Data Aggregation Using Dynamic Selection of Aggregation Points ...

Wireless Pers Commun (2015) 80:611–633 DOI 10.1007/s11277-014-2031-5

Data Aggregation Using Dynamic Selection of Aggregation Points Based on RSSI for Wireless Sensor Networks Azlan Awang · Shobhit Agarwal

Published online: 5 September 2014 © Springer Science+Business Media New York 2014

Abstract In wireless sensor networks (WSNs), due to dense deployment, sensory data gathered by sensor nodes in close proximity tend to exhibit high correlation and therefore redundant. Transmitting such redundant data is not practical in the energy-constrained WSNs. Data aggregation offers a key solution to reduce such redundancy by allowing intermediate nodes to aggregate raw data streams before routing them toward a sink node. This in turn reduces transmission energy consumption. Prior work in data aggregation often rely on node’s location for selecting an aggregator node, a fusion point. In this work, we propose two data aggregation mechanisms where aggregator nodes are determined opportunistically without dependency on global knowledge of data flow, network topology and nodes’ geographical location. These mechanisms aggregate and route data packets based on Received Signal Strength Indicator (RSSI). An aggregation identification (Agg_ID) is associated with each data packet generated by a sensor node. The RSSI and Agg_ID are used in the RSSIBased Fowarding for favoring nodes closer to sink to be an aggregator and also a relay node. We show via simulation the performance of the proposed mechanisms in terms of normalized number of transmissions, total number of packets transmissions and receptions, average energy consumed per data packet, network lifetime, end-to-end delay and packet loss probability. Keywords RSSI · Data aggregation · CTS response · Structure-free · Cross-layer · Network lifetime

A. Awang (B) · S. Agarwal Electrical and Electronic Engineering Department, Universiti Teknologi PETRONAS, Bandar Seri Iskandar, 31750 Tronoh, Perak, Malaysia e-mail: [email protected]; [email protected] S. Agarwal e-mail: [email protected]

123

612

A. Awang, S. Agarwal

1 Introduction Wireless sensor networks (WSNs) are widely recognized as powerful means for in situ observations of events and environments over long period of time [3,4]. A sensor network consists of small sensor nodes usually deployed in dense within a region of interest. Since nodes are typically powered with non-rechargeable batteries, energy is limited. Often the sensory data gathered by the sensor nodes are spatially and temporally correlated [3,4]. Transmitting such correlated and redundant data to a sink node is not practical in the resource constrained WSNs. Data aggregation offers a key solution to curtail the network load and hence reduces redundant packet transmission. This in turn reduces transmission energy consumption in WSNs. In data aggregation, raw data from multiple sensor nodes are combined into an aggregated packet at an aggregator node before relaying it to a sink node, as illustrated in Fig. 1a. The data aggregation mechanism consists of two main tasks: data processing at intermediate nodes followed by data routing as shown in Fig. 1b. Data processing consists of the task to combine raw data using an aggregation function.1 This requires the availability of data from multiple nodes at an aggregator point, at a particular instance of time. The availability of data from different source nodes at a single aggregator node requires convergence of data from different locations. This is termed as spatial convergence [12]. The availability of data from multiple nodes also requires convergence of data in time, i.e., temporal convergence. The spatial and temporal convergence of data are two necessary conditions for the effectiveness of the data aggregation process. This concept will be elaborated in Sect. 2. Synchronization of data transmission is essential for convergence of data at an aggregator node. Generally, a sink node or an aggregator node is responsible for synchronizing transmissions between nodes. An aggregator node (or an aggregation point) is usually elected based on some parameters, e.g., the number of neighbors, residual energy and location of the nodes. It is responsible for aggregating data from multiple source nodes. Most of the existing data aggregation protocols are based on the idea of organizing nodes in the network in a structured way where the optimal location of an aggregator node can be determined. However, this is not possible without the global knowledge of data flow and underlying network topology. In WSNs, the network topology in a region of interest is dynamic in nature where some nodes may enter sleep mode to conserve energy or may run out of energy. Therefore, routing data packets over a fixed structure is difficult as it requires periodic reconstruction of the routing structure that is used for aggregation. In this work, we propose two topology independent data aggregation mechanisms, namely, Class-based CTS Response and Alpha-based CTS Response data aggregation mechanisms. These mechanisms are based on the idea that building a structure for optimal data aggregation is not possible without some global knowledge of the network. Thus in these mechanisms, data aggregation points are selected opportunistically in such a way that nodes closer to sink and having similar Agg_ID are prioritized more than the others to be an aggregator and a relay node. The proposed data aggregation mechanisms in this article are part of the future work highlighted in RSSI-Based Forwarding (RBF) [5,6]. In [5,6], no data aggregation is implemented. RBF is a cross-layer Medium Access Control (MAC) and routing protocol for WSNs that uses an RSSI value each node maintains with respect to a sink node. RBF uses the four-way 1 Choice of aggregation function depends on application under consideration. Some common aggregation

function are summation (SUM), average (AVG), maximum (MAX), minimum (MIN), COUNT. Detailed discussion about aggregation function can be found in [9,14,24,27].

123

Dynamic Selection of Aggregation Points

613

Environment Data Acquisition 1

Sensor Nodes

2 3

7

f{1,2,3,4} 8 f{6,7}

10 9

f{5,8,9}

f{10,11} 12

5

Aggregator Node

9

Source Node Flow of data

Data Aggregation

5 4

6

Data Collection/ Routing Data processing + Data aggregation at intermediate Sensor Nodes Data Collection/ Routing

11 Sink

(a)

Sink

(b)

Fig. 1 Illustration of data aggregation mechanism in multihop wireless sensor networks a Data gathering at aggregator nodes and relaying toward a sink node. b The components of data aggregation

Request To Send (RTS)–Clear To Send (CTS)–DATA–Acknowledgment (ACK) handshake for joint access and routing toward a sink node. The Enhanced CRT (CRT, CTS Response Time) mechanism proposed in [5,6] favors forwarding candidates closer to sink to be a relay node, i.e., with higher probability to respond first a CTS packet to the sender node that initiates an RTS transmission. The class-based and alpha-based data aggregation mechanisms proposed in this work, uses an aggregation ID (Agg_ID), associated with each sensed data. This Agg_ID is used to prioritize a CTS response in RBF such that data packets with the same Agg_ID are aggregated together when the routes meet. In Class-based CTS Response, priority is given to forwarding candidates having similar Agg_ID by allowing them to contend using a window of higher priority time slots (class of contention window) when responding a CTS packet. The contention window classification proposed in [12] requires node’s distance calculation from the sink using coordinates as part of the aggregation process. Differ from [12], our approach does not require location awareness to prioritize a CTS response. In Alpha-based CTS Response, the parameter α as defined in RBF [5,6] is used as a key parameter for prioritizing a CTS response as part of the process to select an aggregator and also a relay node. The remainder of this paper is organized as follows. Section 2 provides a review of prior work in data aggregation. We also highlight some important points regarding the effectiveness of data aggregation process. The spatial and temporal convergence of data are discussed. In Sect. 3, we present our proposed class-based and alpha-based data aggregation mechanisms. Section 4 describes the network model and simulation parameters considered in this work, followed by the simulation results and discussion. Finally, in Sect. 5, we conclude our work with a summary of the results and the suggested future work.

2 Related Work In WSNs, the cost of transmitting a bit is much higher than the computation cost [10,18]. Therefore, reducing communication cost improves energy efficiency in the network. Thus data aggregation is considered as one of the primary techniques to reduce transmission energy

123

614


consumption that leads to a prolonged network lifetime. Various data aggregation protocols for WSNs have been proposed in the past. These protocols are classified based on network architecture (flat, hierarchical networks), network flow-based, or to meet certain quality of service (QoS) requirement [26]. The authors in [22] classify the routing schemes with data fusion into different categories: routing-driven algorithms, coding-driven algorithms and fusion-driven algorithms. Some authors in [12] classify network architecture based data aggregation techniques into: structure-based and structure free. In structure-based data aggregation protocols, nodes are aware of their next-hop nodes for routing data. Generally, nodes are organized in a fixed structure, i.e., a routing path from a source node to a sink node is established even when a node does not have data for transmission. These protocols are classified based on organization of nodes in the network structure: flat networks and hierarchical networks. In flat networks, all the sensor nodes behave in a similar fashion (all nodes can aggregate and forward data) and have approximately the same amount of energy. In these protocols, data centric routing is used to propagate query in the network, which facilitates data aggregation. Generally, sink node or base station takes initiative of starting the aggregation process by propagating a query. Directed diffusion [17], rumor routing [7], SPIN [19] are some of the data aggregation protocols for flat networks. These protocols use flooding algorithm [29] which increases message duplication and wastage of bandwidth. Aggregation protocols for hierarchical networks are further classified as either clusterbased, tree-based or chain-based protocols. Low Energy Adaptive Clustering Hierarchy (LEACH) [16] is a distributed cluster-based data aggregation protocols and one of the first protocols based on hierarchical organization of nodes. It combines the idea of energy-efficient clustered routing along with medium access. Over the years, researchers have proposed many variants of LEACH like Low Energy Adaptive Clustering Hierarchy-Centralized (LEACHC) [16], Vice-Low Energy Adaptive Clustering Hierarchy (V-LEACH) [34], Hybrid Energy Efficient Distributed Clustering (HEED) [35], Energy-Low Energy Adaptive Clustering Hierarchy (E-LEACH) [33], to overcome the drawback associated with LEACH. Protocol like Voting based Clustering Algorithm (VCA) [25] uses multiple criteria to optimize cluster head (CH) selection. However, all these protocols require global knowledge for CH selection, without which the clusters produced by them are not uniformly distributed. In tree-based data aggregation protocols, data is forwarded over a tree structure and are good for in-network aggregation and query-based approaches. Power Efficient Data gathering and Aggregation Protocol (PEDAP) [28], Energy-Aware Distributed Heuristic (EADAT) [11], Tiny AGgreggation (TAG) [23], Dynamic and Scalable Tree data aggregation (DST) [30] are some of the tree-based protocols. These protocols require periodic reconstruction of tree and have low resilience to link failures. Power-Efficient Gathering in Sensor Information Systems (PEGASIS) [20] is a chain-based protocol and suffers from increased end-to-end delay. In structure-free data aggregation protocols, nodes are not aware of their next-hop nodes, i.e., there are no fixed routing paths in the network. Next-hop node is determined only at the time when data need to be transmitted. In [12], Data-Aware Anycast (DAA) with Randomized Waiting (RW) was the first structure-free protocol proposed for data aggregation. DAA associates the sensed data with an ID and uses anycasting [36] in order to forward data toward the sink, such that data packets with the same ID are aggregated. It prioritizes a CTS response used in anycasting based on the assigned ID and location of nodes to increase spatial convergence. Location of nodes can be calculated by equipping the nodes with a global positioning system (GPS) receiver or by using some localization algorithm [8,21]. We may argue that equipping each node with a GPS receiver is not practical, as it increases

123

Dynamic Selection of Aggregation Points Fig. 2 Illustration of spatial convergence of data due to exchange of Agg_ID along with RTS in a multi-hop WSNs

615

1

Sink

2

I 5

4

1

Sink

5

2

II

1

Sink

3

5

4

2

III

3

3

4

the cost of sensor nodes, whereas, determining the location of sensor nodes by localization algorithms consumes extra energy. Tree-on-DAG (ToD) (DAG, Directed Acyclic Graph) [13] is a semi-structured data aggregation protocol. This protocol uses DAA in phase one to aggregate the data followed by dynamic forwarding on a structure, ToD, in phase two. The two level tree structures with dynamic forwarding increases scalability of the network. However, it uses location aware routing which requires each node to be equipped with GPS transceivers or a localization protocol [8,21] to be executed before ToD can work, thus increasing cost of the network. In [32], authors have proposed a prediction-based data aggregation scheme for environmental monitoring. It exploits temporal correlation of sensed data to reduce total number of transmissions between sensor nodes, which results in enhanced network lifetime. This protocol uses Kalman filtering and Grey Model to predict data values. In [15], authors have proposed a feature selection based data aggregation protocol using particle swarm optimization (PSO) and back propagation neural networks (BPNN). In this protocol, PSO was used to select the features that are being acquired from different surroundings in WSNs, and BPNN was used to create prediction model for the same surroundings. Both of these protocols are complex in nature and require extra transmissions. The effectiveness of the aggregation operation is often measured in terms of aggregation ratio [22]. This is defined as the ratio of outgoing data to that of incoming data at an aggregation point. When an outgoing packet length is the same as each incoming packet length, full aggregation occurs with an aggregation ratio of 1/l, where l is the number of incoming packets. When the outgoing packet length equals the summation of all incoming packet lengths, zero aggregation occurs. In most applications, partial aggregation is observed. The amount of reduction depends on the degree of data correlation determined by the monitored and aggregation processes. As discussed earlier, spatial and temporal convergence of data are two necessary conditions for the effectiveness of the data aggregation. Figure 2 shows an example of spatial convergence of data, illustrated in three steps: step I, II and III. In this figure, a colored circle represents node with data whereas a white circle represents nodes without data and the available routing paths are indicated by arrows. In step I, node 2 sends its data to node 1 as no other alternative route exists. In step II, data from node 3 can either be sent to node 2 or 4 based on which

123

616

A. Awang, S. Agarwal 1

2

3

4 Sink

Fig. 3 Illustration of temporal convergence of packet forwarding from node 1 to sink. The packet is aggregated along the route by increasing temporal convergence

Frame Control

RSSI

Agg_ID

Receiver Address

Transmitter Address

FCS

Fig. 4 Format of RTS packet. A sending node includes its aggregation ID (Agg_ID) in the RTS transmission

node wins the contention process to be a relay node. However, data will only be aggregated if node 4 is selected as the next-hop. If node 4 is informed by an incoming control packet transmission from node 3 that there will be some data with similar aggregation ID to be aggregated, we can then prioritize node 4 to become a next-hop relay node instead of node 3. Thus spatial convergence of data occurs if node 4 becomes a relay node and data packets can be aggregated at node 4 in step III. In a multi-hop network, aggregation occurs only if a node receives data from its predecessors prior to transmitting its own data. This often requires nodes to delay (aggregation delay) the transmission of their own data. During this delay, a node gathers data from its predecessor nodes. Figure 3 shows an example of data being forwarded to sink in a multi-hop network. In this network, each node delays its data transmissions and delayn represents the aggregation delay of a node n. For example, data aggregation occurs at node 2 only if data from node 1 reaches node 2 before it transmits its data, thus delay2 should be greater than delay1 . Similarly, for data aggregation to occur at node 3, it requires the data from node 2 to reach node 3 before it transmits its data, thus delay3 should be greater than delay2 . Thus for aggregating data packets coming from nodes 1, 2 and 3 at node 4, data from all the nodes must be available at node 4 at the same time. Therefore, to improve temporal convergence of data, nodes closer to sink should have higher aggregation delay than nodes that are away from the sink. For example, as illustrated in Fig. 3, node 4 should have maximum delay and node 1 should have minimum delay (delay4 > delay3 > delay2 > delay1 ) in order to have temporal convergence.

3 RSSI-Based Data Aggregation Mechanisms This section describes the class-based and alpha-based data aggregation mechanisms. In these mechanisms, a next-hop node for each transmission is determined based on path loss and whether or not the node has data packets for aggregation. In addition to RSSI (or path loss) field as defined in RBF [5,6] (see Fig. 4), an Agg_ID is included in an RTS packet. In this work, data packet generation time is used as an Agg_ID and this is associated with each data packet. Some other parameters can also be chosen as an Agg_ID depending on application and usage. When a node has data packet to send, it broadcasts an RTS packet containing its RSSI value and the Agg_ID. At an aggregator node, data packets with similar Agg_ID are aggregated into one packet. Since an Agg_ID is already incorporated in an RTS packet used in RBF, no extra cost incurred to have another control packet.

123


617

For this paper to be self-explanatory, we briefly describe the RBF protocol [5,6]. A node responds to an RTS packet only if its RSSI value is greater than that of the sender (or if the path loss ratio between the forwarding candidate and the sender is less than one). In response to an RTS transmission from source node i having path loss L i , forwarding candidate j with path loss L j sends a CTS packet by selecting a slot k within a contention window of size W , based on probability pk = q2 q1k , where q1 and q2 , as defined in [5,6] are given in (1) and (2) for convenience. 1 − b2 L j 1 α (1) q1 = b + b Li r 1 − q1 q2 = (2) 1 − q1W where, α ∈ (0, 1], b ∈ (0, 1) are tunable constants and r is the percentage of node’s residual energy. For further details of the effects of b and path loss on the probability pk , readers can refer to [5,6]. In this work, we assume r = 1 and full aggregation is achievable at an aggregator node. In [5,6], no data aggregation is considered. In this work, α in (1) is used as part of the aggregation mechanism in alpha-based data aggregation which will be explained in details in Sect. 3.2. In RBF, data packets routed from a node may take different paths to reach a sink node depending on the path loss of a node. Therefore, to increase spatial convergence (the availability of data from different locations at one node) of data, we extend the approach used in [12]. In our proposed mechanisms, Agg_ID is carried in the RTS packet transmission, thus the exchange of RTS-CTS informs the nodes about data packet that is going to be transmitted. The use of Agg_ID together with RTS-CTS mechanism in RBF leads to spatial convergence of sensed data. To enhance temporal convergence, the proposed mechanisms use variable delay based approach. This delay is implemented as a function of RSSI (path loss) with respect to sink node, specifically, aggregation delay is considered as an inverse function of path loss [adopted from (1) and is given in (3)]. Thus nodes that are farther away from sink will have smaller delay (larger path loss or smaller RSSI) and nodes that are closer to the sink will have larger delay (smaller path loss or larger RSSI). Figure 5 shows the variation of delay with respect to RSSI using (3). RSS In α 1 − b2 delayn = b + (3) b RSS Imin where, α ∈ (0, 1], b ∈ (0, 1) are constants, RSS In is the RSSI value of node n and RSS Imin is the minimum RSSI value. 3.1 Class-Based CTS Response Data Aggregation Mechanism In this mechanism, CTS response is prioritized by dividing the contention window into groups, namely, Group A [0, W A − 1] and Group B [W A , W − 1], as shown in Fig. 6. Group A time slots are reserved for intermediate nodes (forwarding nodes) that have similar Agg_ID as the sending nodes, whereas, the rest of the contenders participate in contention by random selection of a time slot from Group B. Appropriate value of W A is determined in such a way that the best value of energy efficiency is obtained. This will be explained later but in order to understand its principles we give first the description of the class-based data aggregation approach.

123

618

A. Awang, S. Agarwal 0.06

Aggregation Delay (s)

0.05

0.04

0.03

0.02

0.01

0 −25

−55

−64.03 −69.31 −73.06 −78.34 −80.35 −83.62 −84.86

RSSI values (dBm) Fig. 5 Variation of aggregation delay (delayn ) versus RSSI values using (3). Aggregation delay for each node is inversely proportional to its path loss

0

1

...

...

WA-1

WA

...

...

Group A

...

...

W-1

Group B

Fig. 6 Time slots (contention window) are divided in two groups. Contenders with the same Agg_ID participate in contention by selecting a time slot within Group A and the rest choose a time slot from Group B

3 4

R

10

8 5

2

12

Sink

6

11

9 13

1

14

7 R

Fig. 7 A scenario to illustrate data packet aggregation and forwarding toward the sink node. Node 5 has data packet and broadcasts it with an Agg_ID

In [5,6], a forwarding candidate may choose a time slot based on uniform distribution of time slots [Uniform CTS Response Time (CRT)] or Enhanced CRT if a time slot is selected based on probability pk using (1) and (2). As such, we classify in similar ways this classed-based data aggregation into two specific mechanisms: Class-based uniform CRT Data Aggregation (class-uniCRT-Agg) and Class-based Enhanced CRT Data Aggregation (class-enCRT-Agg).

123


619

As an illustration, in Fig. 7, an event occurs that triggers nodes 5 and 9 at the same time. Since data packets are generated at the same time, the sensed data are assigned with the same Agg_ID. Nodes 5 and 9 attempt to broadcast an RTS if the medium is idle for a DIFS time and if their network NAVs are zero. Suppose that node 5 senses the medium to be idle first, it then broadcasts an RTS packet. Forwarding candidates that receive the RTS compare their RSSI value with the one in the RTS packet. If their RSSI value is higher, they then compare the received Agg_ID with the IDs within their Agg_ID list. Those with higher RSSI value participate in contention to respond a CTS packet by participating in contention with random selection of a time slot from Group A (if they have Agg_ID similar to that in RTS) or Group B, otherwise. Other nodes having lower RSSI value such as nodes 1, 2 and 3 are put to sleep mode. We suppose in this figure that nodes 4, 6, 7, 8 and 9 are eligible to participate in contention to respond a CTS packet. However, using (1) and (2), we favor nodes closer to sink to respond first with CTS. In this example, nodes 8 and 9. Since node 9 has the same Agg_ID as node 5, it is given a priority to choose randomly a time slot from Group A and node 8 will choose one from Group B. If a CTS packet is not received after an RTS timer is set in node 5, it will transmit an RTS again. When node 5 receives a CTS successfully, DATA and ACK packet exchange follow between the two specific nodes. Node 9 aggregates the data packet with its own data and waits for the random aggregation delay (depending on path loss of the node). Within an aggregation delay limit, a forwarding node is able to gather more data packets coming from its neighbors before attempting to transmit the data further. The process is repeated until the data is transferred to sink node.

Algorithm 1 gives the pseudo code for the algorithm of a sending node. In the mechanism, RTS_retrans _count is used to track the number of retransmissions of an RTS packet. Maximum retransmissions are limited by RTS_retrans_limit. Similarly, data_retrans_count helps to track retransmissions of a data packet and data_r etrans_limit gives maximum number of data packet retransmissions. Algorithm 2 provides the pseudo code for the algorithm of a forwarding candidate. The window size in Group A, W A is determined via simulation that best meets the energy efficiency. We define energy efficiency as the inverse of energy used per data packet. Thus, a lower value of energy used per data packet refers to a more energy efficient communication. A small W A limits the choice of time slots selection and this may increase the probability of packet collisions as several contending nodes may select the same time slot. This in turn increases packet retransmissions and hence increases energy consumption. Endto-end (ETE) delay may also increase as nodes may find the wireless medium is occu-

123

620


pied due to retransmissions before it could attempt again its RTS transmission. A large W A may lead to unnecessary back off and this introduces higher ETE delay as well as energy consumption. Furthermore, nodes may not always have data packets to aggregate. Smaller number of data packets available for aggregation increases transmissions energy cost.

To determine the best value of W A , a simulation scenario consisting of uniformly distributed nodes was designed using Castalia [2]. The simulations were runs for 30 iterations using different random seeds and for a simulation duration of 100 s. The impact of varying network sizes is studied (number of nodes equals 20, 40, 60 and 80). Figure 8a shows the graph for total energy consumed in transmission and reception, whereas, graph for average energy consumed per data packet delivered to the sink node is shown in Fig. 8b for different W A values. From these figures, we observe as W A is increased from 2 to 10, energy consumed in transmission and reception as well as energy consumed per data packet decreases. This occurs because with increase in the number of time slots in Group A, the

Tx energy class−uniCRT−Agg Rx energy class−uniCRT−Agg Tx energy class−enCRT−Agg Rx energy class−enCRT−Agg

5.2 5 4.8 4.6 4.4 4.2

Avg. Energy Consumed per Data Packet (Avg J/Data Pkt)

Energy Consumed in Transmission and Reception

−4

8.6 x 10

5.4

Class−uniCRT−Agg Class−enCRT−Agg

8.4 8.2 8 7.8 7.6 7.4

2

5

10

15

Number of Slots in Group A

(a)

20

25

2

5

10

15

20

25

Number of Slots in Group A

(b)

Fig. 8 The impact of varying W A on the a transmission and reception energy consumption, and b energy efficiency for class-uniCRT-Agg and class-enCRT-Agg data aggregation mechanisms

123

0 1 2 3 4 5 6 7 8 9 Time slot, n

(a)

05

0.3 0.25 0.2 0.15 0.1 0.05 0

621

Probability

0.3 0.25 0.2 0.15 0.1 0.05 0

Probability

Probability


0 1 2 3 4 5 6 7 8 9 Time slot, n

(b)

0 75

0.3 0.25 0.2 0.15 0.1 0.05 0

0 1 2 3 4 5 6 7 8 9 Time slot, n

(c)

10

L Fig. 9 The impact of varying α on the probability distribution pk = q2 q1k for L j = 0.05, b = 23 , W = 10. i

The probability to choose a lower time slot to respond a CTS packet increases as α increases (adopted from [5,6]) a α = 0.5 b α = 0.75 c α = 1.0

probability of forwarding candidates to select the same time slot decreases, thereby, reducing the number of possible collisions between aggregated packets. Total number of packets transmissions and receptions are reduced when more data packets are aggregated within an aggregation delay limit which in turn reduces total transmission and reception energy consumption. When W A value is varied greater than 10, the total energy consumed in transmission and reception and average energy consumed per data packet start to increase. This is because nodes take longer time for sending a CTS response and this leads to smaller number of data packets being available for aggregation. Thus number of aggregations is reduced which in turn increases the number of transmissions and receptions. According to our analysis, the best value of W A is found to be 10 for a contention window size W equals 64. 3.2 Alpha-Based CTS Response Data Aggregation Mechanism In this section, we describe the Alpha-based CTS response data aggregation mechanism (alpha-enCRT-Agg). In contrast with the class-based aggregation mechanism described earlier, alpha-enCRT-Agg mechanism does not divide the contention window into groups. Instead, the mechanism extends the use of parameter α in (1) (α ∈ (0, 1]), as defined in [5,6] as a key parameter for the aggregation mechanism. Specifically, α is implemented as a function of the number of data packets with the same Agg_ID gathered at a forwarding candidate. Forwarding candidates with higher number of data packets having similar Agg_IDs will have higher value of α. Higher value of α increases the probability to select a smaller time slot for sending a CTS packet using (1) and (2). The impact of α on probability pk using (1) and (2) is described as follows. When a forwarding candidate j is closer to sink as compared to a sending node i, path loss ratio is very small. In this case, the value of q1 increases with an increase in α value. This increases the probability pk of forwarding candidates to select lower time slots, as illustrated in Fig. 9. Thus nodes with more number of data packets having the same Agg_ID (higher α) will select lower time slots. In this mechanism, each forwarding candidate assigns an initial value of α to data packets with different Agg_IDs. All forwarding candidates maintain a list of α values corresponding to unique Agg_IDs (maps unique Agg_IDs to α values). Every time a data packet is received, the forwarding candidate checks whether or not it has data packets associated with the same Agg_ID within its data list. If it does not have any data associated with the received Agg_ID, it updates its list and assigns an initial value of α. However, if data packet with the same Agg_ID already exists, it increments the corresponding value of α by a constant c, such that α ∈ (0, 1].

123

622


The aggregation technique in alpha-based CTS response mechanism is almost similar to class-based CTS response mechanism stated above. The main difference between the two mechanisms is in the process of selecting a time slot within the contention window. In alpha-enCRT-Agg, each node maintains a list of α values along with the list of Agg_IDs. To illustrate this aggregation process, we consider the network example given in Fig. 7. When nodes 4, 6, 7, 8 and 9 receive an RTS packet from node 5, they check within their Agg_ID list for data packets associated with that Agg_ID. Since node 9 has data packet with the same Agg_ID, therefore, the value of α for node 9 is greater than all other nodes. Thus, using (1) and (2), it selects a lower time slot as compared with other contending nodes. After data packet is received by node 9, it aggregates the received data with its own data and updates the corresponding value of α in its list. Algorithm 3 gives the pseudo code for the algorithm of a forwarding candidate. Pseudo code for a sending node is similar to the one described in Algorithm 1 (for class-based enhanced CRT). The only difference is when a node assigns an Agg_ID to sensed data, it maps the Agg_ID with α and updates the corresponding α values in the alpha list.

4 Simulation Results 4.1 Simulation Model and Parameters In this study, we consider a wireless data collection model that consists of large number of sensor nodes and one sink node. Sensor nodes generate data periodically. Each sensor node is assumed to have a limited amount of energy and a fixed transmission power, hence data packets are sent to the sink via multi-hop communications. Sink is assumed to have an unlimited energy and a high transmission power so as to cover the entire sensor field. In RBF [5,6], the work was studied using OPNET® discrete event simulator. As the OPNET® licence is not freely available, we assess the performance of the proposed data

123

Dynamic Selection of Aggregation Points Table 1 Simulation parameters

Parameter

623 Value

Standard deviation of shadowing

5 dB

Path loss exponent

3

Path loss (L d0 ) at distance d0

55 dB

Packet reception threshold

−100 dBm

Sensor node transmit power

0 dBm

SIFS

20 µs

Slot duration

32 µs

Max. number of slots (W )

64

Slots in Group A (W A )

10

Alpha (α) for class-based mechanism

1

b

2/3

Initial (α) for alpha-based mechanism

0.5

Incremental step (c) for α

0.05

Maximum aggregation delay

0.01 s

RTS retransmit limit

7

DATA packet retransmit limit

9

Number of sensor nodes

60, 80, 100, 120, 140

Number of sensor nodes (for

80

aggregation delay analysis) Location of sink node

(50, 50)

Packet interarrival time

Uniform [0,1]

Simulation time

100 s

aggregation mechanisms using Castalia [2]. The scenario consists of nodes that are distributed randomly in a network topology with an area of 100 m × 100 m. The simulation was performed for 30 iterations using different random seeds. The impact of varying node density on the performance of the aggregation mechanisms is studied using an RSSI-based aggregation delay which has a maximum delay of 0.01 s. Each node in the network computes its aggregation delay whenever it receives a beacon packet from the sink. The delay is computed such that nodes closer to sink will have larger aggregation delay whereas nodes farther away from the sink will have smaller aggregation delay. The impact of varying aggregation delay on the performance of the aggregation mechanisms is studied using an 80-node random network topology. Table 1 gives some parameters used in the simulation. The packet sizes (transmission times, bytes) of Data and control packets are given in Table 2. Each node was given an initial energy of 1 J. The energy consumption is computed based on the power consumption rate for each sensor node state according to [1] and these values are given in Table 3. Nodes’ time spending in each state is computed and energy consumption is determined accordingly using the power consumption rates. The performances of the proposed class-based and alpha-based data aggregation mechanisms are also compared with RBF scenario without data aggregation [5,6]. The class enhanced CRT mechanism with no aggregation (class-enCRT-NoAgg) and alpha enhanced CRT mechanism with no aggregation (alpha-enCRT-NoAgg) are compared. The proposed class-enCRT-Agg and alpha-enCRTAgg are also compared with Data-Aware Anycast with Randomized Waiting (DAA+RW) [12].

123

624 Table 2 Packet size

Table 3 Power consumption rate for different modes of sensor node using CC2420 radio [1]

A. Awang, S. Agarwal Packet type

Packet size

RTS/CTS

0.000416 s (13 bytes)

DATA

0.001536 s (48 bytes)

ACK

0.00016 s (5 bytes)

Node state

Current (mA)

Power (mW)

Transmitting (0 dBm)

17.4

52.2

Receiving

19.7

59.1

Listening

0.365

1.095

Sleep

0.001

0.003

4.2 Results and Discussion This section presents our simulation results. As discussed earlier, number of aggregations directly effects the total number of transmissions and reception. However, number of packet transmissions can either be reduced due to early aggregation or loss of packets, whereas, number of aggregations may increase due to data aggregation taking place after several number of hops. Therefore, in this study, normalized number of packet transmissions is used along with the number of packets transmissions and receptions as a parameter to analyze the effectiveness of the aggregation process. We define the normalized number of transmissions as the ratio of the total number of packet transmissions to the total number of sources, as in [12]. Packet transmissions include RTS, CTS, DATA and ACK packets and number of sources are the total number of nodes that generate data at any particular instance of time. 4.2.1 The Impact of Varying Node Density Figure 10a gives the total number of RTS, CTS, DATA and ACK packets that are transmitted versus number of nodes. The number of packets transmitted depends on aggregation thus the curves for class- and alpha-based mechanisms without aggregation are almost the same. It is observed that the class-uniCRT-Agg, class-enCRT-Agg and alpha-enCRT-Agg have almost the same number of packet transmissions when the number of nodes in the network are less but with an increase in the number of nodes, alpha-enCRT-Agg performs better than the class-uniCRT-Agg, class-enCRT-Agg. This is because when there are less number of nodes (node density = 0.006 nodes/m2 ), data having the same Agg_ID are less. Therefore, the number of aggregations become less and all the three mechanisms behave similarly. But with an increase in the number of nodes, more data packets with the same Agg_ID are available for aggregation and we can observe that alpha-enCRT-Agg performs better than the classbased and alpha-based mechanisms. The number of packet receptions is directly related to the number of packet transmissions, hence the same kind of behavior is observed in Fig. 10b that gives the total number of packet receptions versus number of nodes. The DAA+RW exhibits about similar performance (slightly lower transmissions/receptions) with alpha-enCRT-Agg mechanism. Figure 11 depicts the normalized number of transmissions versus number of nodes. It can be observed that the curves for all the mechanisms (both with aggregation and without aggregation) increases gradually with an increase in the number of nodes due to the

123


3 2.5

5

7 class−enCRT−NoAgg alpha−enCRT−NoAgg class−uniCRT−Agg class−enCRT−Agg alpha−enCRT−Agg DAA+RW

Total number of receptions

Total number of transmissions

x 10 3.5

625

2 1.5 1 0.5 50

x 10


6 5 4 3 2 1

60

70

80

90

100

110

120

130

140

50

150

60

70

80

90

100

110

120

130

140

150

Number of nodes

Number of nodes

(a)

(b)

Fig. 10 The impact of varying node density on: a Total number of RTS, CTS, DATA, ACK packet transmissions, and b total number of RTS, CTS, DATA, ACK packet receptions, using a maximum aggregation delay of 0.01 s

Normalized number of transmissions


2800 2600 2400 2200 2000 1800 1600 1400 50

60

70

80

90

100

110

120

130

140

150

Number of nodes

Fig. 11 Normalized number of packet transmissions versus number of nodes using a maximum aggregation delay of 0.01 s

increase in the total number of transmissions (see Fig. 10a). In the case of no aggregation (class-enCRT-NoAgg and alpha-enCRT-NoAgg), both mechanisms have almost the same number of normalized transmissions since without aggregation both class-based and alphabased mechanisms have similar functioning. Comparing among the aggregation based mechanisms, i.e., class-uniCRT-Agg, class-enCRT-Agg and alpha-enCRT-Agg, it can be observed that alpha-enCRT-Agg performs better than the others. This is because in alpha-based mechanism, nodes find the best slot (lowest for that data packet transmission) in the contention window thus increasing the availability of data packets with similar Agg_ID. Whereas, in class-based mechanisms, nodes transmit by selecting the best possible slot within a group, however, this slot may not be the best slot when contention window is considered as a whole. The DAA+RW exhibits slightly lower number of normalised transmissions than alpha-enCRT-Agg. Figure 12a shows the packet loss probability versus number of nodes. Packet loss probability gives a reliability indicator of any protocol and a low packet loss probability implies a more reliable protocol. Packet loss in WSNs generally occurs due to packets collision

123

626

A. Awang, S. Agarwal −2

−3

1.5

class−enCRT−NoAgg alpha−enCRT−NoAgg class−uniCRT−Agg class−enCRT−Agg alpha−enCRT−Agg DAA+RW

Energy consumed per data packet (Avg. J/data pkt)

Packet loss probability

10

−3

10

−4

10

50

60

70

80

90

100

110

120

130

140

x 10

1.3 1.2 1.1 1 0.9 0.8 0.7 0.6 50

150

class−enCRT−NoAgg alpha−enCRT−NoAgg class−uniCRT−Agg class−enCRT−Agg alpha−enCRT−Agg DAA+RW

1.4

60

70

80

Number of nodes

90

100

110

120

130

140

150

140

150

Number of nodes

(a)

(b) 80

115

alpha−enCRT−Agg class−enCRT−Agg class−uniCRT−Agg DAA+RW alpha−enCRT−NoAgg class−enCRT−NoAgg

class−uniCRT−Agg class−enCRT−Agg alpha−enCRT−Agg DAA+RW class−enCRT−NoAgg alpha−enCRT−NoAgg

75

End−to−end delay (ms)

Network lifetime (s)

120

110 105 100

70 65 60 55 50 45

95

40 50

60

70

80

90

100

110

Number of nodes

(c)

120

130

140

150

50

60

70

80

90

100

110

120

130

Number of nodes

(d)

Fig. 12 The impact of varying node density on: a Packet loss probability, and b average energy consumed per data packet, and c network lifetime, and d end-to-end delay, using a maximum aggregation delay of 0.01 s

that mostly depends on the number of transmissions and receptions within the network. In Fig. 12a, packet loss probability for class-enCRT-NoAgg and alpha-enCRT-NoAgg mechanisms increases, whereas, in the case of class-uniCRT-Agg, class-enCRT-Agg and alphaenCRT-Agg, the curves increase gradually from 60-node network to 80-node network but tend to become almost constant later. This behavior can be explained by correlating this figure with the figure for normalized number of packet transmissions (see Fig. 11). Initially, normalized number of transmission increases thus we observe an increase in packet loss probability. However, when there are large number of nodes, the normalized number of packet transmissions tends to increase gradually thus we observe that packet loss probability remains almost constant. It is observed that alpha-enCRT-Agg exhibits the most superior performance (lowest packet loss) among the mechanisms (including DAA+RW) in terms of packet loss probability. The average energy consumed per data packet received by sink versus number of nodes in shown in Fig. 12b. This can be considered as the inverse of energy efficiency. Hence, a lower value refers to a more efficient communication. The increase in the number of nodes leads to an increase in the average communication cost (in terms of number of total number of transmission and receptions) and packet loss probability (see Fig. 12a), and hence an increase in the average energy consumed per data packet delivered to the sink, as depicted in Fig. 12b. For class-enCRT-NoAgg and alpha-enCRT-NoAgg, the increase is continuous and steady, whereas, for mechanisms employing aggregation the increase is gradual. This is because the increase in normalized number of transmissions (see Fig. 11) is gradual, whereas, in the case of no aggregation the increase is much higher. We observe that DAA+RW does not give better

123


627

performance than the rest of aggregation mechanisms (class-uniCRT-Agg, class-enCRT-Agg and alpha-enCRT-Agg) in terms of energy efficiency. Figure 12c shows the network lifetime versus number of nodes. Network lifetime is defined as the time until the first node death, i.e., the time until the first node exhausts its energy. As shown in Fig. 11, the normalized number of transmissions for alpha-enCRTAgg is less than class-enCRT-Agg and class-uniCRT-Agg. DAA+RW exhibits the lowest normalized number of transmissions (Fig. 11) but its packet loss probability is higher (Fig. 12a) and lesser energy efficient than the other aggregation mechanisms (Fig. 12b). Therefore, alpha-enCRT-Agg is expected to have a higher network lifetime than the others, as depicted in Fig. 12c. The mechanisms employing aggregation have higher network lifetime as compared to the mechanisms without aggregation since aggregation reduces communication cost. The class-enCRT-Agg improves the network lifetime by 16 % as compared to class-enCRT-NoAgg, whereas, alpha-enCRT-Agg improves by 16–17 % when compared to alpha-enCRT-NoAgg. The alpha-enCRT-Agg exhibits superior network lifetime among the mechanisms. The impact of varying node density on End-to-end (ETE) delay is analyzed. The ETE delay consists of delay due to queuing, propagation, and transmission and processing. It is the difference between times when a data packet is generated to the reception of the data packet by the sink. For each simulation iteration, the average ETE delays experienced by all data packets received at the sink is computed. Figure 12d shows the graph for average ETE delay versus number of nodes. In a data aggregation system, data packets at the aggregator nodes have to wait for packets from other nodes for aggregation. Therefore, the curves for mechanism with aggregation will have a higher ETE delay as compared to the one without aggregation. In can be observed in Fig. 12d that the curves for all the techniques increase as number of nodes in the network increases. The alpha-enCRT-Agg reduces the average ETE delay by 7–11 % when compared to class-uniCRT-Agg and 4–8 % when compared to classenCRT-Agg. The class-enCRT-Agg increases average ETE delay by 22–24 % as compared to class-enCRT-NoAgg whereas alpha-enCRT-Agg increases by 21–25 % when compared to alpha-enCRT-NoAgg. It is observed that ETE delay for DAA+RW is lesser than alphaenCRT-Agg by up to approximately of 8 % (in 140-node network). The percentage reduction in the normalized number of packet transmissions and average energy consumed per data packet are analyzed from Figs. 11 and 12b, respectively. Figure 13a gives the percentage reduction for normalized number of packet transmissions for different network sizes. The alpha-enCRT-Agg reduces the normalized number of transmissions by 12–20 % when compared to class-uniCRT-Agg and 6–10 % when compared to class-enCRT-Agg. The class-enCRT-Agg reduces the normalized number of transmission by 25–27 % as compared to class-enCRT-NoAgg whereas alpha-enCRT-Agg reduces by 26– 33 % when compared to alpha-enCRT-NoAgg. Figure 13b gives the percentage reduction in the average energy consumed per data packet for different network sizes. The alphaenCRT-Agg reduces the average energy consumed per data packet by 17–22 % when compared to class-uniCRT-Agg and 5–10 % when compared to class-enCRT-Agg. The classenCRT-Agg reduces the average energy consumed per data packet by 27–32 % as compared to class-enCRT-NoAgg whereas alpha-enCRT-Agg reduces by 29–33 % when compared to alpha-enCRT-NoAgg. 4.2.2 The Impact of Varying Aggregation Delay The impact of varying aggregation delays to the performance of data aggregation mechanisms is studied in a 80-node random network topology. In this study, when there is no aggregation,

123

50 class−enCRT−Agg vs. class−enCRT−NoAgg alpha−enCRT−Agg vs. alpha−enCRT−NoAgg alpha−enCRT−Agg vs. class−uniCRT−Agg alpha−enCRT−Agg vs. class−enCRT−Agg

45 40 35 30 25 20 15 10 5 0

60

80

100

120

Percentage reduction in energy consumed per data packet


Percentage reduction in normalized number of transmissions

628 50

class−enCRT−Agg vs. class−enCRT−NoAgg alpha−enCRT−Agg vs. alpha−enCRT−NoAgg alpha−enCRT−Agg vs. class−uniCRT−Agg alpha−enCRT−Agg vs. class−enCRT−Agg

45 40 35 30 25 20 15 10 5 0

140

60

80

100

120

140

Number of nodes

Number of nodes

(b)

(a)

Fig. 13 Percentage reduction in: a Normalized number of packet transmissions, and b average energy consumed per data packet versus number of nodes using a maximum aggregation delay of 0.01 s 5

5

x 10

2.6 2.4 2.2

x 10

4

class−enCRT−NoAgg alpha−enCRT−NoAgg DAA+RW class−uniCRT−Agg class−enCRT−Agg alpha−enCRT−Agg

Total number of receptions

Total number of transmissions

2.8

2 1.8 1.6 1.4

3.5


3 2.5 2

1.2 2

4

6

8

10

12

14 −3

Aggregation delay (s)

(a)

x 10

2

4

6 8 10 Aggregation delay (s)

12

14 −3

x 10

(b)

Fig. 14 The impact of varying aggregation delay on: a Total number of RTS, CTS, DATA, ACK packet transmissions, and b Total number of RTS, CTS, DATA, ACK packet receptions in an 80-node random network topology

i.e., for class-enCRT-NoAgg and alpha-enCRT-NoAgg, the aggregation delay is zero. The curves for these two mechanisms versus aggregation delay are constant values as none of the parameter is varying. We have shown these values in our analysis so as to make it easy for the reader to analyze the effect of the proposed mechanisms on the network. Figure 14a, b give the total number of RTS,CTS, DATA and ACK packets transmitted and received versus aggregation delay, respectively. Number of packet transmissions and receptions reduces with an increase in aggregation delay as more data packets are available for aggregation. Thus these graphs follow the same trend as normalized number of packet transmissions shown in Fig. 15a. A data packet is received by more than one node, thus number of packet receptions is higher than the number of packet transmissions. Therefore, we observe that the decrease in the number of packet receptions is slightly higher than the number of packet transmissions. Figure 15a gives the normalized number of packet transmissions versus aggregation delay. The normalized number of packet transmissions for class-uniCRT-Agg, class-enCRT-Agg and alpha-enCRT-Agg decreases with an increase in aggregation delay. This is because with an increase in aggregation delay, more data packets are available at a particular node and at the same time for aggregation, thus reducing the normalized number of packet transmissions accordingly. The normalized number of packet transmissions reduces rapidly until aggregation delay equals 0.008 s. After this point, the decrease in the normalized number of packet

123


629


3000

2500

2000

1500 2

4

6

8

10


(a)

12

14 −3

x 10

Energy consumed per data packet (Avg. J/data pkt)

Normalized number of transmissions

−3

x 10

3500


1.3 1.2 1.1 1 0.9 0.8 0.7 0.6 2

4

6

8

10


12

14 −3

x 10

(b)

Fig. 15 The impact of varying aggregation delay on: a Normalized number of packet transmissions, and b average energy consumed per data packet in an 80-node random network topology

transmissions tends to become gradual. Data transfer between two nodes takes about 0.007 s and when the aggregation delay increases above this value, most of the nodes have already gone through at least one round of data transfer. Therefore, number of data packets being transmitted to a node from its one-hop neighbors, decreases after 0.007 s. Thus we observe a comparatively gradual decrease in the normalized number of packet transmissions after 0.008 s. Energy consumed in the network is directly related to the total number of transmissions and receptions. Since the number of packets transmitted and received decreases, therefore, the average energy consumed should also decrease. This is depicted in Fig. 15b for the average energy consumed per data packet delivered to sink node. The curves in this figure follow a similar trend as that in Figs. 14a, b and 15a. This is because with an increase in an aggregation delay, more data packets are aggregated by each aggregator node. Figure 16a shows the network lifetime versus aggregation delay. We can observe from Fig. 15b that with a decrease in average energy consumption, there is an increase in network lifetime in Fig. 16a. As we keep on increasing the aggregation delay, we can observe that the slope for the curves of class-uniCRT-Agg, class-enCRT-Agg and alpha-enCRT-Agg decreases. This emphasizes that there is an upper bound on the network lifetime in WSNs, as studied in [31]. The class-enCRT-Agg improves the network lifetime by 12–15 % as compared to classenCRT-NoAgg whereas alpha-enCRT-Agg improves by 11–17 % when compared to alphaenCRT-NoAgg. The alpha-enCRT-Agg also exhibits superior performance than DAA+RW in terms of network lifetime. Figure 16b gives the average ETE delay versus aggregation delay. As expected, the curves for class-uniCRT-Agg, class-enCRT-Agg and alpha-enCRT-Agg increase with an increase in the aggregation delay. The class-enCRT-Agg increases the average ETE delay by 12– 36 % as compared to class-enCRT-NoAgg whereas alpha-enCRT-Agg increases by 15–36 % when compared to alpha-enCRT-NoAgg. Figures 15b and 16b reveal that there is a trade-off between energy saving and ETE delay in the aggregation process. Efficient aggregation in WSNs will result in high energy saving at the expense of increased ETE delay. However, depending on applications, the trade off between energy consumed and ETE delay of the system can be varied to meet certain objectives. We compute the percentage reduction in the normalized number of packet transmissions and average energy consumed per data packet from Fig. 15a, b, respectively. Figure 17a gives the percentage reduction in the normalized number of packet transmissions versus aggregation delay. The alpha-enCRT-Agg reduces the normalized number of packet trans-

123

630

A. Awang, S. Agarwal 75 alpha−enCRT−Agg class−enCRT−Agg class−uniCRT−Agg DAA+RW alpha−enCRT−NoAgg class−enCRT−NoAgg

115

110

105

100 95 2

4

class−uniCRT−Agg class−enCRT−Agg alpha−enCRT−Agg DAA+RW class−enCRT−NoAgg alpha−enCRT−NoAgg

70

End−to−end delay (ms)

Network lifetime (s)

120

6

8

10

12


14

65 60 55 50 45 40

2

4

−3

6

8

10

12

(a)

14 −3


x 10

x 10

(b)

55 class−enCRT−Agg vs. class−enCRT−NoAgg alpha−enCRT−Agg vs. alpha−enCRT−NoAgg alpha−enCRT−Agg vs. class−uniCRT−Agg alpha−enCRT−Agg vs. class−enCRT−Agg

50 45 40 35 30 25 20 15 10 5 0

0

0.0025

0.0050

0.0075

0.0100


(a)

0.0125

Percentage reduction in energy consumed per data packet

Percentage reduction in normalized number of transmissions

Fig. 16 The impact of varying aggregation delay on: a Network lifetime, and b End-to-end delay in an 80-node random network topology 50 class−enCRT−Agg vs. class−enCRT−NoAgg alpha−enCRT−Agg vs. alpha−enCRT−NoAgg alpha−enCRT−Agg vs. class−uniCRT−Agg alpha−enCRT−Agg vs. class−enCRT−Agg

45 40 35 30 25 20 15 10 5 0

0

0.0025

0.0050

0.0075

0.0100

0.0125


(b)

Fig. 17 Percentage reduction in: a Normalized number of packet transmissions, and b average energy consumed per data packet versus aggregation delay in an 80-node random network topology

missions from 20 to 40 % when compared to alpha-enCRT-NoAgg whereas class-enCRT-Agg reduces the normalized number of packet transmission by 18–35 % when compared to classenCRT-NoAgg. The alpha-enCRT-Agg reduces normalized number of packet transmissions by 12–26 % as compared to class-uniCRT-Agg and 7–13 % as compared to class-enCRTAgg. As for the average energy consumed per data packet, the percentage reduction versus aggregation delay is shown in Fig. 17b. The alpha-enCRT-Agg reduces the average energy consumed per data packet by 14–24 % as compared to class-uniCRT-Agg and 5–9 % as compared to class-enCRT-Agg. The class-enCRT-Agg reduces the average energy consumed per data packet by 22–34 % as compared to class-enCRT-NoAgg whereas alpha-enCRT-Agg reduces by 19–35 % as compared to alpha-enCRT-NoAgg. 5 Conclusion Data aggregation offers a key solution to curtail network loads, reduce redundant packets transmission and hence reduces transmission energy consumption in WSNs. In this paper, two topology independent data aggregation mechanisms are presented. These mechanisms use RBF for aggregating and forwarding data toward a sink node. An aggregator node is prioritized using a CTS response based on the availability of data for aggregation at contending nodes. Specifically, the performance of alpha-enCRT-Agg, class-enCRT-Agg, class-uniCRTAgg, alpha-enCRT-NoAgg, class-enCRT-NoAgg and DAA+RW mechanisms are compared

123


631

in terms of the total number of packet transmissions and receptions, normalized number of packet transmissions, energy consumed per data packet, network lifetime, packet loss probability and ETE delay for different network sizes and aggregation delays. It was found that alpha-based CTS response mechanism gives superior performance than the other class-based CTS response mechanisms. The results also indicate there is a trade-off between energy saving and ETE delay in the aggregation process. Efficient aggregation in WSNs will result in high energy saving at the expense of increased ETE delay due to temporal convergence requirement. However, depending on applications, the trade off between energy consumed and ETE delay of the system can be compromised to meet certain performance objectives. As part of future work, we would like to investigate the impact of varying network load on the performance of the aggregation mechanisms. We may also consider the node’s residual energy as part of key parameters in the contention to select an aggregator node among the forwarding candidates. Acknowledgments This work has been supported by Universiti Teknologi PETRONAS, Malaysia under the Short Term Internal Research (STIRF) grant (STIRF Code No: 64/2011).

References 1. (2013). CC2420 Data Sheet. http://focus.ti.com/docs/prod/folders/print/cc2420.html. Accessed April 12, 2013. 2. (2013). Wireless sensor networks simulator. http://castalia.research.nicta.com.au/index.php/en/. Accessed April 12, 2013. 3. Akyildiz, I., Su, W., Sankarasubramaniam, Y., & Cayirci, E. (2002a). A survey on sensor networks. IEEE Communications Magazine, 40(8), 102–114. 4. Akyildiz, I., Su, W., Sankarasubramaniam, Y., & Cayirci, E. (2002b). Wireless sensor networks: A survey. Computer Networks (Elsevier), 38(4), 393–422. 5. Awang, A. (2011). A cross-layer MAC/routing protocol for wireless sensor networks. PhD thesis, Institut TÉLÉCOM / TÉLÉCOM Bretagne, France. 6. Awang, A., Lagrange, X., & Ros, D. (2009). RSSI-based forwarding for multihop wireless sensor networks. In 15th Open European summer school and IFIP TC6.6 workshop (EUNICE2009). The internet of the future, LNCS (Vol. 5733, pp. 138–147). 7. Braginsky, D., & Estrin, D. (2002). Rumor routing algorithm for sensor networks. In 1st ACM international workshop on wireless sensor networks and applications (WSNA) (pp. 22–31). doi:10.1145/570738. 570742. 8. Bulusu, N., Heidemann, J., Estrin, D., & Tran, T. (2004). Self-configuring localization systems: Design and experimental evaluation. ACM Transactions on Embedded Computing Systems, 3(1), 24–60. 9. Cohen, E., & Kaplan, H. (2004). Spatially-decaying aggregation over a network: Model and algorithms. In ACM SIGMOD international conference on management of data (pp. 707–718). 10. Culler, D., Estrin, D., & Strivastava, M. (2004). Overview of sensor networks. In IEEE Computer Society’04 (pp. 41–49). 11. Ding, M., Cheng, X., & Xue, G. (2003) Aggregation tree construction in sensor networks. In IEEE 58th vehicular technology conference (pp. 2168–2172). 12. Fan, K. W., Liu, S., & Sinha, P. (2007). Structure-free data aggregation in sensor networks. IEEE Transactions on Mobile Computing, 6(8), 929–942. 13. Fan, K. W., Liu, S., & Sinha, P. (2008). Dynamic forwarding over tree-on-DAG for scalable data aggregation in sensor networks. IEEE Transactions on Mobile Computing, 7(10), 1271–1284. 14. Fasolo, E., Rossi, M., Widmer, J., & Zorzi, M. (2007). In-network aggregation techniques for wireless sensor networks: A survey. IEEE Wireless Communications Magazine, 14(2), 70–87. 15. Guo, W., Xiong, N., Vasilakos, A., Chen, G., & Cheng, H. (2011). Multi-source temporal data aggregation in wireless sensor networks. Wireless Personal Communications, 56(3), 359–370. 16. Heinzelman, W., Chandrakasan, A., & Balakrishnan, H. (2002). An application-specific protocol architecture for wireless microsensor networks. IEEE Transactions on Wireless Communications, 1(4), 660–670. 17. Intanagonwiwat, C., Govindan, R., Estrin, D., Heidemann, J., & Silva, F. (2003). Directed diffusion for wireless sensor networking. IEEE/ACM Transactions on Networking, 11(1), 2–16.

123

632


18. Karl, H., & Willig, A. (2005). Protocols and architectures for wireless sensor networks. New York: Wiley. 19. Kulik, J., Heinzelman, W., & Balakrishnan, H. (2002). Negotiation-based protocols for disseminating information in wireless sensor networks. Wireless Networks, 8(2/3), 169–185. 20. Lindsey, S., Raghavendra, C., & Sivalingam, K. (2002). Data gathering algorithms in sensor networks using energy metrics. IEEE Transactions on Parallel and Distributed Systems, 13(9), 924–935. 21. Liu, Y., Yang, Z., Wang, X., & Jian, L. (2010). Location, localization, and localizability. Journal of Computer Science and Technology, 25(2), 274–297. 22. Luo, H., Liu, Y., & Das, S. (2007). Routing correlated data in wireless sensor networks: A survey. IEEE Networks, 21(6), 40–47. 23. Madden, S., Franklin, M. J., Hellerstein, J. M., & Hong, W. (2002). TAG: A Tiny AGgregation service for ad-hoc sensor networks. ACM SIGOPS Operating Systems Review, 36(SI), 131–146. 24. Nath, S., Gibbons, P., Seshan, S., & Anderson, Z. (2004). Synopsis diffusion for robust aggregation in sensor networks. In 2nd international conference on embedded networked sensor systems (pp. 250–262). 25. Qin, M., & Zimmermann, R. (2007). VCA: An energy-efficient voting-based clustering algorithm for sensor networks. Journal of Universal Computer Science, 13(1), 87–109. 26. Rajagopalan, R., & Varshney, P. (2006). Data-aggregation techniques in sensor networks: A survey. IEEE Communications Surveys and Tutorials, 8(4), 48–63. 27. Sharaf, A., Beaver, J., Labrinidis, A., & Chrysanthis, K. (2004). Balancing energy efficiency and quality of aggregated data in sensor networks. The VLDB Journal, 13(4), 384–403. 28. Tan, H. O., & Körpeoˇglu, I. (2003). Power efficient data gathering and aggregation in wireless sensor networks. ACM Newsletter SIGMOD Record, 32(4), 66–71. 29. Tanenbaum, A. (2002). Computer networks (4th ed.). Upper Saddle River: Prentice Hall Professional Technical Reference. 30. Villas, L., Guidoni, D., Boukerche, A., Araujo, R., & Loureiro, A. A. F. (2011). Dynamic and scalable routing to perform efficient data aggregation in WSNs. In IEEE international conference on communications (ICC) (pp. 1–5). 31. Wang, Q., & Zhang, T. (2009). Bottleneck zone analysis in energy-constrained wireless sensor networks. IEEE Communications Letters, 13(6), 423–425. 32. Wei, G., Ling, Y., Guo, B., Xiao, B., & Vasilakos, A. V. (2011). Prediction-based data aggregation in wireless sensor networks: Combining grey model and Kalman filter. Computer Communications, 34(6), 793–802. 33. Xiangning, F., & Yulin, S. (2007). Improvement on LEACH protocol of wireless sensor network. In International conference on sensor technologies and applications (pp. 260–264). 34. Yassein, M., Al-Zou’bi, A., Khamayseh, Y., & Mardini, W. (2009). Improvement on LEACH protocol of wireless sensor networks (VLEACH). International Journal of Digital Content Technology and its Applications, 3(2), 132–136. 35. Younis, O., & Fahmy, S. (2004). HEED: A hybrid, energy-efficient, distributed clustering approach for ad hoc sensor networks. IEEE Transactions on Mobile Computing, 3(4), 366–379. 36. Zorzi, M., & Rao, R. (2003). Geographic random forwarding (GeRaF) for ad hoc and sensor networks: Energy and latency performance. IEEE Transactions on Mobile Computing, 2(4), 349–365.

Azlan Awang received his Ph.D. degree from Institut TÉLÉCOM/ TÉLÉCOM Bretagne and University Rennes 1, Rennes, France in 2011. He received his B.Sc. and M.Sc. degrees in Electrical Engineering from Polytechnic Institute of NYU, Brooklyn, New York, USA in 1989 and 1990, respectively. He was with Motorola (M) Sdn Bhd (1991), Schlumberger Overseas S.A. (1992–1993), Alcatel Networks Systems (M) Sdn Bhd (1994–2001) and Universiti Teknologi MARA (2002–2004). He joined Universiti Teknologi PETRONAS, Malaysia in 2004 and currently is a Senior Lecturer in the Electrical and Electronic Engineering department. His research interests are in the area of cross-layer MAC and Routing for Wireless Sensor Networks and Wireless Body Area Networks. He is a member of IEEE, Eta Kappa Nu and Tau Beta Pi engineering honor societies.

123


633

Shobhit Agarwal received his Bachelors in Technology degree from Amity University Noida, India in 2009. Currently, he is pursuing his MSc degree in the Electrical and Electronic Engineering department, Universiti Teknologi PETRONAS, Malaysia. His research interests include cross-layer designs, data aggregation, MAC and routing protocols in Wireless Sensor Networks.

123