Sivram-base Paper IEEE

8/8/2019 Sivram-base Paper IEEE

1/8

2328 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 5, MAY 2009

Performance Evaluation of Efficient and ReliableRouting Protocols for Fixed-Power Sensor Networks

Peter Kok Keong Loh, Hsu Wen Jing, and Yi Pan

AbstractFixed-power wireless sensor networks are prevalentand cost-effective. However, they face mote failures, RF interfer-ence from environmental noise and energy constraints. Routingprotocols for such networks must overcome these problems toachieve reliability, energy efficiency and scalability in messagedelivery. Achievement of these requirements, however, posesconflicting demands. In this paper, we propose an efficient andreliable routing protocol (EAR) that achieves reliable and scal-able performance with minimal compromise of energy efficiency.The routing design of EAR is based on four parameters -expected path length and a weighted combination of distancetraversed, energy levels and link transmission success history, todynamically determine and maintain the best routes. Simulation

experiments of EAR with four existing protocols demonstratethat a design based on a combination of routing parametersexhibits collectively better performance than protocols based on

just hop-count and energy or those using flooding.

Index TermsEnergy efficiency, hop count, reliable routing,routing protocol, wireless sensor network.

I. INTRODUCTION

WIRELESS sensor networks (WSNs) open up new ap-plication areas such as tactical surveillance, intelligent

environmental and structural monitoring and target tracking

[18]. In a WSN, large numbers of tiny nodes (sensor motes)

may be deployed in an ad hoc manner. These nodes automati-cally configure a topology by communicating and coordinating

with each other [13].Nodes assume the roles of both sensing device and router.

Messages are relayed to other nodes or to a hub in a multi-

hop fashion. Multi-hop routing in an energy-constrained WSN

has been shown to give rise to significant gains in network

performance [27]. With more nodes, the area being monitored

can be increased or with the same area, the increase in node

density gives more precise and timely data and also provides

a degree of operational reliability. Power-controlled networks

have nodes with variable RF power transceivers that provide

greater routing performance at the expense of higher powerconsumption and costs [8], [17], [25]. Fixed-power networkshave cheaper motes with fixed-power RF transceivers but may

be more prone to communication disruptions [2], [14]. Despite

advances in Micro-Electro Mechanical Systems technology,

Manuscript received October 12, 2006; revised January 12, 2007; acceptedJanuary 14, 2007. The associate editor coordinating the review of this paperand approving it for publication was Y. B. Lin.

P. K. K. Loh and W. J. Hsu are with Nanyang Technological University,School of Computer Engineering, Nanyang Avenue, Singapore 639798 (e-mail: {askkloh, hsu}@ntu.edu.sg).

Y. Pan is with Georgia State University, Department of Computer Science,34 Peachtree Street, Suite 1450, Atlanta, GA 30302-4110 USA (e-mail:[email protected]).

Digital Object Identifier 10.1109/TWC.2009.060772

energy limitations still limit sustainability of operations in

WSNs and new, low energy devices are still experimental[2], [15], [21]. Several routing protocols in fixed-power, multi-

hop WSNs use shortest-path routing [3], [8]. Since operation

is often over long unattended periods, the protocol must be

energy efficient. The environment is also unpredictable and

often disrupts operation. As such, routing protocols must

ensure that the WSN can reconfigure [6], [13], be energy

efficient and resilient to failures [14], [21]. These non-trivial

requirements pose conflicting demands on protocol design [5],

[10], [12], [16], [26],[ 30][31]. With these issues in mind,

we propose an efficient and reliable routing protocol (EAR)that routes messages to one or more hubs for data-aggregationapplications. EAR takes into account the expected path length

and a weighted combination of distance traversed, energy level

and past performance of an RF link for its routing decisions.

Control overheads in EAR are low.

The rest of this paper is organised as follows. Section II

surveys some reliable and efficient routing protocols for multi-

hop WSNs. Section III highlights our motivation. Section IV

describes the detailed design of EAR. Simulation results and

some open problems are presented and discussed in Section

V. Section VI concludes this paper followed by the references

and acknowledgements.

I I . RELATED WORK

Here we discuss four recently proposed routing protocols

for reliable and efficient many-to-one routing in multi-hopWSNs: Dynamic Source Routing (DSR) [19-20], Gradient-

Based Routing (GBR) [26], Gradient Broadcast (GRAB) [23],

[28] and Adhoc On-Demand Vector routing (AODV) [11],

[24]. These routing protocols are similar in that they use

neighbourhood information to route packets to the hub. In

DSR, nodes dynamically discover a complete route acrossmultiple hops to the required destination in the network. To

do this, each packet header contains the complete, ordered listof traversed nodes (including the source node). The significant

information overheads here are partially offset by not requiring

periodic route updates and the monitoring of only routes inuse. If an intermediate node is not the destination or it does

not have any route to the destination in its route cache, it will

initiate a route discovery process via request broadcasts to its

neighbors. If available, the complete route to the destination is

found and returned to the initiator. Otherwise, the neighbour

appends its address to the route record and re-broadcasts the

route request to its neighbors. In noisy environments, route

discovery incurs large overheads. When routes become invalid,

DSR adapts by sending a route error packet to the source

1536-1276/09$25.00 c 2009 IEEE


2/8

LOH et al.: PERFORMANCE EVALUATION OF EFFICIENT AND RELIABLE ROUTING PROTOCOLS FOR FIXED-POWER SENSOR NETWORKS 2329

node, which stops using the route. For better reliability, DSR

maintains multiple route entries in its routing table.

In GBR, hop-counts to the hub are computed for each

node and the difference between a nodes hop count and that

of its neighbour is the gradient of that link. Gradients arethus established from nodes to the hub and all messages will

flow in the direction of the greatest gradient. Thus, packet

routing is similar to distance-based techniques utilised by

several existing protocols. When a node detects that its energylevel has decreased to 50% or less, it lowers its gradient to

discourage other nodes from routing packets through it. This

change in gradient must be propagated as far as needed to

keep other gradients consistent. Although this process incurs

significant overhead, it is done only once for each node. If

a link fails or is disrupted, the RF link with the next highest

gradient is chosen. No additional route management overheadsare incurred. With multiple neighbours having links with the

same gradient, one is randomly chosen to route the message.This can spread traffic over time but in WSNs where RF

transmission varies from one node pair to another, even on

optimal routes, additional routing information is required toidentify better performing links.

In GRAB, routing is based on cost and credit. A nodes

cost is the minimum energy overhead to route a packet to

the hub. The hub has a cost of 0 and the value increases

with node distance from the hub. A new packets header is

initialised with the source nodes cost to the hub and assigns

a credit value to the packet before broadcasting it. Intermediate

nodes include their own costs in the packet header but do not

need to designate next hop nodes as flooding is used. This

mechanism provides good reliability at the expense of highenergy consumption, latency and increased packet collisions

especially in a noisy and/or high traffic volume WSN. Onlyneighbours with lower costs may continue routing packets. A

packets credit is consumed at each intermediate node on the

path to the hub. When a packet has enough credit, the node

will forward the packet else it will be dropped. By assigning an

appropriate amount of credit to each packet, duplicate copies

of that packet may travel in multiple paths in a mesh from the

source to the hub. However, this increased in probability of

delivery depends greatly on the fault locations relative to the

width of the broadcast mesh with the latter being widest at

the source. It is also offset by the increased packet collisions.

In AODV, the on-demand route discovery process is similar

to DSR except that each node maintains a routing table withonly one entry for each destination. As a result, AODV incurssimilar information overheads as DSR with lower reliability.

Each table entry records the next hop to that destination

and a sequence number generated by the destination that

indicates whether this information is current or dated. Each

entry also records the addresses of active neighbors through

which packets for the specified destination can be routed.

Whenever a link in use is no longer valid, the upstream node

of that link immediately notifies the active neighbors of the

link. This is repeated until the source nodes using that link are

reached. Route recovery is then initiated. Unlike DSR, there

is reduced recovery overhead as the route request does notrecord traversed nodes but only the traversed node count. A

node that receives route replies, updates its routing table entry

and propagates the reply back towards the source only if the

reply has either a greater destination sequence number (more

current), or the same sequence number with a smaller hopcount. In noisy environments, multi-round discovery may be

needed to establish a route, increasing overheads.

III. MOTIVATION FO R CURRENT WOR K

Routing based on hop-count and node energy levels doesnot account for dynamically varying RF link conditions in

the operating environment. An optimal route need not be the

shortest and a more stable RF link may be more efficient

and reliable. Duplicate packets that are routed to enhance

delivery success incur additional energy consumption and

increased latency. Keeping track of the entire message route

dynamically is expensive, especially in a noisy, disruptiveoperating environment. The onus is to prolong the operational

lifetime of the WSN while tolerating mote failures and RF

disruptions. The performance of the routing protocol also has

to scale with network size. The challenge then is to develop a

routing protocol that can meet these conflicting requirementswhile minimizing compromise.

IV. DESIGN OF EAR

The procedural design of EAR may be divided into three

phases: setup, route selection, and data dissemination. These

are detailed in the following sections.

A. Setup Phase

When a hub is powered on, it broadcasts an Advertisement

(ADV) packet indicating that it wants to receive RPT packets.When a neighbouring node around the hub receives this ADV

packet, it will store the route to the hub in its routing table.Nodes do not propagate the ADV packet received. When a

node is powered on, it delays for a random interval of time

before starting an initialisation process. A node starts the

initialisation process by broadcasting a Route Request (RREQ)

packet asking for a route to a hub. When a hub receives

a RREQ packet, it will broadcast a Route Reply (RREP)

packet. Similarly, when a node receives a RREQ packet, it will

broadcast a RREP packet if it has a route to a hub. Otherwise,

it will ignore the RREQ packet. Nodes do not propagate RREQ

packets. When a node receives a RREP packet, it will store theroute in its routing table. When it has at least one route to thehub it skips the initialisation process. By introducing random

delay for each node to begin initialisation process, a portion

of nodes will receive a RREP packet before they have begun

their initialisation process. This enables faster propagation of

routes and saves on the amount of control packets generated in

the setup phase. A node may store more than one route to the

hub. A route in the routing table is indexed using the next hop

nodes ID - that is the ID of the neighbouring node. A node

keeps only one route entry for a neighbour that has a route to

the hub even though that neighbour could have multiple routes

to the hub. For each route entry in the route table, only thebest route is stored. The selection of best routes is described

next.


3/8


B. Route Selection Phase

Ideally, the best route is the shortest as it incurs the lowest

latency and consumes the least energy. In an actual environ-

ment, the performance of an RF link varies with physical dis-tance and the terrain between nodes and should be accounted

for in routing decisions. In EAR, shortest routes are initially

admitted into the routing table based on hop-count. As RPT

packets flow through these links, less desirable ones will start

to exhibit high packet loss rate and are eventually blacklistedand omitted from the routing table. Links that are omitted

from the routing table may be re-admitted again only after

a period of time. Some RF links are affected by temporary

external disruption and should be given the chance to be re-

admitted. This allows for adaptiveness. The mechanism uses

a sliding window that keeps track of the last N attempts toroute packets on a specified link. If a link fails to relay all

packets in the last N consecutive attempts, then it will beblacklisted and omitted from the table. A metric, LinkScore,is defined as LinkScore = (PE WE + PT WT), wherePE energy level of the next hop node (0.0 to 100.0), WE assigned weight for PE (0.0 to 1.0), PT transmissionsuccess rate (0.0 to 100.0) and WT assigned weight forPT (0.0 to 1.0). Weights, WE and WT, may be determinedempirically but their sum must equal 1. For example, in a low

noise environment, the probability of successful transmission

is higher. In this scenario, WE = 0.7 and WT = 0.3may be chosen allowing routing decisions to focus more

on energy conservation in path selection. Conversely, in anoisy environment, WT = 0.7 and WE = 0.3 could bechosen instead, giving higher emphasis to the selection of high

reliability paths over energy conservation. LinkScore thentakes on a value from 0 to 100 and a higher value indicates

a better link. An arbitrary value is initially assigned to PT asthe link performance is unknown. PT rises (or drops) whensubsequent packet transmissions succeed (or fail). PE startsat 100 and drops as a node consumes its energy resources.

LinkScore is used when there are two links of differentroutes with the same hub distance competing to be admitted

to the routing table. When a new link is received and therouting table is full, link replacement is initiated. The search

ignores blacklisted links and targets the link with the lowest

LinkScore to be replaced. When there is more than one entrywith the same LinkScore, the entry with the longest lengthis chosen to be replaced. This worst route is then compared

against the incoming route and the shorter route is admittedinto the routing table. If there is a tie in route length, then

the route with the higher LinkScore is admitted. Since eachnode stores and maintains the best available RF links in its

routing table, packets travel on the best route from a node to

a hub at the given time. Since the routing table contains only

one entry to the next hop node, its size scales slowly with

network size and multiple hubs.

C. Data Dissemination

Sensor nodes generate RPT packets at periodic intervals

or sleep, waiting for some event to happen. An RPT packetcontains information of interest to network users and has two

fields in its header: ExpPathLen and N umHopT raversed.

A

C

HubB

D

E

F

G

ls=48

ls=75ls=94

Fig. 1. Illustration of forwarding based on LinkScore metric.

The first field is the expected number of hops the packet will

have to traverse before it reaches the hub. It is defined as:

ExpPathLen = N H , where 0.0 < 1.0, N His the number of hops from this node to the hub for the

route selected. The route selected need not be the shortest

but ExpPathLen is bounded by the network diameter. is an assigned weight such that 0.0 < 1.0 since the

minimum number of hops to reach the hub is at least 1.NumHopTraversed is the distance a packet has traversedand is initialised as 0. The packet is forwarded to the next

node in the route. When the next node receives the packet, it

will increment N umHopT raversed by one and compare itwith ExpPathLen. The algorithm is as follows:

By assigning >> 0.0, a packet may favour a route withbetter performing links rather than just the shortest route (Fig.1). If the number of hops that a packet has traversed exceeds

the expected number, there must be changes in the network

topology. During this period of instability, the packet will take

the shortest route to the hub. To prevent potential deadlocks

from occurring, a variable BufUtilLvl is used at each node tostore the current utilisation level of the packet output buffer. Athreshold value, BLThreshold, is defined where BLThreshold< Bmax (max size of buffer). If BufUtilLvl is greater thanBLThreshold, the packet will be relayed on the shortest routeto the hub. This buffer control mechanism ensures that new

packets will not be injected when the buffer is almost fulland there will always be at least one buffer space for transit

packets to be routed.

Proposition 1: EAR is deadlock-free in a connected wire-

less sensor network.

Proof: We recall that ExpPathLen = N H . Let hbe the number of hops traversed by a packet.

Case 1: ExpPathLen h: the data disseminationalgorithm selects the shortest route in the routing table for

forwarding. In the case of a tie, the route with the highest

LinkScore is selected. Selecting the shortest route at everyintermediate node will lead the packet to a hub and no node

is revisited.

Case 2: ExpPathLen > h: the data dissemination algo-rithm selects the route with the highest LinkScore in therouting table for forwarding. If there is a tie in LinkScore,the shorter route is chosen. In this case, the neighbouring

node with the highest LinkScore may have the same hopcount to the hub as the sender node and this may resultin the packet being re-transmitted to the sender. Although

a temporary loop may be possible, a deadlock can never


4/8


1 2

(1) Node 1 sends RTS

1 2

(2) Node 2 sends CTS

1 2

(3) Node 1 sends data packet

Nodes with latest

route info

Nodes without latest

route info

Legend

Fig. 2. Piggybacking on RTS/CTS packet.

occur since BLThreshold < Bmax and buffers will neverbe full. A packet will be forwarded on the shortest route if

the packet output buffer exceeds BLThreshold, thus breakingthe loop. ExpPathLen is bounded by the network diameter.If h ExpPathLen, the packet will travel on the shortestroute to hub, as in case 1.

Proposition 2: The distance a packet travels before reaching

a hub is bounded by N H+ D, where D is the distance of thefurthest node from the hub in terms of hops.

Proof: Consider a packet that is generated at node

A. When is set to 1, ExpPathLen = N H. WhenExpPathLen > NumHopTraversed, the packet is for-warded to its neighbour node B which has the highestLinkScore. Node B may not be on the shortest route inthe routing table and in the worst case, the packet will be

relayed along the longest route at every intermediate node

and the packet will traverse N H hops until ExpPathLen =NumHopTraversed. In the worst case, the packet will havebeen routed to the furthest node from the hub. Now, the packet

will be forwarded on the shortest route and the packet willneed a maximum of D hops to reach the hub.

Corollary 1: The routing algorithm of EAR is livelock-free.

Proof: In a non-ideal environment, loops formed are

temporary and by Theorem 1, will eventually be exited. By

Theorem 2, a packet eventually arrives at a hub after a

maximum of N H + D hops, where D is the distance of thefurthest node from the hub in hops.

D. Route Update

Sensor nodes continually update best routes in the routing

table. Instead of explicit control packets, EAR uses the hand-shaking mechanism at the MAC layer. Route information ispiggybacked onto both RTS and CTS packets. Fig. 2 illustrates

this scenario. Nodes in blue have received updated route

information from either node Xs RTS or node Ys CTS

packet or both. RTS and CTS packets have to be received

and processed by all nodes as part of the collision avoidance

mechanism employed by the MAC protocol [29]. Hence, util-

ising RTS-CTS handshaking instead of separate DATA-ACK

would result in more current route information for a node.

As an example, EAR can use S-MAC (Sensor MAC) [17]

that has energy saving mechanisms. S-MAC uses the same

four-way handshaking mechanism as IEEE 802.11 to achievereliable link-to-link transmission. One of the energy-saving

mechanisms known as Overhearing Avoidance specifies that

Packet Delivery Ratio (PDR) with 10% Source Nodes

0

10

20

30

40

50

60

70

80

90

100

0 50 100 150 200 250 300 350 400

Number of nodes

%o

fpa

cketsdeliveredsuccessfully

AODV

DSR

EAR

GBR

GRAB

Fig. 3. Packet Delivery Ratio results (with 10% active sources).

Packet Delivery Ratio (PDR) with 50% Source Nodes

0

10

20

30

40

50

60

70

80

90

100

0 50 100 150 200 250 300 350 400

Number of nodes

%o

fpacketsdeliveredsuc

cessfully

AODV

DSR

EAR

GBR

GRAB

Fig. 4. Packet Delivery Ratio results (with 50% active sources).

nodes upon hearing a RTS or CTS packet that is not addressed

to them will go into sleep mode. Periodically exchanging

routing information between nodes is costly in terms of energy

consumption and bandwidth usage. Piggybacking route infor-

mation onto existing RTS and CTS packets incurs additional

energy consumption as the packet size increases. However,

no extra packets need be generated and additional costs are

negligible compared to the cost incurred in relaying explicitroute information (control) packets.

Cost analysis of route update: We show the energy

and latency costs in piggybacking route information on the

RTS/CTS packet compared to sending an explicit control

packet. The energy consumed is proportional to packet size

and we assume the energy cost to transmit 1 bit is 1 unit. Let

K be the size of the route information in bits, W be the sizeof the routing control packet in bits and the bandwidth be Bbits per second. Then, we have:

Energy cost of piggybacking = K units, Latency cost ofpiggybacking = K/B seconds,

Energy cost of explicit control = (SRTS + SCTS + K+W + SACK) units, and

Latency cost of explicit control = (SRTS/B) +(SCTS/B) + ((K+ W)/B) + (SACK/B) seconds

where SRTSRT Spacket size in bits, SCTSCT S packetsize in bits and SACK ACK packet size in bits.


5/8


Packet Latency with 10% Source Nodes

0

0.5

1

1.5

2

2.5

0 50 100 150 200 250 300 350 400

Number of nodes

Latencyperpacket(s)

AODV

DSR

EAR

GBR

GRAB

Fig. 5. Packet Latency results (with 10% active sources).

Packet Latency with 50% Source Nodes

0

1

2

3

4

5

6

7

8

0 50 100 150 200 250 300 350 400

Number of nodes

Latencyperpacket(s)

AODV

DSR

EAR

GBR

GRAB

Fig. 6. Packet Latency results (with 50% active sources).

V. SIMULATION

GloMoSim [1] was used to emulate a WSN with crossbow

MICA2 motes [9] for data aggregation. In the simulation, all

nodes generate data packets that are routed to the hub located

in the centre of the WSN. To focus on performance variationsdue to design differences among the routing protocols tested,

we used a simulation model that precludes effects such asfading, shadowing and multipath losses. To ensure consistency,

the same model is applied to all routing protocols tested.

The routing protocols are subjected to a series of tests to

evaluate their performances in a realistic WSN environment

that is compounded with noise and node failures. We simulated

network sizes from 50 to 400 nodes with 10% and 50% activesource nodes. Every node except the hub takes on a random

noise factor between 10% and 50%. The noise factor specifies

the probability that packets to be received by that node are

corrupted or lost. Also, 50% of randomly selected nodes fail

at random times within the simulation duration. Results wereaveraged over 30 runs each with a different seed and are shown

in Figs. 38. The following standard metrics [10], [13], [23]

were used:

Packet delivery ratio (PDR) measures the percentage

of data packets generated by nodes that are successfully

delivered, expressed as

Total number of data packets succesfully delivered

Total number of data packets sent 100%

Energy Consumption with 10% Source Nodes

0

10

20

30

40

50

60

70

80

90

100

0 50 100 150 200 250 300 350 400

Number of nodes

Energy

expendedperpacket(mJ)

AODV

DSR

EAR

GBR

GRAB

Fig. 7. Energy Consumption results (with 10% active sources).

Energy Consumption with 50% Source Nodes

0

50

100

150

200

250

0 50 100 150 200 250 300 350 400

Number of nodes

Energyexpendedperp

acket(mJ)

AODV

DSR

EAR

GBR

GRAB

Fig. 8. Energy Consumption results (with 50% active sources).

Packet latency measures the average time it takes to route

a data packet from the source node to the hub. It is expressedas

Individual data packet latency

Total number of data packets delivered.

Energy Consumption: This measures the energy expended

per delivered data packet. It is expressed as

Energy expended by each node

Total number of data packets delivered.

We calculate energy expended in transmission and receptionby the nodes RF transceivers. This metric also indirectly

measures the amount of control packet overhead of a routing

protocol.Fault Tolerance and Scalability: Values for the above

metrics are recorded for 10% and 50% active source nodes

with a specified fault model: 50% of randomly selected nodes

fail and surviving nodes are exposed to a random noise factor

ranging from 10% to 50% to disrupt RF communications.

A. Performance Evaluation

Packet Delivery Ratio: DSR operates in promiscuous mode

such that nodes can obtain the latest routing information

and packets are routed on valid paths with high probability.

Multiple paths are kept in the routing table giving DSR agood degree of reliability. DSR exhibits moderately high PDR

(Figs. 3, 4). Besides path length, GBR considers a nodes


6/8


energy level to distribute packet traffic more uniformly among

all nodes. The improvement over DSR and GRAB is evident

with larger networks and higher traffic levels (Fig. 4). Moreactive sources contribute to a higher rate of packet collisions.

Spreading of network traffic reduces the chance of collisions.

However, GBRs selection of the next hop node does not

account for the stability of RF links that may vary over time.

GRAB has the second highest PDR on average as it floods

packets in a mesh to the hub. With 50% active source nodes,however, GRAB is less robust (Figs. 3, 4). Additional packet

loss is contributed by collisions as flooding generates many

redundant packets. A packet broadcasted by a source node may

also be lost for good during transmission as broadcast does not

provide for link-to-link transmission and acknowledgement.

Although the route discovery process in AODV is similar to

DSR, each node only maintains a single routing table entry for

each destination. A single route discovery in AODV reveals

less information than in DSR. Hence, within the same time,fewer routes are discovered with the consequence that the

number of packets delivered is less (Figs. 3, 4).

With both 10% and 50% active source nodes, EAR achievedabout 97% PDR on average compared to 82% for GRAB and70% for GBR. Besides expected path length, EARs routing

decisions are based on a weighted combination of distance

traversed, energy and link performance history. EAR blacklists

RF links with consistently poor performance and discards

them from the routing table. The performance is consistenteven with larger networks up to 400 nodes (Figs. 3, 4).

Packet Latency: GRAB and DSR incur the highest laten-cies, especially on networks with 200 nodes upwards (Figs.

5, 6). GRAB floods packets in a mesh to the hub. With 50%source nodes, packet latencies increase exponentially, off the

graph (Fig. 6). Values recorded are: 36.6s (10% active), 94.3s(200 nodes, 50% active), 138.9s (300 nodes, 50% active), and

318.9s (400 nodes, 50% active). Number of control packets

produced by GRAB fluctuates significantly across simulations

and therefore it shows widely fluctuation performance. Largeerror bars have also been observed in an independent study [3].

DSR exhibits large packet latencies because its route discovery

takes more time. Every intermediate node tries to extract and

record information before forwarding a reply. The same thing

happens when a data packet is routed from node to node.Hence, while route discovery in DSR yields more information

for delivery, packet transmission slows down. Although route

discovery process in AODV is similar to DSR, each nodemaintains only one routing table entry for each destination.

Unlike DSR, AODV does not record traversed nodes, thus

reducing route update latencies. As network size increases,

this is more evident (Figs. 5, 6).

EARs packet latency is 0.05 seconds higher than GBR on

average. With network sizes of 350 nodes upwards, the packet

latency is 0.4 seconds higher than GBR on average (Figs.

5, 6). This is due to the higher control overhead compared

to GBR, despite the piggy-backing. However, this is good

compromise resulting in appreciably higher PDR (Figs. 3, 4).

GBR exhibits the lowest packet latencies of all protocols (Figs.

5, 6). In GBR, when a nodes energy reserves fall below 50%,the node reduces its gradient with respect to its neighbours,

discouraging further routing through it. The gradient change

TABLE IENERGY CONSUMED PER PACKET FOR AODV AN D GRAB

Energy Expended Per Packet (mJ) for 10% active sources

50 100 150 200 250 300 350 400

AODV 7.102 20.777 111.394 175.474 256.966 272.406 331.473 683.139

GRAB 61.741 112.126 114.814 300.578 183.041 206.479 235.023 449.239

Energy Expended Per Packet (mJ) for 50% active sources

50 100 150 200 250 300 350 400

AODV 29.033 104.697 182.426 293.484 367.358 510.777 585.768 824.233

GRAB 69.713 131.59 100.48 214.843 148.559 215.705 186.362 314.066

is progressively propagated to neighbouring nodes to maintain

consistency. However, this process is initiated only once by

each node. There is no need for regular updates of routingtables and hence, latency is reduced significantly due to lower

control overheads.

Energy Consumption: Route discovery in AODV is energy-intensive. The data packet carries pointers to the full route in

itself, which incurs additional energy overheads during routing

compared to data packets of routing protocols that carry only

neighbourhood information. The additional energy consumed

is proportional to network size. With an operating environmentthat is plagued with dynamically occurring link disruptions

and node failures, it may be very difficult to establish a full

route from source to the destination at a given point in time.

The source will keep sending the route discovery packet but

will not receive a definite route response from the destination.

Route discovery packets will accordingly flood the network

consuming more energy. The result is exacerbated energy

overheads, especially manifested in larger networks above 150nodes (Table 1).

GRAB incurs high levels of energy consumption as it floods

packets in a mesh to the hub. Energy is also expended to

build up the costfi

eld. Excepting AODV, the consumptionlevels of GRAB are at least an order of magnitude higher

than the other protocols. With larger networks from about

200 nodes upwards, the additional redundant routes available

to packets appreciably increase packet delivery success over

AODV (Fig. 3, 4). Less control packets are needed compared

to AODV especially for route discovery. Hence, GRABs

energy overheads show an improvement over AODVs for

larger network sizes (Fig. 8). Despite the large error bars, the

appreciably higher energy consumption levels of both AODV

and GRAB relative to the other protocols are evident fromthe resulting trends. DSR makes use of promiscuous mode to

constantly update route information. This reduces total energy

consumption as less re-transmissions and control packets are

needed. As in AODV, however, route discovery broadcasts in

DSR can lead to significant energy consumption especially inlarger networks. As an improvement over AODV, DSR uses a

route cache to reduce route discovery costs (Fig. 7, 8).

GBR and EAR exhibit the lowest energy overheads (Fig. 7,

8). As GBR directly monitors each nodes energy reserves, it

permits a more gradual and uniform rate of energy consump-

tion over the network. Though the progressive propagation

of gradient changes to neighbouring nodes consume energy,

the process is done only once for each node when its

energy level goes below the 50% threshold. Energy overheadsof EAR are competitive with that of GBR. The slight im-

provement over GBR with larger networks may be attributed


7/8


in part to EAR dynamically accounting for variable link

performance and forwarding packets only on best routes. By

piggybacking on RTS/CTS packets, current route informationis disseminated quickly with low energy costs. As the cost

analysis in Section IV-D shows, the energy cost of a single

route update is appreciably lower when piggybacking than

using explicit control packets.

Fault Tolerance and Scalability: AODV and DSR utilise

on-demand route discovery based onflooding the network withroute requests and are more suitable for a small- to moderate-

sized network with minimal noise and RF disruptions. As

network size scales upwards and in the presence of appreciable

node and link failures, costs associated with route repair

scale exponentially. The mesh-broadcast feature of GRABmay increase the robustness of data delivery in the presence

of faulty nodes and noisy channels. Like AODV and DSR,

however, GRAB is more suitable for small to moderate-sized

networks with a small proportion of active source nodes. With

larger networks and heavy traffic, overheads incurred and

packet collisions encountered rise disproportionately leading

to poor performance scalability. Fault tolerance exhibited byGRAB depends very much on the location of node failuresand link disruptions. When these occur nearer the source node,

theres a higher chance of avoiding it as the mesh is wider.

The width of the mesh is controlled by the amount of creditthat is used up as packets get closer to the hub. Packet latency

and energy consumption exhibited by GBR and EAR scale

well with network size in the presence of node faults and link

disruptions. GBR latency and energy consumption figures are

essentially independent of the number of active sources and

also scale well with network size. Its performance is largely

due to its one-off approach in monitoring energy levels and its

use of shortest routes whenever possible. For packet latencyand energy consumption, EAR exhibits performance close to

GBR. By also accounting for link performance history, EAR

exhibits consistently higher PDR than GBR over all network

sizes.

B. Open Issues

In the proposed network setup, nodes relay all data packets

to the hub. In networks with a single, centrally-located hub,neighbouring nodes to the hub route more packets than other

nodes. These nodes will drain their energy-reserve at a muchfaster rate. One possible solution to this problem could be to

adopt a clustering approach wherein cluster node membersroute only to the cluster heads and cluster heads may be

implemented with variable-powered motes.

VI . CONCLUSION

This paper has proposed a viable routing protocol, EAR, for

data aggregation in fixed-power WSNs. Routes are managed

based on expected path length and a weighted combination

of distance traveled, energy level and RF link performance

history. Simulation results have shown that it performs com-

petitively against existing routing protocols in terms of packet

delivery ratio, packet latency, scalability and energy con-sumption while operating in a noisy wireless environment

where network traffic, link disruptions and node failure rates

are high. Future work includes studying the open issues

and investigating the security aspects of EAR, especially in

clustered WSNs.

ACKNOWLEDGEMENT

We would like to thank the Nanyang Technological Univer-

sity and Georgia State University for supporting our research

efforts. Yi Pans work was also supported by the 111 project

of China under the grant No. 111-2-14. We acknowledge

Long Say Huans help in the development and testing ofEAR. We also thank the editors and anonymous reviewers

recommendations that have significantly improved this paper.

REFERENCES

[1] A. R. Bagrodia, L. Bajaj, M. Gerla, and M. Takai, GloMoSim: ascalable network simulation environment," technical report 990027,UCLA, Computer Science Dept., 1999.

[2] I. F. Akyildiz, E. Cayirci, Y. Sankarasubramaniam, and W. Su, A surveyon sensor networks," IEEE Trans. Commun., pp. 102114, 2002.

[3] T. Bokareva, N. Bulusu, and S. Jha, A performance comparison of datadissemination protocols for wireless sensor networks," in Proc. IEEEGlobecom, 2004, pp. 8589.

[4] A. Bourkerche, Performance evaluation of routing protocols for ad hocwireless networks," Mobile Networks Applications, vol. 9, no. 4, pp.333342, 2004.

[5] A. Bourkerche, R. W. N. Pazzi, and R. B. Araujo, Fault-tolerantwireless sensor network routing protocols for the supervision of context-aware physical environments," J. Parallel Distributed Computing, vol.66, no. 4, pp. 586599, Apr. 2006.

[6] N. Bulusu, J. Heidemann, D. Estrin, and T. Tran, Self-configuringlocalization systems: design and experimental evaluation," ACM Trans.Embedded Computing Systems, vol. 3, no. 1, pp. 2460, 2004.

[7] A. Chandrakasan, S.-H. Cho, N. Ickes, R. Min, E. Shih, A. Sinha,and A. Wang, Physical layer driven protocol and algorithm design forenergy-efficient wireless sensor networks," in Proc. 7th Annual Intl.

Conf. Mobile Computing Networking, 2001, pp. 272287.[8] J.-H. Chang and L. Tassiulas, Maximum lifetime routing in wireless

sensor networks," IEEE/ACM Trans. Networking, vol. 12, no. 4, pp.609619, Aug. 2004.

[9] CrossBowMICA2 mote specs [Online]. Available:http://bullseye.xbow.com/Products/productdetails.aspx?sid=192

[10] S. R. Das, C. E. Perkins, E. M. Royer, and M. K. Marina, Performancecomparison of two on-demand routing protocols for ad hoc networks,"

IEEE Personal Commun., vol. 8, no. 1, pp. 1629, Feb. 2001.[11] S. R. Das, C. E. Perkins, and E. M. Royer, Ad hoc on-demand distance

vector (AODV) routing," IETF Internet draft, draft-ietf-manet-aodv-13.txt, Feb. 2003 (work in progress).

[12] G. Deepak, A. Cerpa, Y. Yu, W. Ye, J. Zhao, and D. Estrin, Networkingissues in sensor networks," J. Parallel and Distributed Computing, vol.64, no. 7, pp. 799814, 2004.

[13] D. Estrin and A. Cerpa, ASCENT: adaptive self-configurable sensor

networks topologies," IEEE Trans. Mobile Computing, vol. 3, no. 3, pp.272285, July 2004.

[14] D. Estrin and J. Elson, Wireless sensor networks: a bridge to thephysical world," Wireless Sensor Networks, pp. 121, 2004.

[15] D. Estrin, L. Girod, G. Pottie, and M. Srivastava, Instrumenting theworld with wireless sensor networks," in Proc. International Conf.

Acoustics, Speech Signal Processing (ICASSP), May 2001, pp. 14.[16] D. Estrin, C. Intanagonwiwat, R. Govindan, J. Heidemann, and F. Silva,

Directed diffusion for wireless sensor networking," IEEE/ACM Trans.Networking, vol. 11, no. 1, pp. 216, 2003.

[17] A. Goldsmith and S. Chua, Variable-rate variable-power M-QAM forfading channels," IEEE Trans. Commun., vol. 45, pp. 12181230, Oct.1997.

[18] Z. J. Haas and T. Small, A new networking model for biologicalapplications of ad hoc sensor networks," IEEE/ACM Trans. Networking,vol. 14, no. 1, pp. 2740, Feb. 2006.

[19] Y. C. Hu, J. G. Jetcheva, D. B. Johnson, and D. A. Maltz, The dynamicsource routing protocol for mobile ad hoc networks," Internet draft,draft-ietf-manet-dsr-09.txt, Apr. 2003.


8/8


[20] D. B. Johnson, D. A. Maltz, and J. Broch, DSR: the dynamic sourcerouting protocol for mulit-hop wireless ad hoc networks," Ad Hoc

Networking, Charles E. Perkins, ed., Chap. 5, pp. 139172. Addison-Wesley, 2001.

[21] J. M. Kahn, R. H. Katz, and K. S. J. Pister, Next century challenges:mobile networking for smart dust," in Proc. 5th Annual ACM/IEEE

International Conf. Mobile Computing Networking, 1999, pp. 271278.[22] LAN MAN Standards Committee of the IEEE Computer Society,

Wireless LAN medium access control (MAC) and physical layer (PHY)specification," New York, USA, IEEE Std 802.11-1997.

[23] F. Ye, G. Zhong, S. Lu, and L. Zhang, Gradient broadcast: a robustdata delivery protocol for large scale sensor networks," ACM WirelessNetworks, vol. 11, no. 2, pp. 285298, 2005.

[24] C. E. Perkins, E. M. Belding-Royer, and S. Das, Ad hoc on demanddistance vector (AODV) routing," IETF Internet draft, draft-ietf-manet-aodv-12.txt, Nov. 2002 (work in progress).

[25] V. Rodoplu and T. H. Meng, Minimum energy mobile wireless net-works," IEEE J. Select. Areas Commun., vol. 17, no. 8, pp. 1333-1344,Aug. 1999.

[26] C. Schurgers, V. Raghunathan, and M. B. Srivastava, Power manage-ment for energy-aware communication systems," ACM Trans. EmbeddedComputing Systems, vol. 2, no. 3, pp. 431-447, Aug. 2003.

[27] S. Toumpis and A. J. Goldsmith, Capacity regions for wireless ad hocnetworks," IEEE Trans. Wireless Commun., vol. 2, no. 4, pp. 736748,July 2003.

[28] F. Ye, G. Zhong, S. Lu, and L. Zhang, Gradient broadcast: a robustdata delivery protocol for large scale sensor networks," ACM Wireless

Networks, vol. 11, no. 2, pp. 285298, 2005.[29] W. Ye and J. Heidemann, Medium access control in wireless sensor

networks," Wireless Sensor Networks, pp. 73-91, 2004.[30] Q. Zhao and G. Mohan, Topology knowledge range control for lifetime

maximization in sensor networks with data aggregation," in Proc. 2ndACM Intl Workshop Performance Evaluation Wirelesss Ad Hoc, Sensor

Ubiquitous Networks, 2005, pp. 8491.[31] Q. Zhao and L. Tong, Energy efficiency of large-scale wireless net-

works: proactive versus reactive," IEEE J. Select. Areas Commun., vol.23, no. 5, pp. 11001112, 2005.

Peter K. K. Loh is currently an Associate Pro-fessor of the School of Computer Engineering,Nanyang Technological University of Singapore.He has authored and co-authored several papers inwireless sensor networks, parallel and distributedsystems, and computer gaming. He has previouslyheld positions as Head of Software Cluster Labsand the Parallel Processing Lab. He obtained hisPh.D. from Nanyang Technological University, theMSc (Comp.Sc.) from Manchester Univeristy, theMSc (Elect.Engg.) and the BEng from the National

University of Singapore. He is also a registered Professional Engineer(Singapore), Chartered Engineer (UK), and a Senior Member of IEEE.

Wen-Jing Hsu is presently with the School of Com-puter Engineering, Nanyang Technological Univer-sity. He is also a Faculty Fellow of the Singapore-MIT Alliance (SMA) Computer Science Programsince 2001. At NTU, he has initiated and com-pleted several large scale externally funded researchprojects. He has authored or co-authored many pa-pers on parallel and distributed processing; amongthem four papers were nominated for Best PaperAwards (with two awarded). He is presently Deputy

Director of the Maritime Research Center, NTU andhas published a book Anatomy of HCTS- a High-capacity Container HandlingSystem for Mega Vessels. He was previously Director of the Center ofAdvanced Information Systems, and Deputy Director of Financial Engineeringat NTU. He also held positions on the faculty of Michigan State University,USA, and was a Visiting Scientist at the IBM T. J. Watson Research Center,Yorktown Heights, and the IBM Palo Alto Scientific Center.

Yi Pan received his B.Eng. and M.Eng. degreesin computer engineering from Tsinghua University,China, in 1982 and 1984, respectively, and his Ph.D.degree in computer science from the University ofPittsburgh, USA, in 1991. Currently, he is the chairand a full professor in the Department of Com-puter Science at Georgia State University. Dr. Pansresearch interests include parallel and distributedcomputing, optical networks, wireless networks, and

bioinformatics. Dr. Pan has published more than 80journal papers with 28 papers published in various

IEEE journals. In addition, he has published over 90 papers in refereedconferences (including IPDPS, ICPP, ICDCS, INFOCOM, and GLOBECOM).He has also co-edited 13 books (including proceedings) and contributedseveral book chapters. His pioneering work on computing using reconfigurableoptical buses has inspired extensive subsequent work by many researchers, andhis research results have been cited by more than 100 researchers worldwidein books, theses, journal and conference papers. He is a co-inventor ofthree U.S. patents (pending) and five provisional patents, and has receivedmany awards from agencies such as NSF, AFOSR, JSPS, IISF and theMellon Foundation. His recent research has been supported by NSF, NIH,NSFC, AFOSR, AFRL, JSPS, IISF, and the states of Georgia and Ohio.He has served as a reviewer/panelist for many research foundations/agenciessuch as the U.S. National Science Foundation, the Natural Sciences andEngineering Research Council of Canada, the Australian Research Council,

and the Hong Kong Research Grants Council. Dr. Pan has served as an editor-in-chief or editorial board member for eight journals including three IEEETransactions and as a guest editor for seven special issues. He has organizedseveral international conferences and workshops and has also served as aprogram committee member for several major international conferences suchas INFOCOM, GLOBECOM, ICC, IPDPS, and ICPP. Dr. Pan has deliveredover 50 invited talks, including keynote speeches and colloquium talks atconferences and universities worldwide. Dr. Pan is an IEEE DistinguishedSpeaker (20002002), a Yamacraw Distinguished Speaker (2002), a Shell OilColloquium Speaker (2002), and a senior member of IEEE. He is listed in

Men of Achievement, Whos Who in the Midwest, Whos Who in America,Whos Who in American Education, Whos Who in Computational Scienceand Engineering, and Whos Who of Asian Americans.

Documents

Sivram-base Paper IEEE