular path. The calculation of r for the last

Figure 4. Pseudo code for the proposed algorithm.

hop node m is as follows:

Trequired

m= Qm

RT = 1

rµ − λRT(

pm + ∑∀k∈Nodes pk

)r = m

µTrequired+ λRT

µ

(pm +

∑∀k∈Nodes

pk

)(3)

By considering the optimization problem above, we pro-pose the algorithm shown in figure 4, to find a least-cost path,which meets the constraints and maximizes the throughputfor non-real-time data. The algorithm calculates the cost foreach link, line 1 of figure 4, based on the cost function de-fined in Section 3.2. Then, for each node the least cost pathto the gateway is found by running Dijkstra’s shortest pathalgorithm in line 2. Between lines 5–15, appropriate r -valuesare calculated for paths from imaging sensors to the gateway.For each sensor node that has imaging capability, an r -valueis calculated on the current path (line 5). If that value is notbetween 0 and 1, extended Dijsktra algorithm for K-shortestpath is run in order to find alternative paths with bigger costs(line 9). K different least-cost paths are tried in order to finda proper r -value between 0 and 1 (lines 10–13). If there isno such r -value, the connection request of that node to thegateway is rejected.

The algorithm might generate different r -values for differ-ent paths. Since, the r -values are stored in a list; the maximumof them is selected to be used for the whole network (line 17).That r -value will satisfy the end-to-end delay requirement forall the paths established from imaging sensors to the gateway.

In order to find the K least cost paths (i.e. K shortest paths),we modified an extended version of Dijkstra’s algorithm givenin [8]. Since, the algorithm can suffer loops during execution;we modified the algorithm in such a way that each time a newpath is searched for a particular node; only node-disjoint pathsare considered during the process. This eliminates loops andensures simplicity and efficiency. This might also help findinga proper r -value easily since that node-disjoint path will not


inherit the congestion in the former path. Interested reader isreferred to [8] for further information.

2.5. Multi-r mechanism

Since the single-r mechanism is just an approximation to theoptimal solution of allocating r-values for each node by as-suming a unique r-value for each node, we extended the modelso that it will allow different r -values to be assigned to sensornodes for better resource allocation. In order to find differentr-values, each node’s r -value is calculated by setting max-imum allowable queuing delay for every node on the pathproportional to arrival rate of real-time traffic to that node.The least-cost path is picked. The gateway calculates a de-lay factor “�d” by dividing the value of the end-to-end delay“d” by the accumulative arrival rates of real-time traffic at allnodes on the path. The gateway then broadcasts the value of“�d” to all nodes on the path so that they can use it to derivetheir r -value.

Then �d will be calculated as follows:

�d = Trequired∑∀i∈Path

(pi + ∑qi

j=1 p j) ∗ λRT

(4)

Each sensor node i will calculate its r-value ri by using �das follows:

From [1],

�d = 1

riµ − λRT ∗ (pi + ∑qi

j=1 p j)

ri = 1

�d ∗ µ+ λRT

µ∗

(pi +

qi∑j=1

p j

)(5)

Then the problem will be to maximize the total throughput oneach particular path:

Max

( ∑∀i∈Path

(1 − ri )

)where 0 ≤ ri < 1.

3. Experimental results

The effectiveness of the energy-aware QoS routing approachis validated through simulation. This section describes the per-formance metrics, simulation environment, and experimentalresults.

3.1. Performance metrics

We have used the following metrics to capture the performanceof our QoS routing approach:

� Time to first node to die: When the first node runs out ofenergy, the network within the cluster is said to be par-titioned. The name network partitioning reflects the factthat some routes become invalid and cluster-wide rerout-ing may be immanent.

� Average lifetime of a node: This gives a good measure ofthe network lifetime. A routing algorithm, which maxi-mizes the lifetime of the network, is desirable. This met-ric also shows how efficient is the algorithm in energyconsumption.

� Average delay per packet: Defined as the average timea packet takes from a sensor node to the gateway. Mostenergy aware routing algorithms try to minimize the con-sumed energy. However, the applications that deal withreal-time data is delay sensitive, so this metric is impor-tant in our case.

� Network Throughput: Defined as the total number of datapackets received at the gateway divided by the simulationtime. The throughput for both real-time and non-real-timetraffic will be considered independently.

3.2. Environment setup

In the experiments we have considered a network of 100 ran-domly placed nodes in a 1000 × 1000 meter square area. Thegateway position is determined randomly within the bound-aries of deployment area. A free space propagation channelmodel is assumed [2] with the capacity set to 2 Mbps. Packetlengths are 10 Kbit for data packets and 2 Kbit for routingand refresh packets. Each node is assumed to have an initialenergy of 5 joules. The buffers for real-rime data and normaldata have default size of 20 packets [12]. A node is considerednon-functional if its energy level reaches 0. For the term CF1

in the cost function, we used the linear discharge curve of thealkaline battery [29].

For a node in the sensing state, packets are generated ata constant rate of 1 packet/sec. This value is consistent withthe specifications of the Acoustic Ballistic Module from Sen-Tech Inc. [6]. The real-time packet generation rate (λRT) forthe nodes, which have imaging capability is greater than thenormal rate. The default value is 6 packets/sec. A service rate(µ) of 20 packets/sec is assumed. Each data packet is time-stamped when it is generated to allow the calculation of aver-age delay per packet. In addition, each packet has an energyfield that is updated during the packet transmission to calculatethe average energy per packet since our cost function definedfor each link is using remaining energy as part of the cost.A packet drop probability is taken to be 0.01. This is used tomake the simulator more realistic and to simulate the deviationof the gateway energy model from the actual energy modelof nodes. We assume that the network is tasked with a target-tracking mission in the experiment. The initial set of sensingnodes is chosen to be the nodes on the convex hull of sen-sors in the deployment area. The set of sensing nodes changesas the target moves. Since targets are assumed to come fromoutside the area, the sensing circuitry of all boundary nodesis always turned on. The sensing circuitry of other nodes areusually turned off but can be turned on according to the targetmovement. We also assume that each sensor node is capableof taking the image of a target to identify it clearly and canturn on its imaging capability on demand. During simulation,


a small subset of current active nodes, which are the closestnodes to the target, are selected to turn on their imaging capa-bility. Therefore, the imaging sensor set may change with themovement of the target.

The packet-generation rate for imaging sensors is big-ger than the normal sensors; hence more packets are gen-erated when imaging sensors are employed [7]. These pack-ets are labeled as real-time packets and treated differentlyin sensor nodes. The r -value is initially assumed to be 0but it is recalculated as imaging sensors get activated. Tar-gets are assumed to start at a random position outside theconvex hull. Targets are characterized by having a constantspeed chosen uniformly from the range 4 meters/s to 6 me-ters/s and a constant direction chosen uniformly depend-ing on the initial target position in order for the target tocross the convex hull region. It is assumed that only onetarget is active at a time. This target remains active until itleaves the deployment region. In this case, a new target isgenerated.

3.3. Performance results

In this section, we present some performance results obtainedby the simulation. Different parameters are such as buffer size,packet drop probability and real-time data generation rates areconsidered in order to capture the effects on the performancemetrics defined earlier in this section.

Performance comparison of three different protocolsAs a baseline approach, we have used the same cost functionwith same routing algorithm (i.e. Dijkstra) without doing anyservice differentiation. That is, we have not differentiated be-tween packets and have used only one queue in each sensornode, which accommodates all kinds of packets. Therefore,no bandwidth sharing on any path is performed. We have com-pared this approach with our single-r and multi-r mechanismsby looking at the average delay per packet, average lifetime ofa node and time to first node to die. When we compare the av-erage delay per real-time packets generated in our model withthe average delay per packet generated in single queue model,we observed that both multi-r and single-r mechanisms haveless average delay (See figure 5). This is due to the priority

Figure 5. Average delay per packet with different real-time data rates.

Figure 6. Average lifetime of a node with different real-time data rates.

Figure 7. Time for first node to die with different real-time data rates.

given to real-time packets when transmitting to the gateway.On the other hand, multi-r mechanism performs better thansingle-r mechanism as expected. Because, every particularnode adjusts its r -value based on the resources it has. This ismore efficient than the single-r case in which a unique r -valueis imposed by the gateway for all the nodes. Furthermore, inall cases the average delay per packet increases for higher ratesand real-time data causes more queuing delay at each sensornode.

In figures 6 and 7, we have looked at the energy usage of theprotocols. The average lifetime of a node and the time for firstnode to die decreases when real-time data increases, causingthe nodes to sense and transmit more packets. Since the samecost function is used for all protocols, the lifetime of the nodesand the time for first node to die are very close to each other asconfirmed by figures 6 and 7. However, the energy usage of thesingle-r mechanism is slightly less than the others. This can beexplained by looking at the throughput. For the single-r mech-anism, sometimes an r -value for the whole network cannotbe found; causing the rejection of some connections. This de-creases the throughput hence fewer packets are relayed. This isnot the case for the baseline protocol. On the other hand, for themulti-r mechanism, it is easier to find an r -value for a particu-lar node. Furthermore, the efficiency in the usage of resourcesfor multi-r mechanism causes an increase in the throughputespecially for non-real-time data as seen in figure 8. Suchincrease incurs a little more energy consumption in the sensornodes.


Figure 8. Non-real-time data throughput for different real-time data rates.

Figure 9. Non-real-time packet delay for different real-time data rates.

Effect of real-time data rate on throughput and delayIn order to study the performance of the algorithm for dif-ferent real-time data rates, we ran simulation for differentvalues of real-time packet data rates. The results are depictedin figures 8 and 9. First, we have looked at the non-real-timedata throughput. While the number of real-time packets in-crease, it gets more difficult to satisfy increasing number ofQoS paths. Hence, this can cause rejection of paths or packetdrops for non-real-time data causing throughput for such datato decrease. However, such decrease is very less, becomingconstant after a certain point (See figure 8).

We restricted r -value to be strictly less than 1, causing thethroughput for non-real-time data ((1 − r )µ) to stay greaterthan 0. Hence, the algorithm does not sacrifice the throughputfor non-real-time data for the sake of real-time data. Multi-rmechanism has greater throughput than the single-r since theresources are handled more efficiently.

Figure 9 shows the effect of real-time data rate on aver-age delay per non-real-time packet. The delay increases withthe rate since packets incur more queuing delay and sharethe same amount of bandwidth. It is interesting to note thatthe average packet delay for non-real-time packets in thecase of multi-r mechanism is bigger than the single-r mech-anism. In multi-r mechanism, the increase in the throughputof non-real-time packets cause extra queuing delay on thenodes; leading non-real-time packets to have more end-to-enddelay.

Effect of end-to-end delay requirement and real-time dategeneration rate on r-valuesIn order to see how the algorithm behaves under stringent con-ditions, we varied the end-to-end delay requirement and moni-

Figure 10. Network r -value with different end-to-end delay values.

Figure 11. Network r -value with different real-time data rates.

tored how this change affects the network r -value. The resultsare depicted in figure 10. The network r -value goes downwhile the end-to-end delay requirement gets looser. Since thedelay is not too strict, the nodes will be able to meet the end-to-end delay requirement with a smaller r -value as expectedfrom equation (3). On the other hand, while we congest thenetwork with more real-time data packets by increasing thereal-time data generation rate, more bandwidth will be re-quired for real-time packets. This will cause the r -value toincrease so that each node can serve more real-time packets(See figure 11).

Effect of packet drop probability on delayand average lifetime of a nodeTo study the effect of packet drop probability on performance,we varied the probability of packet drop from 0.01 to 0.05. Theresults are depicted in figures 12 and 13. The average delayper packet decreases with the increasing probability. This canbe explained by noting that as the number of hops the packettraverse increases, the probability that it will be dropped in-creases. This means that the packets that arrive to the gatewayare most probable to take a small number of hops and thus in-curring less delay. As expected, the throughput decreases dueto lost packets. The average node lifetime increases since notall packets reach their destination and thus the node energy isconserved.

Effect of buffer size on delay and average lifetime of a nodeSince, the queuing model we employed uses buffers in eachnode and there is a limit on the size of those buffers, we varied


Figure 12. Average delay per RT packets for different packet drop probabil-ities.

Figure 13. Average lifetime of a node for different packet drop probabilities.

Figure 14. Average delay per RT packets for different buffer size.

the buffer size to see if this has any effect on the performanceof the algorithm. The results are shown in figures 14 and15. The average delay per packet increases with the buffersize since the throughput increases. Packets are not droppedwhen there is enough space in the buffers. This will increasethe number of packets arriving to the gateway. The packetsfrom far nodes will be also able to reach the gateway. Morepackets from far nodes mean more delay, which eventuallyincreases the average delay per packet. The increasing numberof packets arriving to the gateway will also increase the energy

Figure 15. Average lifetime of a node for different buffer size.

consumption by increasing the number of transmission andreception costs, therefore decreasing the average lifetime of anode.

It is worth noting that, for both delay and average lifetimemetrics, multi-r mechanism performs better because of themore efficient adjustment of the packet service rates on sensornodes as depicted in figures 12–15.

4. Conclusion

In this paper, we presented a new energy-aware QoS rout-ing protocol for sensor networks. The protocol finds QoSpaths for real-time data with certain end-to-end delay re-quirements. In order to support both best effort and real-timetraffic at the same time, a class-based queuing model is em-ployed. The queuing model allows service sharing for real-time and non-real-time traffic. A ratio r is defined as an initialvalue set by the gateway and is used to calculate the amountof bandwidth to be dedicated to the real-time and non-real-time traffic on a particular outgoing link. The selected queu-ing model for the protocol allows the throughput for normaldata not to diminish by utilizing that service rate on eachnode.

Two different mechanisms, namely single-r and multi-r ,for setting that service rate on each node are presented. Single-r mechanism sets a network wide r -value for every sensornode. In the multi-r mechanism, the gateway broadcasts thenecessary information to the sensor nodes in order for them tocalculate their own r -value. The effectiveness of the protocolfor both mechanisms is validated by simulation. Simulation re-sults have shown that our protocol consistently performs wellwith respect to QoS metrics, e.g. throughput and average de-lay, in comparison to a baseline non-QoS aware protocol thatuse the same link cost. The multi-r mechanism has providedbetter end-to-end delay for real-time packets with a slight in-crease in energy usage. It has also increased the throughput fornon-real-time data packets, which has extended the queuingdelay on the nodes causing an increase in the average delayper non-real-time packets.

While our proposed protocol fits a fixed gateway model,we plan on addressing issues related to the relocation and


mobility of the gateway under QoS traffic as a future work. Insuch cases, the frequent update of the position of the gatewayand the propagation of that information through the networkmay excessively drain the energy of nodes. We plan to extendto model in order to handle the overhead of mobility andtopology changes.

References

[1] I.F. Akyildiz et al., Wireless sensor networks: A survey, Computer Net-works 38 (2002) 393–422.

[2] J. Andersen et al., Propagation measurements and models for wire-less communications channels, IEEE Communications Magazine 33(1)(1995).

[3] A. Buczak and V. Jamalabad, Self-organization of a heterogeneous sen-sor network by genetic algorithms, in: Intelligent Engineering SystemsThrough Artificial Neural Networks, C.H. Dagli, et al. (eds.), Vol. 8(ASME Press, New York, 1998), pp. 259–264.

[4] S. Chen and K. Nahrstedt, Distributed quality-of-service routing in ad-hoc networks, IEEE Journal on Selected areas in Communications 17(8)(1999).

[5] E. Crowley et al., A framework for QoS based routing inthe Internet, Internet-draft, draft-ietf-qosr-framework-06.txt (Aug.1998).

[6] Data sheet for the Acoustic Ballistic Module, SenTech Inc.,http://www.sentech-acoustic.com/

[7] K. Danilidis et al., Real-time tracking of moving objects with an activecamera, Real-Time Imaging Journal 4(1) (1998) 3–20.

[8] Q.V.M. Ernesto et al., The K shortest paths problem, Research Report,CISUC (June 1998).

[9] D. Estrin et al., Next century challenges: Scalable coordination in sen-sor networks, in: Proc. of MobiCom’99, Seattle, Washington (Aug.1999).

[10] G. Feng et al., Performance evaluation of delay-constrained least-costrouting algorithms based on linear and nonlinear lagrange relaxation,in: Proc. of ICC’2002, New York (April 2002).

[11] S. Floyd and V. Jacobson, Link sharing and resource management mod-els for packet networks, IEEE/ACM Transactions on Networking 3(4)(1995) 365–386.

[12] M. Gerla et al., Wireless, mobile ad-hoc network routing, in: Proc. ofIEEE/ACM FOCUS’99, New Brunswick, NJ (May 1999).

[13] W. Heinzelman et al., Adaptive protocols for information disseminationin wireless sensor networks, in: Proc. of MobiCom’99, Seattle, WA(Aug. 1999).

[14] W. Heinzelman et al., Energy-efficient communication protocols forwireless microsensor networks, in: Hawaii International Conferenceon System Sciences (Jan. 2000).

[15] C. Intanagonwiwat et al., Directed diffusion: A scalable and robust com-munication paradigm for sensor networks, in: Proc. of MobiCom’00,Boston, MA (Aug. 2000).

[16] D. Julian et al., QoS and fairness constrained convex optimization ofresource allocation for wireless cellular and ad hoc networks, in: Proc.of IEEE INFOCOM, NY (June 2002).

[17] R.H. Katz et al., Mobile networking for smart dust, in: Proc. of Mobi-Com’99, Seattle, WA (Aug. 1999).

[18] W.C. Lee et al., Routing subject to quality of service con-straints integrated communication networks, IEEE Network (July/Aug.1995).

[19] C.R. Lin, On demand QoS routing in multihop mobile networks, IEICETransactions on Communications (July 2000).

[20] Q. Ma and P. Steenkiste, Quality-of-service routing with performanceguarantees, in: Proc. of the 4th Int.Workshop on Quality of Service (May1997).

[21] R. Min et al., Low power wireless sensor networks, in: Proc. of Int.Conference on VLSI Design, Bangalore, India (Jan. 2001).

[22] G.J. Pottie and W.J. Kaiser, Wireless integrated network sensors, Com-munications of the ACM 43(5) (2000) 51–58.

[23] R. Querin and A. Orda, QoS-based routing in networks with inaccurateinformation: Theory and algorithms, in: Proc. IEEE INFOCOM’97,Japan (1997) pp. 75–83.

[24] J.M. Rabaey et al., PicoRadio supports ad hoc ultra low power wirelessnetworking, IEEE Computer 33 (2000) 42–48.

[25] R. Shah and J. Rabaey, Energy aware routing for low energy Ad hocsensor networks, in: Proc. of IEEE Wireless Communications and Net-working Conference, Orlando, FL (March 2002).

[26] R. Sivakumar et al., Core extraction distributed ad hoc routing specifi-cation, IETF Internet draft draft-ietf-manet-cedar-spec-00.txt (1998).

[27] K. Sohrabi et al., Protocols for self-organization of a wireless sensornetwork, IEEE Personal Communications 7(5) (2000) 16–27.

[28] Z. Wang and J. Crowcraft, QoS-based routing for supporting resourcereservation, IEEE Journal on Selected Area of Communications (Sept.1996).

[29] M. Woo et al., Power aware routing in mobile ad hoc networks, in:Proc. of MobiCom’98, Dallas, TX (1998).

[30] M. Younis et al., Energy-aware routing in cluster-based sensor net-works, in: Proc. of MASCOTS2002, Fort Worth, TX (Oct. 2002).

[31] L. Zhang et al., RSVP: A New Resource ReServation Protocol, IEEENetwork (Sept. 1993).

[32] C. Zhu and M.S. Corson, QoS routing for mobile ad hoc networks, in:Proc. of IEEE INFOCOM (2002).

Kemal Akkaya received his B.S. degree in ComputerScience from Bilkent University, Ankara, Turkey in1997 and MS degree in Computer Science from Orta-dogu Technical University (ODTU), Ankara, Turkeyin 1999. He worked as a software developer at anautomation project of Siemens and World Bank inAnkara, Turkey in 2000. He is currently a Ph.D. can-didate at University of Maryland, Baltimore County(UMBC), Baltimore, MD. His research interests in-clude energy aware routing, security and quality ofservice issues in ad hoc wireless networks.E-mail: [email protected]

Mohamed F. Younis received B.S. degree in com-puter science and M.S. in engineering mathematicsfrom Alexandria University in Egypt in 1987 and1992, respectively. In 1996, he received his Ph.D.in computer science from New Jersey Institute ofTechnology. He is currently an assistant professorin the department of computer science and electri-cal engineering at the university of Maryland Bal-timore County (UMBC). Before joining UMBC, hewas with the Advanced Systems Technology Group,an Aerospace Electronic Systems R&D organizationof Honeywell International Inc. While at Honeywellhe led multiple projects for building integrated faulttolerant avionics, in which a novel architecture andan operating system were developed. This new tech-nology has been incorporated by Honeywell in multi-ple products and has received worldwide recognitionby both the research and the engineering commu-nities. He also participated in the development theRedundancy Management System, which is a keycomponent of the Vehicle and Mission Computerfor NASA’s X-33 space launch vehicle. Dr. Younis’technical interest includes network architectures andprotocols, embedded systems, fault tolerant comput-ing and distributed real-time systems. Dr. Younis hasfour granted and three pending patents. He served onmultiple technical committees and published over 40technical papers in refereed conferences and journals.E-mail: [email protected]

Documents

Energy and QoS Aware Routing in Wireless Sensor Networks