Cluster Communication Synchronization in Delay-SensitiveWireless Sensor Networks

Xi Deng and Yuanyuan YangDepartment of Electrical and Computer Engineering

Stony Brook University, Stony Brook, NY 11794, USA

Abstract—Clustering has been widely used in wireless sensornetworks to increase scalability, improve energy efficiency andprovide QoS guarantees. In such networks, frequent interactionsbetween the intra-cluster communication and the inter-clustercommunication are inevitable, which may severely downgradethe communication efficiency and hence the network performanceif not handled properly. This is especially problematic in delay-sensitive data gathering applications. Thus, proper synchroniza-tion among these two types of communications is required. Inthis paper, we propose two approaches to schedule the com-munications in clustered wireless sensor networks aiming atdelay-sensitive applications. In the first approach, an efficientcycle-based synchronous scheduling is proposed to achieve lowaverage packet delay and high throughput by optimizing the cyclelength and transmission order. In the second approach, a novelclustering structure is introduced to eliminate the necessity ofcommunication synchronization so that packets are transmittedwith no synchronization delay, yielding very low end-to-endpacket delay. Our extensive experimental results demonstrate thesuperior performance of both approaches. The distinct behaviorof two scheduling approaches enables them to support a widerange of data gathering applications with different performancerequirements in clustered wireless sensor networks.

Index Terms—Clustering, wireless sensor networks, communi-cation synchronization, delay-sensitive, data gathering.

I. INTRODUCTION

Clustering has been widely used in wireless sensor networks(WSNs) to increase scalability, improve energy efficiency andprovide QoS guarantees. With clustering, sensor nodes areorganized into clusters and a cluster head (CH) node is selectedfor each cluster according to certain rules, while other nodesact as members in the clusters. In cluster-based data gathering,data collected by cluster members are first sent to CHs, whichin turn deliver the data to the data sink either by directcommunication or through relays on intermediate CHs. Whileclustering is initially introduced to achieve energy efficiency, itcan also help maintain low packet latency in delay-sensitivedata gathering. This is because that packets from differentmembers can be combined as aggregated packets at CHs toreduce the transmission overhead of packet headers and controlpackets (e.g., ACK packets), leading to shortened transmissiondelay. In addition, clustering simplifies the routing from thesource node to the sink, and shorter routing paths reducenetwork traffic as well.

However, cluster-based WSNs encounter a new communica-tion synchronization problem due to their more complex com-munication patterns compared to WSNs with a flat topology. Ingeneral, the communication in a cluster-based WSN includesintra-cluster communication among sensors in the same cluster

and inter-cluster communication among different CHs. Intra-cluster communication in each cluster is usually controlledby the CH with a Time-Division-Multiple-Access (TDMA)based protocol to avoid transmission collisions. For inter-clustercommunication, CHs can be considered to form a smallerrelay network where either TDMA or Carrier-Sense-Multiple-Access (CSMA) based protocols can be utilized. To avoidinterference between intra- and inter-cluster communications,different channels are used for two types of communications,which implies that CHs have to switch between two channelsaccordingly as most of sensors can operate on only one radiochannel at a time. Let i-state and o-state denote that a CH usesthe channel for intra-cluster communication and inter-clustercommunication, respectively. Such state switching thus incursa synchronization problem which is critical for delay-sensitiveapplications: the sending CH and the receiving CH should be ino-state simultaneously and any inter-cluster packet transmissionmay endure an unacceptable long delay before the receiverswitches to o-state. In such a case, the inter-cluster packet thatconsists of multiple sensing packets may become useless andbe discarded, causing severe performance loss.

Previous work that targets at the energy efficiency of cluster-ing handles this synchronization problem by simply groupingthe intra- and inter-cluster communications involved in allclusters into two global and non-overlapping periods. In thisapproach, since the intra- and inter-cluster communicationsare guaranteed to be performed separately, the synchronizationproblem can be avoided. However, by intentionally separatingtwo types of communications, such an approach may causelow channel utilization and hence long end-to-end delays,rendering it not suitable for delay-sensitive applications. In thispaper, we propose two communication scheduling approachesto solving the synchronization problem from different anglesand support delay-sensitive data gathering applications withdifferent requirements. To the best of our knowledge, this isthe first work that tackles the communication synchronizationwhile considering packet delays in cluster-based WSNs.

We first propose a TDMA based, synchronous schedulingapproach to achieve low end-to-end packet delay by convertingthe cluster synchronization problem into a scheduling problemin generic wireless networks. Due to the NP-completenessof the scheduling problem, we propose an efficient heuris-tic scheduling algorithm. Compared to other cycle-based ap-proaches in the literature, our approach owns three uniquefeatures. First, the cluster heads can individually decide theirintra-cluster packet collection time, rather than a globally

2013 IEEE International Conference on Distributed Computing in Sensor Systems

978-0-7695-5041-1/13 $26.00 © 2013 IEEE

DOI 10.1109/DCOSS.2013.68

36

synchronized collection time. Second, since all packets are sentto the sink, we schedule transmissions according to the order ofnodes in the routing path to minimize the queueing delay. Third,we efficiently overlap the transmissions so that the cycle lengthand the average end-to-end packet delay can be reduced asmuch as possible. Experiments show our approach can achieve50% shorter average packet delay than the existing approach.

The cycle based approach may have some restrictions ina harsh environment due to its vulnerability to the clockdrift, topology changes and irregular interferences occurred inWSNs [14]. We thus propose the second approach that solvesthe communication synchronization problem asynchronously.The approach is constructed on a new clustering structurewith a new type of node, called relay node, rather than theconventional CH-member structure. The relay nodes stay ino-state and replace the CHs to receive and forward the aggre-gated packets. With the assistance of relay nodes, inter-clustercommunications are automatically synchronized. Compared tothe first approach, the asynchronous approach better utilizesthe wireless channel and yields even lower packet delay whileconsuming a reasonable amount of energy. The performanceof both approaches has been verified through extensive ns-2simulations.

The rest of the paper is organized as follows. Section IIdiscusses the related work. Section III and Section IV describethe design of synchronous and asynchronous approaches, re-spectively. Section V presents the experimental results for theproposed approaches. Finally, Section VI gives the conclusion.

II. RELATED WORK

Clustering is a popular topology control approach to achiev-ing energy efficiency and scalability in WSNs. In this section,we briefly review the cluster formation algorithms and then dis-cuss some existing work concerning communication protocolsin cluster-based networks.

Cluster formation algorithms have been extensively studiedin the literature. Their primary purpose is to consider loadbalance and energy efficiency to prolong network lifetime.While some algorithms consider a heterogeneous environmentwhere CHs are more powerful than regular sensor nodes [1],[4], other algorithms consider a homogeneous environmentwhere CHs are ordinary sensors [5]. In this paper, we mainlyfocus on clustering in a homogeneous environment. Typicalalgorithms in this category include LEACH [2] and HEED [3]LEACH selects CHs randomly and distributes energy consump-tion evenly among all nodes by cluster head rotation. HEEDselects CHs by considering the residual energy in the nodes andthe communication cost. A comprehensive survey on differentclustering algorithms can be found in [5].

There is also some work on communication protocols incluster-based networks. While intra-cluster communication inthese protocols is always TDMA-based, the inter-cluster com-munication adopts different approaches in different works. In[6], a round-based data collection scheme with direct sinkaccess was proposed. It was assumed that CHs can directlyaccess the sink, which has the capability of multiple packet

reception. The intra- and inter-cluster packet delay was consid-ered separately and the end-to-end delay was not studied. Asimilar round-based protocol was proposed in [7], where therouting path for inter-cluster communication consists of eithercluster heads or a combination of CHs and members. In [8], apure TDMA-based scheme was proposed to achieve optimizedenergy efficiency and minimum delay, in which the packetdelay is directly associated with the length of TDMA frame. Inaddition, the MAC protocol defined in IEEE 802.15.4 can beutilized in cluster-based networks [9]. In particular, a cluster-tree topology is constructed with each CH corresponding toa coordinator, which maintains a superframe with 16 slots.The members are allowed to communicate with the CH in anyslot in the superframe. In general, superframes for differentcoordinators do not overlap so that the interference amongdifferent clusters can be avoided.

In the above protocols, communication synchronization ishandled by either setting a complete transmission schedulefor every CH, or globally separating the intra- and inter-cluster communications. As will be seen in the performanceevaluation section later, compared to our proposed approaches,these approaches do not perform well in delay-sensitive datagathering.

III. SYNCHRONOUS SCHEDULING

In this section, we present Cycle-Based Scheduling (CBS),a TDMA based, synchronous scheduling approach. To beginwith, we describe the assumptions for the system, which arealso shared by the asynchronous scheduling approach.

A. Assumptions

The network considered is a WSN with 𝑛 sensor nodesrandomly distributed in a 2D region. We consider a typical datagathering application in WSNs where all sensor nodes sendcollected data to a single sink. We also make the followingassumptions on the WSN.

∙ The clustering topology is pre-constructed by a clusteringalgorithm, such as the algorithms mentioned in Section II,which indicates that the size of the different clusters maybe different. In addition, we assume the topology of therelay network is stable during the data gathering. This isreasonable if the CHs are properly selected with adequateenergy.

∙ Sensors can transmit on different radio channels. However,they can only transmit or receive packets on one channelin any instant. Different radio channels do not interferewith each other.

∙ Sensors have the same sensing rate 𝜆 and sensing packetsare of the same length. The packet generation process isassumed to be Poisson.

Under the above assumptions, which are applicable in manyreal-world networks, next we describe the details of CBS.

B. Basic Communication Cycle

CBS schedules communications in consecutive cycles andeach node is assigned some fixed conflict-free intervals to

37

transmit and receive packets in each cycle. Nodes only wakeup in the assigned intervals and sleep otherwise to reduceenergy consumption. Each node is assigned a single intervalfor transmission so that the synchronization overhead betweenthe transmission pair is minimized. The goal of the schedulingis to minimize the average end-to-end packet delay. We considerthis problem by separating the intra- and the inter-clustercommunications.

1) Intra-cluster communication: Intra-cluster communica-tion includes all transmissions from cluster members to theCH. Since interferences from other clusters can be avoidedby assigning different radio channels to adjacent clusters, thecommunications within a cluster are independent and hence itis reasonable to only consider a single cluster.

We limit all communications for a cycle in a consecutiveperiod so that the CH needs not to frequently switch betweenintra- and inter-cluster communications. As will be seen later,the duration of this period is relatively short compared to thecycle length, we simply consider a general TDMA schemefor the intra-cluster period. The whole period is divided intomultiple identical time slots whose length 𝜏 is set equal tothe time required for a packet transmission. Packets are sentin these time slots directly from cluster members to the CH.Each node is assigned the same 𝑘 time slots given their samepacket generation rate. For simplicity, we assume the CH is alsoassigned 𝑘 time slots for necessary control packets. Assume thecluster has 𝑚 nodes, the duration of the intra-cluster period isthus 𝑚 ⋅ 𝑘 ⋅ 𝜏 .

For such scheduling, we are concerned with the determina-tion of 𝑘 and the packet collection delay, which is defined asthe elapsed time between the packet is generated and the endof the intra-cluster period in which the packet is collected.

Lemma 1: The lower bound of 𝑘 is ⌈𝜆𝑇 ⌉, where 𝑇 is thecycle length.

Proof: The expected number of packets generated by eachnode in a cycle is 𝜆𝑇 . Since a cluster member can transmitone packet in a time slot, in order to collect all packets in oneintra-cluster period, it must satisfy 𝑘 ≥ 𝜆𝑇 . Since 𝑘 can onlybe an integer, the lower bound of 𝑘 is ⌈𝜆𝑇 ⌉.

Lemma 2: If 𝑘 is large enough for collecting all packets ina cycle, the expected collection delay is 𝑇+𝑚⋅𝑘⋅𝜏

2 .Proof: Consider a packet generated by node 𝑖, whose time

slots assigned end at 𝑠𝑖. Since 𝑘 is large enough for collectingall packets in a cycle, this packet must be generated between theend of slot 𝑠𝑖 of two consecutive intra-cluster periods and thisinterval is 𝑇 . Since the packet generation process is Poisson, thetime a particular packet is generated within a fixed interval isuniform [15], thus the expected generation time in this intervalis 𝑇

2 and the expected collection time is 𝑇2 + (𝑚 ⋅ 𝑘 − 𝑠𝑖)𝜏 .

Therefore, the expected collection time for all packets will be

𝐸(𝐷𝑐) =𝑇

2+𝑚 ⋅ 𝑘 ⋅ 𝜏 − 𝐸(𝑠𝑖) =

𝑇 +𝑚 ⋅ 𝑘 ⋅ 𝜏2

(1)

Lemma 2 indicates that the intra-cluster period can be placedat any position in the cycle without affecting the collection

delay, which is only dependent on the cycle length and theperiod duration. It also suggests that 𝑘 can be selected at itslower bound ⌈𝜆𝑇 ⌉ to minimize the collection delay, whichmonotonically increases with 𝑘.

2) Inter-cluster Communication: Inter-cluster communica-tion includes transmissions in the relay network, which consistsof CHs and the sink. For data gathering, as the relay networkis stable, the CHs are organized into a fixed routing tree rootedat the sink at the same time when the clusters are formed. Theconstruction of the routing tree is independent of our schedulingapproach and thus is not discussed in this paper. Within a cycle,each CH is assigned an interval to send all packets, includingpackets collected by itself and packets received from otherCHs, to its parent. The practical length of this interval shouldbe slightly longer than the transmission time of all packets toaccommodate the necessary control packets such as ACK andpotential synchronization errors. However, since our paper isfocusing on the cycle scheduling, we set the length equal tothe transmission time of all packets for simplicity.

The relay network can be viewed as a general wirelessnetwork except that the CH is not available during intra-clusterperiod. We introduce an intra node for each CH to representthe intra-cluster packet collection. Intra node 𝑖 generates 𝑚𝑖 ⋅𝑘packets in each cycle with zero queueing delay before thepackets are sent out. It transmits packets to the CH within aninterval whose duration equals 𝑚𝑖 ⋅𝑘 ⋅𝜏 . The transmission doesnot affect other nodes except for the associated CH. Thus theconsidered problem becomes to schedule intervals for all nodesin the transformed network.

C. Interval Schedule

To obtain an efficient interval schedule, we first present ananalytical model for the problem and then show an illustrationexample. Guided by the example, we will propose our schedul-ing approach.

1) Mathematical Model: The network is represented by agraph 𝐺 = (𝑉,𝐸). 𝑉 is the set of nodes, including the sink, theCHs and the corresponding intra nodes. Denote 𝑝𝑖 as the parentof node 𝑖 in the routing tree and (𝑖, 𝑗) as a transmission linkbetween node 𝑖 and node 𝑗, then 𝐸 = {(𝑖, 𝑝𝑖)∣𝑖 ∈ 𝑉 }. Sinceevery node has a fixed parent, it is easy to see ∣𝑉 ∣ = ∣𝐸∣+ 1.

To model the interference of transmissions, we construct aconflict graph 𝐺′ = (𝑉 ′, 𝐸′). 𝑉 ′ represents all the transmissionlinks in 𝐸. For simplicity, we use 𝑖 to represent link (𝑖, 𝑝𝑖)such that 𝑉 ′ = 𝑉 ∖{𝑣𝑠}, where 𝑣𝑠 represents the sink. 𝐸′

is constructed such that if (𝑖, 𝑗) ∈ 𝐸′, nodes 𝑖 and 𝑗 cannottransmit at the same time due to that the distance between anytwo of nodes 𝑖, 𝑗, 𝑝𝑖 and 𝑝𝑗 is within the transmission radius.Such construction is valid if we assume that the receiver maysend ACK packets and transmissions will not only be affectedby the sender, but also by the receiver. For an intra node 𝑖,there is only one conflict edge (𝑖, 𝑝𝑖) corresponding to the factthat the CH cannot send packets during the transmission of itscorresponding intra node.

The interval scheduling problem is to find a feasible interval(𝑠𝑖, 𝑓𝑖) for each node 𝑖 in 𝑉 ′, where 0 ≤ 𝑠𝑖 ≤ 𝑓𝑖. The cycle

38

length is then set as 𝑡 = max𝑖∈𝑉 ′ 𝑓𝑖. Here we normalize thecycle into time slots with length 𝑘 ⋅ 𝜏 so that the actual cyclelength 𝑇 = 𝑘 ⋅ 𝜏 ⋅ 𝑡. Since the interval equals the transmissiontime of all packets, we have 𝑓𝑖 = 𝑠𝑖 + 𝑛𝑖, where 𝑛𝑖 is themaximum number of packets node 𝑖 sends in a cycle and canbe obtained by

𝑛𝑖 =

{𝑚𝑖 if 𝑖 is an intra node∑

𝑗∈𝐶𝑖𝑛𝑗 if otherwise

Here, 𝐶𝑖 denotes the child set of node 𝑖. For an interval ofnode 𝑖 to be feasible, its transmission link should not conflictwith any other transmission links, thus ∀(𝑖, 𝑗) ∈ 𝐸′, 𝑠𝑖 ≥ 𝑓𝑗or 𝑠𝑗 ≥ 𝑓𝑖.

The end-to-end packet delay can be broken down into trans-mission delay and queueing delay. The transmission delay froman intra node to the corresponding CH is simply the collectiondelay in the intra-cluster period while the transmission delayfrom CH 𝑖 to its parent 𝑝𝑖 is 𝑛𝑖. The queueing delay, definedas the waiting time of a packet at a node before it is sent out ,will be (𝑠𝑝𝑖

− 𝑓𝑖 + 𝑡) mod 𝑡 for a parent 𝑝𝑖. Thus the averagepacket delay is

𝐷 = 𝑘 ⋅ 𝜏 ⋅∑

𝑖∈𝑉 ′ 𝑛𝑖(𝑑𝑖 + (𝑠𝑝𝑖− 𝑓𝑖 + 𝑡) mod 𝑡)

𝑛𝑣𝑠

(2)

where

𝑑𝑖 =

{𝑡+𝑚2 if 𝑖 is an intra node

𝑛𝑖 if otherwise

The optimal scheduling problem was proved to be NP-completeby reducing the K-Colorability problem to the schedulingproblem [16]. Therefore, we will design a heuristic algorithm.Before that, we examine an example to reveal some interestingproperty in the scheduling problem.

2) Example of Chain Topology: This example considers anetwork with chain topology as shown in Fig. 1(a). Each CHhas a corresponding intra node, which generates one packet ina cycle. Thus, node 𝑖 will send 𝑖 packets in a cycle.

Fig. 1(b) shows an intuitive scheduling, where node 𝑖 issequentially assigned an interval of 𝑖 and all intra nodes areassigned an interval of 1 at the beginning of the cycle. This isactually a not-so-bad scheduling as it eliminates the queueingdelay in the relaying: a CH will immediately send out all thepackets upon its receiving from its child. The average packetdelay 𝐷 = 32.5.

The intuitive scheduling does not consider the fact that pack-ets have zero queueing delay on the intra node. An improvedscheduling in Fig. 1(c) schedules the intervals for intra nodesas close as possible to the corresponding CHs. The averagepacket delay 𝐷 = 29.2.

In the improved scheduling, the collection delay on intranode, which is 11.5 according to Eq. (1), takes a large part in thetotal delay. Since the delay mainly depends on the cycle length,we can further reduce the average packet delay by reducingthe cycle length. Fig. 1(d) shows the optimal scheduling thatminimizes the cycle length. In the scheduling, the cycle lengthis exactly the sum of the interval lengths of nodes 4, 5 and 6.

1 2 3 4 5 6

Sink

(a) Chain network consists of 6 CHs.

1

2

3

4

5

6

Cycle

(b) Intuitive scheduling

Cycle

1

2

3

4

6

5

(c) Improved scheduling

1

2

3

4

5

6

Cycle

(d) Optimal scheduling

Fig. 1. An example network with chain topology.

Since these nodes are conflicting with each other, their intervalscannot overlap and thus the cycle length reaches its minimum.On the other hand, the intervals of nodes 1, 2 and 3 are placedat the end of the cycle so that no extra queueing delay isintroduced. The average packet delay 𝐷 = 25.7.

Following this example, we obtain three guidelines to designthe scheduling algorithm.

∙ Interval assignment should follow the order of the nodesin the routing path.

∙ Intervals for intra nodes should be as close as possible tothe intervals for the corresponding CH.

∙ The cycle length can be reduced by overlapping as muchintervals as possible. However, the reduction is bounded bythe intervals that cannot be overlapped due to the conflicts.In fact, although each interval will have a number ofconflicted intervals, it is the longer intervals that decidethe lower bound of the cycle length.

3) Algorithm: The algorithm is summarized in Tables 1 and2. Table 1 describes a basic scheduling algorithm that strictlyschedules the nodes according to their orders in the routingpath: a node is always scheduled before its parent. From thesecond guideline, we see that nodes should be scheduled aslate as possible. Thus, the algorithm actually schedules nodesin the reverse order so that nodes can be scheduled closer tothe end of the cycle. For illustration, we define function 𝐼(⋅)on nodes such that

39

𝐼(𝑖) =

{1 if 𝑖 is scheduled

0 if 𝑖 is not scheduled

During the scheduling, a set 𝑉𝑡 is maintained to include nodeswhose parents are already scheduled. The algorithm tries tofind the earliest possible scheduling interval for all nodes in𝑉𝑡 and the node with the earliest starting time is scheduled.Then the algorithm schedules the next node with no earlierstarting time until all nodes are scheduled or no intervals can bescheduled within the required range. Then the reverse scheduleis obtained.

TABLE 1BASIC SCHEDULING ALGORITHM

Input: node set to be scheduled 𝑉𝑠, schedule range (𝑆, 𝐹 )Output: updated schedule𝑉𝑡 = {𝑖∣𝑖 ∈ 𝑉𝑠, 𝐼(𝑖) = 0, 𝐼(𝑝𝑖) = 1}while 𝑉𝑡! = ∅

order 𝑉𝑡 by distance to sinkfind earliest schedulable interval (𝑠𝑖, 𝑓𝑖), 𝑠𝑖 ≥ 𝑆

for each node in 𝑉𝑡.𝑗 = argmin𝑖∈𝑉𝑡 𝑠𝑖if 𝑓𝑗 > 𝐹 returnschedule node 𝑗 with (𝑠𝑗 , 𝑓𝑗)𝑆 = 𝑠𝑗update 𝑉𝑡

end while

Table 2 utilizes this basic algorithm to perform the actualscheduling. The idea is to first determine a tentative cycle lengthand then try to schedule all the intervals within this cycle. Giventhe third guideline, we start to schedule the intervals from thenodes that are closer to the sink. For assistance, we constructtwo node sets 𝑉𝑛 and 𝑉𝑐. Let

𝑉𝑛 = {𝑖∣𝐼(𝑖) = 0, 𝐼(𝑝𝑖) = 1, 𝑖 ∈ 𝑉 ′}.Initially we assume 𝐼(𝑣𝑠) = 1 so that 𝑉𝑛 includes all nodes thatdirectly send packets to the sink. Clearly, these intervals cannotbe overlapped. In addition, their conflicting intervals cannot beoverlapped with these intervals either. For that, we construct

𝑉𝑐 = 𝑉𝑛 ∪ {𝑗∣(𝑖, 𝑗) ∈ 𝐸′, 𝐼(𝑗) = 0, 𝑖 ∈ 𝑉𝑛}.We then schedule 𝑉𝑐 with the basic scheduling algorithm withno range requirement and obtain the tentative cycle length. Forother nodes that are not scheduled, since they are not in thecurrent 𝑉𝑐, it is guaranteed that their scheduled intervals canbe overlapped with intervals for nodes in 𝑉𝑛. Thus we canschedule the rest of nodes from the beginning of the cycle. Thuswe update 𝑉𝑛 and 𝑉𝑐 according to current schedule and repeatthe basic scheduling algorithm. Since nodes that are closer tothe sink have longer intervals, in most cases the updated 𝑉𝑐

can be scheduled at the beginning part of the cycle, leavingthe rest part of the cycle available for further scheduling. Wethen schedule the rest of nodes to fill in the available part of thecycle to avoid queueing delays. This process is repeated until allnodes are scheduled. The finiteness of this process is guaranteedby the construction of 𝑉𝑛, which guarantees that all children ofalready scheduled nodes will be scheduled in the next iteration.In fact, our experiments show that two iterations will sufficein most cases as the tentative cycle length is large enough forthe rest of nodes to schedule sequentially. Notice that since

the basic scheduling algorithm schedules nodes reversely, theactual schedule should be in the exactly reverse order of theobtained schedule.

TABLE 2ACTUAL SCHEDULING ALGORITHM

Input: graph 𝐺 = (𝑉,𝐸) and conflict graph 𝐺′ = (𝑉 ′, 𝐸′).Output: interval schedules for nodes in 𝑉 ′𝑡 = 0construct 𝑉𝑛, 𝑉𝑐

while 𝑉𝑐! = ∅schedule 𝑉𝑐 with range [0,∞)𝑡𝑐 = max𝑖∈𝑉𝑐 𝑓𝑖if 𝑡𝑐 < 𝑡

𝑉𝑢 = {𝑖∣𝑖 ∈ 𝑉 ′, 𝐼(𝑖) = 0}schedule 𝑉𝑢 with range [𝑡𝑐, 𝑡]

end if𝑡 = max(𝑡, 𝑡𝑐)update 𝑉𝑛, 𝑉𝑐

end whilereverse the whole schedule

4) Analysis: With this cycle-based scheduling, we are alsointerested in determining the maximum packet generation rate.Recall 𝑘 ≥ 𝜆𝑇 . We have

𝑘 ≥ 𝜆𝑇 = 𝑘 ⋅ 𝜆𝜏𝑡,𝜆 ≤ 1

𝜏𝑡.

Therefore, the maximum packet generation rate is 1𝜏𝑡 . On the

other hand, when the generation rate does not exceed themaximum rate, it is always satisfied that 𝑘 ≥ 𝜆𝑇 . Thus, wecan always set 𝑘 = 1 to minimize the packet delay.

IV. ASYNCHRONOUS SCHEDULING APPROACH

In this section, we present the second scheduling approach,which is called New Cluster Scheduling (NCS), adopts an asyn-chronous approach that essentially avoids the synchronizationproblem by introducing a new clustering structure. Next wefirst introduce the new clustering structure and then describethe approach in detail.

A. New Clustering Structure

Instead of designing another algorithm to globally scheduleall the inter-cluster communications for synchronization, NCSattempts to simplify the synchronization by changing the com-munication pattern. To achieve this goal, NCS introduces a newclustering structure, which includes a new type of node: relaynode.

The new clustering structure is illustrated in Fig. 2, in whicha cluster contains a CH node, a relay node and multiplecluster members. The relay nodes always stay in o-state andonly participate in inter-cluster communications. During datagathering, while cluster members still send sensing packets tothe corresponding CH, the CH no longer sends the aggregatedpacket to the next-hop CH but sends to the relay node ofits own cluster instead. Upon receiving the packets, the relaynode further combines them with its own sensing packetsand forwards the packets to the next-hop relay node untilthe packets reach the sink. With such communication pattern,the communication synchronization is greatly simplified. CHscan continue intra-cluster data collection immediately after

40

sending out the aggregated packet, reducing the data collectiondelay. In the meanwhile, inter-cluster communication can beperformed without any restrictions, incurring no waiting delaysfor synchronization. The wireless channel thus can be betterutilized and lower packet delay can be achieved.

Relay node

Member node

Cluster head

Sensing packet

Aggregated packet

Fig. 2. The new clustering structure includes CHs, relay nodes and members.The packet transmissions for the center cluster are shown.

On the other hand, the new clustering structure does notsubstantially increase the complexity of the cluster formationprocess. A network with the new clustering structure can besimply converted from a network with the conventional cluster-ing structure by selecting the member with the highest residualenergy as the relay node in each cluster. The routing algorithmsfor creating routes among different CHs in conventional cluster-based networks can also be utilized to create routes among CHsand relay nodes.

B. Approach Details

NCS adopts the same TDMA protocol as CBS for intra-cluster communications and CSMA protocol for inter-clustercommunications. While member nodes and relay nodes arefixed in i-state and o-state, respectively, CHs still need to switchbetween two states, which is the major task of NCS. Sincethere is no synchronization required among different CHs, thestate switching, or the duration at each state, can be determinedindependently for each CH.

We first determine the inter-cluster duration a CH stays in o-state. When a CH switches to o-state, it cannot transmit a packetimmediately. Consider the case that node 1 is sending a packetto node 2 while the CH, a neighbor of node 2, switches to o-state. If the CH is not in the transmission range of node 1, it willnot detect the ongoing transmission, which would be interruptedby any transmission initiated at the CH before the end of thecurrent transmission. In this case, the protection from RTS/CTShandshake fails as they were not received by the CH who wasin i-state then. We call the period in which such collisions mayoccur the blind period and its duration equals the transmissiontime of a packet of a maximum allowable packet length. Afterthis blind period, the CH then sends the aggregated packet to thenext-hop relay node on the routing path. Once the transmissionsare completed, the CH can immediately switch back to i-stateto continue data collection.

Next we consider the intra-cluster duration of a CH ini-state. Since the CH does not participate in inter-clustercommunications for other CHs, the duration in i-state onlyaffects its own collection delay. Intuitively, to minimize thecollection delay, the CH can switch to o-state immediatelyafter the end of the time frame in which a packet is collected.However, such an approach yields relatively small aggregatedpackets, which underutilizes the wireless channel due to theoverhead of packet headers and control packets, lowering themaximum achievable throughput. Alternatively, we use a fixedcollection duration, denoted as 𝑇𝑐. A larger 𝑇𝑐 indicates lessfrequent data collection, yielding a smaller number of largeraggregated packets. Consequently, the channel is better utilizedand higher throughput can be achieved. On the other hand,a larger 𝑇𝑐 also leads to longer collection delay and hencethe end-to-end packet delay. Therefore, adjusting 𝑇𝑐 can obtaindifferent tradeoffs between the packet delay and the maximumachievable throughput.

𝑇𝑐 determines the number of time slots in an intra-clusterperiod. Following a similar analysis to that in Section III-B, wesee that the necessary number of time slots for a member in anintra-cluster period is 𝑘 = ⌈𝜆(𝑇𝑐+𝑇𝑜)⌉, where 𝑇𝑜 represents theduration of the last inter-cluster period. When 𝑇𝑐 > 𝑚 ⋅ 𝑘 ⋅ 𝜏 , aportion of the intra-cluster period is actually wasted. For energyefficiency, we organize the intra-cluster period into time frameswith each consisting 𝑚 time slot, allowing each node to send apacket in a time frame. Then the CH can remain active only inthe last 𝑘 frames and sleep in other times. The entire processfor a CH is described in Fig. 3.

�������

�������

�������

�������

time frameIntra−cluster

STOP packetSTART packet

periodSleepperiod transmission

PacketBlind

Data

Slot for CH

collection

Intra−cluster period Inter−cluster period

Fig. 3. Timing of data gathering at a CH in NCS.

C. Delay Guarantee

Thanks to the relay nodes, NCS avoids the synchronizationdelays during the inter-cluster communications, allowing therelay network to operate similarly to a general WSN. Therefore,although it does not directly provide delay guarantees, it greatlyfacilitates the utilization of real-time routing protocols at theupper layer, such as SPEED [17], which relies heavily on theone-hop packet delays. These delays could be very long and ir-regular in a cluster-based network where communications incursynchronization delays, degrading the performance of the real-time routing protocols. On the contrast, with NCS eliminatingthe synchronization delays, real-time routing protocols can beeasily implemented to provide optimal performance on delayguarantees.

V. EXPERIMENTAL EVALUATIONS

In this section we evaluate the performance of CBS and NCSthrough ns-2 simulations. For comparison purpose, we also

41

evaluate two existing scheduling approaches used in cluster-based WSNs, for which we first give a brief description.

A. Compared Scheduling Approaches

The first approach we compare is a modified version of thescheduling approach used in IEEE 802.15.4, which we simplycall 802.15.4 in this section. To adapt 802.15.4 in a cluster-based network, it is required to construct a cluster tree fromthe routing tree by adding members as the children of thecorresponding CHs. Each CH or coordinator then maintainsa non-conflicting superframe for its children in the cluster tree.For fair comparison, we assume the same cluster tree as inCBS are used and the length of the superframe for node 𝑖equals 𝑛𝑖, which is the maximum number of packets receivedin a cycle in CBS. The superframes are scheduled using thebasic scheduling algorithm in CBS without considering thenode order in the routing tree. In a superframe, we assumethat each slot is collision-free so that a packet transmission ina slot never fails. When multiple children contend in a timeslot, we randomly select a child to transmit while others waitfor the next slot.

The second approach is a simple synchronous approachthat defines a global frame for all clusters. The global frameincludes intra- and inter-cluster periods. In the intra-clusterperiod, CHs collect data from members using the same protocolas in NCS. At the end of the intra-cluster period, all clustersenter the inter-cluster period simultaneously. The duration ofintra-cluster period is also calculated in the same way as inNCS, while the duration of inter-cluster period 𝑇𝑜 is consideredas a parameter in the experiments. Although the schedulingapproach is quite simple, similar ideas of this global framehave already been adopted in practice [6], [7] and we call thisapproach GF in the following evaluation.

B. Experiment Setup

We first describe the cluster formation algorithm adoptedbefore elaborating other network configurations. According tothe system model in Section III-A, there are no restrictionson the cluster formation algorithm. For simplicity, in ourexperiments we form clusters based on their geographicalpositions, which are assumed already known to the nodes.Specifically, we assume the whole region is a unit square, whichis divided into square cells with side length 𝑙 and the sensorsin the same cell form a cluster. The number of clusters isthen ⌈ 1𝑙 ⌉2. The CH and relay node are randomly selected ineach cluster. The transmission range is set to be

√5𝑙, which

is the maximum distance between two nodes in neighbor cells.Such a range allows nodes within a cell or any two neighborcells to communicate with each other and hence guarantees theconnectivity in all clusters and the relay network.

To evaluate the network performance, we consider two net-works with 300 nodes and 1200 nodes randomly scattered in theunit square. The sink is positioned at a corner of the square tocreate relatively long routing paths. The side length 𝑙 is selectedto be 1

5 and 110 , resulting in 25 and 100 clusters, respectively.

The cluster size ranges from 5 to 18. Some approach dependent

parameters are listed in Table 3. Sensing packets have auniform length of 30𝐵 and the transmission bandwidth is setto 1𝑀𝑏𝑝𝑠. The performance metrics evaluated are packet delayand network throughput. Packet delay is defined as the averageend-to-end delay for all packets received at the sink while thenetwork throughput can be interpreted as the maximum packetgeneration rate with which the network can operate steadily.The evaluation time is set to 100 seconds to obtain the networkperformance at the stable state. Each experiment is repeated 10times to obtain the average value.

TABLE 3PARAMETERS OF APPROACHES.

Approach Parameter 300-node network 1200-node networkNCS 𝑇𝑐 0.2𝑠 2𝑠GF 𝑇𝑜 2𝑠 16𝑠

The inter-cluster communication in NCS and GF utilizesthe common IEEE 802.11 MAC protocol. Practical WSNsmay adopt some simplified versions of 802.11, however, thevariation among these versions only affects the inter-clustercommunications but does not substantially affect the overallperformance evaluation. To construct the routing tree, the CH orthe relay node randomly selects a node in the neighbor cells thatare closer to the sink as its next-hop. The aggregated packets atthe source are not further aggregated at the intermediate nodesin the routing tree.

C. Performance Results

Fig. 4 shows the average packet delay among four approacheswith allowable packet generation rates. We first examine theresult in the network with 300 nodes. We can observe thatNCS yields the shortest packet delay when the network isnot saturated under lower packet generation rates. Due to theintroduction of relay nodes, packets are transmitted quickly ateach hop without undertaking any extra delay caused by thestate switching. On the opposite, GF, which also uses 802.11for inter-cluster communications, yields the longest packetdelay. In GF, the synchronization of the inter-cluster periodsfor different CHs incurs many concurrent packet transmissionswith high contentions, eventually resulting in long delay. TwoTDMA based approaches have shorter delay than GF since theycompletely avoid the transmission contention. However, sincethe transmissions in these two approaches are strictly scheduled,packets inevitably incur some queueing delay before they canbe relayed by the intermediate nodes in the routing path. Thusboth have longer delay than NCS. In particular, delay in CBSis about 30% shorter than that in 802.15.4, due to the efficientdesign of the cycle scheduling. While the cycle length of twoapproaches does not have much difference, packets in CBSendure less queueing delay with the ordered interval scheduling.

The performance comparison is similar in the network with1200 nodes, where GF exhibits poor performance with al-lowable generation rate under 0.05, which was not shownin the figure. While NCS still shows the shortest delay, weobserve that CBS obtains a higher performance gain comparedto 802.15.4, whose delay is nearly twice of CBS. This contrast

42

0 1 2 3 40

0.5

1

1.5

2

2.5

Packet generation rate

End

−to

−en

d de

lay

(s)

CBSNCS802.15.4GF

0.2 0.4 0.6 0.8 11

2

3

4

5

6

Packet generation rate

End

−to

−en

d de

lay

(s)

300−node network

1200−node network

Fig. 4. End-to-end packet delay of four scheduling approaches under differentpacket generation rates.

indicates that the scheduling in CBS enjoys more benefits inlarger-scale networks.

Fig. 5 shows the network throughput, or the maximumpacket generation rate for four approaches in both networks.GF has the worst throughput, besides its longest delay asseen in Fig. 4. This is because that GF requires a long inter-cluster period to accommodate long delay and therefore causeslow allowable packet generation rate. NCS also has lowerthroughput compared to two TDMA-based approaches. Thisis mainly due to the intrinsic of CSMA-based 802.11 protocol,which spends much longer time than the actual transmissiontime in transmitting packets. For two TDMA-based approaches,CBS slightly outperforms 802.15.4. The similar performance isdue to the fact that the cycle length, a dominating factor forthe throughput, is similar in both approaches.

CBS NCS 802.15.4 GF0

0.5

1

1.5

2

2.5

3

3.5

4

Pac

ket g

ener

atio

n ra

te

300−node1200−node

Fig. 5. Network throughput with four scheduling approaches.

VI. CONCLUSIONS

In this paper, we have presented two communication schedul-ing approaches CBS and NCS to solve the cluster com-munication synchronization for delay-sensitive data gatheringin WSNs. In CBS, a cycle based schedule for each CH isconstructed based on the pre-determined routing tree. CBSminimizes the cycle length while maintaining the node orderin the routing tree, which minimizes the intra-cluster collectiondelay and allows continuous packet forwarding from the source

to the sink. In NCS, a CH-relay-member structure is proposedto replace the conventional CH-member structure. The introduc-tion of relay nodes releases the CHs from the heavy burden ofpacket relaying so that the intra- and inter-cluster communica-tions can be performed more efficiently. Our simulation resultshave shown that the proposed approaches exhibit much betterperformance than existing scheduling approaches in terms ofpacket delay and throughput. Given the distinct design of CBSand NCS, these two approaches are suitable for delay-sensitivedata gathering in a wide range of applications in WSNs.

ACKNOWLEDGMENTS

This work was supported by US National Science Foundationunder grant number ECCS-0801438 and US Army ResearchOffice under grant number W911NF-09-1-0154.

REFERENCES

[1] G. Gupta and M. Younis, “Load-balanced clustering in wireless sensornetworks,” Proc. of IEEE ICC, May 2003.

[2] W.R. Heinzelman, A. Chandrakasan and H. Balakrishnan, “Energy-efficient communication protocol for wireless microsensor networks,”Proc. of the 33rd Annual Hawaii International Conference on SystemScience, 2000.

[3] O. Younis and S. Fahmy, “HEED: A hybrid, energy-efficient, distributedclustering approach for ad hoc sensor networks,” IEEE Trans. MobileComputing, vol. 3, no. 4, pp. 366-379, 2004.

[4] Z. Zhang, M. Ma and Y. Yang, “Energy efficient multi-hop polling inclusters of two-layered heterogeneous sensor networks,” IEEE Trans.Computers, vol. 57, no. 2, pp. 231-245, 2008.

[5] A.A. Abbasi and M. Younis, “A survey on clustering algorithms forwireless sensor networks,” Computer Communications, vol. 30, no. 14-15,pp. 2826-2841, 2007.

[6] M. Lotfinezhad, B. Liang and E.S. Sousa, “Adaptive cluster-based datacollection in sensor networks with direct sink access,” IEEE Trans. MobileComputing, vol. 7, no. 7, pp. 884-897, 2008.

[7] H.J. Choe, P. Ghosh and S.K. Das, “QoS-aware data reporting control incluster-based wireless sensor networks,” Computer Communications, vol.33, no. 11, pp. 1244-1254, 2010.

[8] L. Shi and A.O. Fapojuwo, “TDMA scheduling with optimized energyefficiency and minimum delay in clustered wireless sensor networks,”IEEE Trans. Mobile Computing, vol. 9, no. 7, pp. 927-940, 2010.

[9] F. Cuomo, E. Cipollone and A. Abbagnale, “Performance analysis ofIEEE 802.15.4 wireless sensor networks: An insight into the topologyformation process,” Computer Networks, 2009.

[10] W. Ye, J. Heidemann and D. Estrin, “An energy efficient MAC protocolfor wireless sensor networks,” Proc. of IEEE INFOCOM, 2002.

[11] G. Lu, B. Krishnamachari and C.S. Raghavendra, “An adaptive energy-efficient and low-latency MAC for tree-based data gathering in sensornetworks,” Wireless Communications & Mobile Computing , vol. 7, no.7, pp. 863-875, 2007.

[12] J. Kim, X. Lin, N.B. Shroff and P. Sinha, “On maximizing the lifetimeof delay-sensitive wireless sensor networks with anycast,” Proc. of IEEEINFOCOM, 2008.

[13] J. Kim, X. Lin and N.B. Shroff, “Optimal anycast technique for delay-sensitive energy-constrained asynchronous sensor networks,” IEEE/ACMTrans. Networking, vol. 19, no. 2, pp. 484-497, 2011.

[14] W. Ye, J. Heidemann and D. Estrin. “Medium access control withcoordinated adaptive sleeping for wireless sensor networks,” IEEE/ACMTrans. Networking, vol. 12, no. 3, pp. 493-506, 2004.

[15] P.A.W. Lewis and G.S. Shedler. “Simulation of nonhomogenous Poissonprocesses with log-linear rate function,” Biometrika, vol. 63, no. 3, pp.501-505, 1976.

[16] S.C. Ergen and V. Pravin. “TDMA scheduling algorithms for wirelesssensor networks,” Wireless Networks, vol. 16, no. 4, pp. 985-997, 2010.

[17] T. He, J.A. Stankovic, C. Lu and T. Abdelzaher, “SPEED: a statelessprotocol for real-time communication in sensor networks,” Proc. ofICDCS, 2003.

43