35th Annual IEEE Conference on Local Computer Networks

LCN 2010, Denver, Colorado

Mitigating the Effect of Interference in Wireless Sensor Networks

Nadeem Ahmed, Salil S. Kanhere and Sanjay Jha
School of Computer Science and Engineering University of New South Wales Sydney, Australia Email: {nahmed, salilk, sanjay}@cse.unsw.edu.au
AbstractPerformance of a deployed Wireless Sensor Network (WSN) is greatly inuenced by the interference it is subject to during operation. Degradation happens due to interference resulting in packet drops, retransmissions, link instability and inconsistent protocol behavior. We have conducted experiments that highlight the fact that interference caused by WiFi and co-channel contention signicantly degrades the network performance of protocols. These potential sources of interference must therefore be accounted for during the design stage of a WSN in order to achieve acceptable network performance. Based on these observations, we have proposed a multi-hop multi-channel topology control protocol RMMTC for WSN that takes into account interference caused by WiFi networks in operation in the vicinity and uses multiple channels at different frequencies to mitigate the effect of co-channel interference. This paper details the design and performance evaluation of our proposed RMMTC protocol using both simulations and empirical experiments. In addition, we have formulated the multiple channel assignment problem as an Integer linear program (ILP) and compared the performance of our distributed protocol with the centralized ILP solution. The simulation results show that RMMTC performs close to the optimal centralized ILP and achieves a nine-fold reduction in the percentage of dropped packets when a dense network is subjected to interference from WiFi and co-channel contention.

I. I NTRODUCTION Wireless Sensor Networks (WSN) have been actively researched over the past decade and many protocols have been proposed covering the MAC, routing and transport layers. Recently, the focus of research in WSN is shifting from performance evaluations using simulations studies to experimental evaluations of proposed protocols in real-world scenarios. It has been observed that most of the proposed protocols do not perform as per design when subjected to real radio environments. One of the major causes of under-performance is the interference issues in WSN. Degradation due to interference results in packet drops, retransmissions, link instability and inconsistent protocol behavior [1] [2] [3] [4] etc. Interference for a ZigBee based WSN1 can be broadly classied as internal, i.e., originating from within the network e.g., multiple links within the same network that can hear each other and external, e.g., WiFi interference. Effect of internal interference is mitigated by careful topology selection, choice of appropriate MAC layer or use of multi-channels. Interference due to WiFi sources, on the other hand, is difcult
1 In this paper, we use ZigBee based WSN and WiFi interference as illustrative examples. In general, the ideas discussed can be applied to any multi-channel WSN operating in presence of external interference.

to anticipate due to several reasons. Firstly, WiFi operate on the same frequency band (2.4 GHz) as ZigBee based WSN and hence the WiFi transmissions appear as noise to ZigBee. Secondly, the WiFi transmissions are likely to be from far more powerful sources as compared with ZigBee based sensor nodes thus considerably lowering the Signal to Noise Ratio (SNR) for the ZigBee transmissions. Considering that WiFi deployments are widespread and most realistic scenarios where WSN would be deployed would very likely be co-located with APs e.g., in urban buildings or factory environment etc., these potential sources of interference must be accounted for during the design stage of a WSN in order to achieve acceptable network performance. One potential solution to mitigate the effect of internal interference is to form a network topology based on multiple noninterfering cliques operating on different channels [5]. Typical WSN devices are capable of channel switching capabilities that provide support for the use of multiple channels operating at different frequencies e.g., CC2420 radios used for MicaZ motes can use 16 different channels in the 2.4 GHz band [6]. Multiple transmissions can take place on these orthogonal channels to increase the spectral efciency. Use of multiple channels in WSN has been explored in previous research efforts and multi-channel MAC protocols has been proposed to improve the network throughput [7] [8] [9] [10]. Most of these protocols assume the presence of multiple orthogonal channels for parallel communications. One problem with use of multichannel in WSN is that channels are not truly orthogonal i.e., multiple simultaneous transmissions on adjacent channels do cause interference due to unwanted power received from transmitters on adjacent channels [2], [5] etc. In [11], we presented a study on the interference issues for multi-channel WSN (partly reproduced in Section III). We studied the effect of interference on off-the-shelf MicaZ motes operating on different ZigBee channels. The study showed that ZigBee transmissions are affected by both external (WiFi) and internal (adjacent channel) interferences, depending on the network topology and the presence of WiFi sources in the vicinity. The study also highlighted that effect of adjacent channel interference is more pronounced when ZigBee transmitters are using variable transmission powers on adjacent channels. Based on the above mentioned study, in order to improve the network performance in presence of both internal and external interference, we have proposed a distributed RSSI based Multihop Multi-channel Topology Control protocol (RMMTC). This

978-1-4244-8389-1/10/$26.00 2010 IEEE

160

protocol selects short, multi-hop communication neighbors with adequate RSSI values to mitigate the effect of external interference, and also reduces co-channel contention and adjacent channel interference by careful channel assignment to form multiple collision domains. In this paper, we formulated the problem of RSSI based multi-channel topology control as an Integer Linear Program (ILP) that serves as a benchmark to compare our proposed heuristic to the optimal solution. We perform a detailed and systematic simulation and empirical study to compare the network performance of our the RMMTC protocol with other multi-hop protocols such as simple hop count based [12] and MintRoute [13]. Simulation results show that for dense topologies, RMMTC protocol considerably improves the network performance by achieving a nine-fold reduction in percentage of drop packets due to interference (from about 45% to less than 5%). This paper is organized as follows; Section II introduces related work in the area of interference issues and multichannel communications in WSN. Section III describes our preliminary experiments that establish the characteristics of WiFi and co-channel interference in ZigBee based WSN. Network model for the topology control problem and the protocol design for the RMMTC protocol is presented in Section IV. Section V details the simulation and experimental study to evaluate the performance of our proposed RMMTC protocol. Finally conclusion is presented in Section VI. II. R ELATED W ORK There is a large body of research exploring the use of multichannels in WSN with the majority focussing on proposing the use of multi-channel MACs to coordinate simultaneous use of multi-channels [7] [8] etc. Wu et.al. in [5] proposed assigning different channels to subtrees to avoid channel switching and use of time-synchronized multi-channel MAC protocols. They proposed sampling the environment to discover ZigBee channels for use that are least effected by WiFi and then create node-disjoint multi-path trees for reducing the internal interference. Our work is different in several ways. Firstly, instead of nding ZigBee channels that are least effected by WiFi, we assume that WiFi interference effects almost all of the ZigBee channels. This assumption is more realistic as WiFi sources operating on different channels can be turned on/off effecting different ZigBee channels at different time instances. Secondly, following the observation in [14] that higher RSSI values for ZigBee based WSN motes have a strong correlation with good packet reception rates, we use RSSI as the threshold for selecting links in the topology that can withstand the WiFi interference. A white paper from CrossBow (manufacturer of MicaZ motes) [6] reports the adverse effect of WiFi on ZigBee transmissions. However, the reported values are much lower than that observed by our link characterization experiments (see Section III). Authors in [3] also conrmed through experiments the adverse effect of WiFi interference on ZigBee based networks and proposed to switch ZigBee channels once interference is detected on the current ZigBee channel using threshold based interference estimator. The base station in their

approach probes the deployed network for RSSI sampling. The selected nodes along the path to the data source reports data back to the base station that then nds the channel with least interference. The base station then directs the node lying on the path to the source to switch the channel. This approach incurs high overhead for a many-to-one data transfer where all the nodes are sending sensed data back to the bases station, which is typical of most WSN. Hauer et.al. [15] recently presented results from an empirical study for the effect of WiFi interference for body area networks. They took multi-channel measurements and proposed noise oor as the estimator for link quality degradation. In [16], the authors proposed channel surng schemes by adapting channel allocations to avoid interference. The proposal is reactive in nature as all nodes initially work on the same channel and switching (of a group or the whole network) only takes place when interference is encountered. In contrast, our work pre-empts the inuence of interference by allocating multiple communications channels to different subtrees rooted at the base station. III. L INK C HARACTERIZATION E XPERIMENTS We have conducted empirical experiments to characterize the link behavior when subjected to WiFi and adjacent channel interference. The initial results appeared in a poster paper [11] and some of the results are re-produced here to put this work in proper context.

Fig. 1.

RF Spectrum Utilization Showing WiFi (high peaks) and 802.15.4

100

Average Packet Reception Rate

90

80

70

60 0 dBm 25 dBm

50

40 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26

ZigBee Channel No

Fig. 2.

Effect of WiFi Interference

We rst study the effect of WiFi interference on the transmission from WSN nodes operating on different ZigBee channels. We conducted our experimental study using the MicaZ platform that provides 16 ZigBee channels operating between 2.405 to 2.480 GHz frequency range (see Figure 1). WiFi, one the other hand, has 11 channels which operate in the same frequency band. As each channel in WiFi occupies

161

a wider RF spectrum than ZigBee (22 MHz as compared with 3 MHz), a single WiFi channel can simultaneously cause interference on four adjacent ZigBee channels e.g., WiFi channel 1 occupies the same RF spectrum as channels 11-14 for ZigBee. The experiments were performed indoors in an ofce building. A pair of transmitter and receiver nodes were placed 1.5m apart on a table raised about 0.5m above the oor level. The transmitter sends a total of 5000 packets at the rate of 20 packets/second with each packet containing a unique sequence number. The receiver logs the sequence number and RSSI of each packet that it receives. For interference purposes, we used FTP clients operating at WiFi channels 1 and 11 with -55 dBm and -62 dBm average received signal power (measured with a spectrum analyzer, co-located with the receiver). We used two power levels (maximum 0 dBm and minimum -25 dBm) for the ZigBee transmitter and repeated the experiment for 16 different ZigBee channels. The average RSSI recorded at the receiver was -62 dBm and -88 dBm for the maximum and minimum transmission level, respectively. We note that techniques for creating reproducible interference proposed in [17] can also be utilized for these experiments. Figure 2 shows that the transmitter is able to maintain a successful packet reception rate (PRR) above 90%, when using any of the ZigBee channels, at maximum transmission power. For the case when the node is transmitting at lowest power level, effect of WiFi interference is more pronounced with average PRR values falling to about 44% (for channel 14).
100

100

Average Packet Reception Rate

90

80

70 0 dBm 25 dBm Mixed Mode ( 0 and 25 dBm)

60

50

19

20

21

22

23

24

25

26

ZigBee Channel

Fig. 4.

WiFi and Multi-Channel Interference (Channel 19-26)

90

80

70

60

0 dBm 25 dBm Mixed Mode (0 and 25 dBm)

50

11

12

13

14

15

16

17

18

ZigBee Channel

Fig. 3.

WiFi and Multi-Channel Interference (Channel 11-18)

We next conducted experiments with eight pairs of nodes, each pair tuned to a separate frequency. Transmitters are placed in one line at a distance of 1.5m from their respective receivers. Simultaneous transmissions occur at multiple frequencies (channel 11-18 and channel 19-26 in separate experiments). We conducted three different set of experiments. In the rst set, all nodes transmit at 0 dBm power. In the second set, the node transmission power was reduced to -25 dBm. The last set, referred to as the mixed mode experiment, nodes using odd number channels transmit at full power while nodes operating at even number channels transmit at lowest power. The WiFi channels 1 and 11 were in operation, both with average received signal power of about -68 dBm. Figures 3 and 4 show that nodes are able to maintain an average PRR close to 100% when transmitting at 0 dBm (with RSSI of about -60 dBm). When the transmission power is

reduced to -25 dBm (with RSSI of about -88 dBm), PRR is dropped to about 90% for channels that are under the inuence of WiFi (channels 11-14 and 22-23 are affected). For the mixed mode experiment, nodes transmitting at 0 dBm are able to maintain the PRR close to 100%. For nodes transmitting at lowest power (even number channels) that are also under the inuence of WiFi interference, the PRR is dropped to 70% 75% (channels 14 and 22). In conclusion, our experimental study reveals the following characteristics of multi-channel interference: 1) ZigBee transmissions are affected by both WiFi and adjacent channel interference, depending on the received signal power levels. 2) ZigBee sources are able to maintain good PRR (> 85%) when the received power from transmitter signal and received power from interference signal differ by less than 25 dB, for WiFi or adjacent channel transmissions. 3) Mixed mode experiments revealed that variable transmission power WSN designed for reducing interference are infact more prone to interference when a high power neighbor is operating on an adjacent channel in addition to interference from WiFi. These characteristics serves as guidelines for designing an efcient multi-channel protocol for WSN e.g., ZigBee channel selection should depend on the topology and whether there are WiFi sources in operation in the environment. Effect of WiFi interference can be mitigated by selecting short, multi-hop communication as compared to long distance transmissions. Similarly, adjacent channels can be utilized for a multi-channel WSN by careful channel assignment. We follow these guidelines in the design of a distributed topology control protocol discussed in the next section. IV. N ETWORK M ODEL AND P ROTOCOL D ESIGN Based on the recommendations from the link characterization experiments discussed in the previous section, we detail here the design of a topology control protocol that can alleviate the effect of external and internal interferences. We rst describe a graph theoretic network model that is used to formulate the multiple channel topology control optimization problem and then detail the design of a heuristic and distributed topology control protocol. A. Network Model We represent the network as an undirected graph G(V, E ) where V is the set of vertices (each v V corresponds

Average Packet Reception Rate

162

to a node in network). E is the set of edges representing the possible wireless links between sensor nodes. Let Rt and Ri denote the xed transmission range and interference range where Ri > Rt . Let d(u, v ) represent the Euclidean distance between two nodes u, v V and K represent the number of available channels. For two nodes u, v V , direct communication is only possible if the d(u, v ) Rt and the two nodes operates on a common channel k K . Let Rthresh represent the minimum threshold RSSI value required to overcome the effect of interference. We rst prune the graph G to remove wireless links with RSSI < Rthresh . Let GR (V, E ) represent the graph induced by RSSI based links selection where GR G. Given the network graph GR (V, E ), we calculate the interference value for a node as follows: Assume node vi transmits to node vj (vi and vj V and both operate on same channel) using transmission power Pi . The signal from vi will reach the receiver vj with signal strength Sij
k Sij

Total number of nodes effecting a receiver j given by

k Cj =
k iUj

k Yij

(6)

During the course of the topology selection, each vertex vi V chooses a vertex vj V, i = j as its parent creating link Lij operating on a channel k for routing to the sink. Lk ij = 1 0 if i, j are assigned channel k otherwise (7)

The output of the topology control protocol is a spanning subgraph GT (V, E ) where GT G. The objective function for multiple channel assignment can be dened as: Find a network channel assignment graph GT (V, E ) G, such that the interference on vertices V is minimized. The problem can be formulated as an Integer Linear Program (ILP) as follows; k Let dest represent the sink, and vj represent a node j V operating at channel k . M inimize
kK v V \{dest} k k Cj vj

= Pi /d(i, j )

(1)

where is the pathloss constant. Let denote the background noise value including the interference from WiFi transmissions. The SNR at the receiver node vj is given by
k / > Rthresh SN R = Sij

(2)

The reception of signal from node vi to vj is also inuenced by co-channel interference. Since interference is an issue at k the receiver, let Uj represent the set of nodes located within Ri distance from the receiver vj and operating on the same k channel k as vj . Co-channel interference Ij at a receiver vj is thus given by
k = Ij
k uUj

subject to the constraints given by Equations 5, 7 and

k vj =

1 if j is assigned channel k 0 otherwise

(8)

Centralized optimization can be performed on a given network Graph G to select GR and subsequentally GT to arrive at an idealized network topology with channel assignment. B. Protocol Design The solution of the ILP formulated problem is not feasible due to its complexity and centralized nature. Hence, we introduce RMMTC, a distributed and heuristic multi-channel protocol, that can be used to nd an approximate solution in polynomial time. We have designed the RMMTC protocol based on the recommendations from our link characterization experiments discussed in Section III. There are two important design considerations. First is the choice of the threshold used for categorizing the external interference (section III characteristic No 1 & 2). As RSSI at the receiver governs the interference, we chose RSSI threshold values for selecting links that are resilient to the effect of external interference. Second consideration is how to avoid the internal interference (co-channel as well as adjacent channel, section III characteristic No 3). We do careful channel assignment for utilizing multiple channels for this purpose. There are two popular choices for the design of a multichannel protocol. One can assign different channels to different links. In this case, each node can use multiple channels. The nodes in the local neighborhood coordinate and decide on the channel allocations using a multi-channel MAC. This approach introduces the channel switching and channel coordination

k Suj

(3)

The signal to interference and noise ratio (SINR) for the receiver node vj combines the background noise interference k and co-channel interference Ij and is represented by
k k /(Ij + ) > Rthresh SIN R = Sij

(4)

Note that background interference including the interference from WiFi is assumed constant for this work for all the nodes in the topology. A node i causes interference at the receiver node j if it operates on the same channel k and is within Ri of j .
k = Yij k 1 if i Uj 0 otherwise

(5)

Let C be the cost function that associates a non-negative cost to each vertex in the network. As we want to minimize the interference in our case, the cost is assumed to be the number of nodes operating on the same channel and causing interference at the receivers along the path from a node to the sink. The total cost can be dened as the sum of the costs of vertices.

163

overheads. The other choice is to form different routing subtrees rooted at the base station and assign a unique channel for transmission to each subtree. All the nodes belonging to a subtree work on the specied channel. Hence any single channel MAC can be used without the need for expensive channel coordination. Similar to the work presented in [5], we used the later approach for the design of RMMTC protocol. RMMTC protocol thus establishes links in the tree topology with two objectives. First, it selects links in the topology that have RSSI values above a certain RSSI threshold. As the most popular data transfer pattern in WSN is the many-toone transfer where all the nodes send data packets to the base station, RMMTC ensures that links in both directions have RSSI values above a certain threshold value. Secondly, RMMTC forms routing subtrees rooted at the base station where each subtree can be assigned a noninterfering communication channel. During the course of the subtree formation, new nodes are added to a tree based on the interference levels observed at their parent nodes. This is done to reduce the internal interference. As described in Section III, ZigBee based WSN has 16 communication channels available for use out of which channel 15, 20, 25 and 26 are least effected by external interference. We have selected channel 26 as the control channel (required for the neighborhood discovery and channel assignment process) for the RMMTC protocol while channel 15, 20, and 25 are the rst channels to be assigned for use by the base station. Once these channels are assigned, the base station uses channels 11, 23, 17, and 13 in this specic order. Remaining channels are only assigned if required. This assignment order makes sure that channels least effected by external interference are utilized rst, followed by alternate channels that would not cause interference. Adjacent channels, prone to internal interference, are only added in the topology if sufcient subtrees are formed that require further unique channel assignment (section III characteristic No 3). We note here that proactive techniques to detect channels that do not experience external interference can also be employed and a list of good quality channels provided to the base station for use [5]. We have not used this technique for two reasons. Firstly, these network measurements are expensive as these need to be taken at several locations in the topology. Secondly, WiFi sources can be introduced or switched off at any time in the environment and therefore these network measurements need to be taken periodically and topology recongured accordingly. We have thus designed the RMMTC protocol without any presumed knowledge of the interference environment. Any such knowledge can be utilized for further enhancing the protocol performance. We now describe the details of the protocol. Once nodes are randomly deployed in the target area, the base station announces its position in the topology by broadcasting HELLO messages. Nodes that receive this HELLO message extract the RSSI values and mark base station as 1 Hop neighbor in the neighbor table. If the RSSI value is more than the Rthresh , the node reply back with an ACK packet. The base station checks the RSSI values of the received ACK messages and in case RSSI is more than the Rthresh , replies each such neighbor with a SELECT message containing a unique channel number.

Fig. 5.

RMMTC Protocol Message Exchanges

Node on reception of the SELECT message notes the channel number and marks the upstream link as valid. These nodes now send a CONFIRM message to the base station. For the rest of the discussion, we refer to the nodes already assigned a channel No as level N nodes and the downstream one hop child nodes as level N+1 nodes. For interference calculations, nodes need to know the location of other nodes operating on the same channel (part of the same subtree). Once a node become a level N node, it broadcasts a REQ message requesting neighbor information from nodes on the same subtree. Nodes that are part of the same subtree, reply back with information about their one-hop neighbors on the k same channel. The level N nodes now discover set Uj (refer Section IV-A) by forming a disc of radius Ri centered at itself. The existing interference Ie is thus equal to the cardinality of k set Uj . Once the Ie is calculated, Level N nodes announce their position and channel number using the HELLO messages to the downstream nodes. Neighboring nodes may receive multiple HELLO messages that pass the RSSI test. These receiver nodes maintain a list of their eligible parents (level N nodes) and send an ACK message to each one of them. Level N nodes do not immediately reply to each of the received ACK messages. Rather it maintain a list of all ACK messages that pass the RSSI threshold. Let Nel represent the number of potential level N+1 nodes. After timeout, level N nodes sends SELECT messages to all the eligible level N+1 nodes. SELECT messages contain the channel number, Ie and Nel values. Nodes receiving the SELECT messages compares the Ie values, selecting the level N node with the lowest interference value. For equal Ie values, level N node with lowest Nel is selected. Ie values control the internal interference while Nel does the load balancing among branches of a subtree. This load balancing mechanism ensures that the tree grows in depth (the chosen channel number continues to be selected down the topology). The level N+1 node now sends a CONFIRM message to their selected parent. Level N nodes now exactly knows how many children have joined it. Figure 5 shows these message exchanges where Node D sends a CONFIRM message to Node B as the received Ie value for Node B is less than that at Node A (2 vs 5). If no CONFIRM message is received by a level N node despite sending a SELECT message to its children, a new SELECT message with a force ag is sent again to one of the children informing that parent node would become a leaf node if it does not join that subtree. Child node receiving

164

Topology Formed by HopCount Assignment

100 91 90 80 70 24 60 50 40 30 92 20 10 42 0 0 120 10 6 20 117 25 76 55 87 51 39 88 30 4 33 40 31 38 8 21 27 30 81 53 118 85 29 2 26 95 22 44 78 61 35 50 46 109 11 101 74 60 43 70 110 108 73 100 32 80 90 100 0 0 16 111 68 48 89 10 19 36 13 54 107 83 66 71 115 116 103 60 52 47 9 99 79 15 106 75 94 97 104 67 69 49 112 102 7 70 77 80 119 64 3 114 72 28 96 40 113 56 100 14 20 90 23 5 63 58 37 12 41 45 59 62 1 84 40 34 30 92 18 93 82 10 42 120 10 6 20 117 25 55 76 17 70 24 86 60 50 31 38 21 80 90 8 99 79 15 91

Topology Formed by RMMTC Assignment

97 94 104 67 69 106 75 27 30 81 53 118 85 29 2 26 95 22 87 51 39 20 88 30 4 33 40 44 78 61 35 50 46 109 11 101 74 60 43 70 110 108 73 100 32 80 90 100 0 0 16 111 68 48 89 10 19 36 13 54 107 83 66 71 115 116 103 60 52 9 49 112 102 7 70 77 80 119 64 3 114 72 28 96 40 113 56 100 14 20 90 23 5 63 58 37 12 41 45 59 62 1 47 84 40 34 30 92 18 93 82 10 42 120 10 20 117 25 55 76 17 70 24 86 60 50 31 38 21 80 90 8 99

Topology Formed by Centralized ILP Assignment

91 94 79 15 106 75 27 30 81 53 118 85 29 2 26 95 22 87 51 6 20 39 88 30 4 33 40 44 78 61 35 50 46 109 11 101 74 60 43 70 110 108 73 100 32 80 90 100 16 111 68 48 89 10 19 36 13 54 107 83 47 66 71 65 93 82 115 103 9 84 60 116 98 97 104 67 69 49 112 102 105 57 7 70 77 80 119 64 3 114 72 28 96 40 113 52 56 14 20 90 23 5 58 63 37 12 41 45 59 62 1 34 18 17

105 98

57

105 98

57

86

65

65

(a) Fig. 6.

(b) Topology Formation for HopCount, RMMTC and Centralized ILP

(c)

this message, sends a LEAVE message to its earlier selected parent. The selected parent replies with SELECT message with information of number of children attached to it. If the parent has more than one child nodes, level N+1 nodes joins the parent who has sent the SELECT message with the force ag by sending CONFIRM message. If no reply is received for the SELECT message with force ag, node can send the same message to other level N+1 nodes from whom ACK were received. This process is now repeated for each new level of nodes till no ACK messages are received by advertising the HELLO messages. These nodes become the leaf of the formed subtrees. Nodes that do not receive any SELECT message in response to their ACK messages, eventually joins the neighbor with the highest RSSI value and become part of a subtree. If no links are detected with RSSI above the threshold value, the protocol has a fall back mechanism to select links with lower RSSI values. V. P ERFORMANCE E VALUATION In this section, we evaluate the network performance of RMMTC by simulations and by experiments on a real testbed. A. Simulation study We implemented our proposed RMMTC protocol in NS2 discrete event simulator to study the performance of our scheme in large networks to investigate its scalability. We compared its network performance with other related topology control protocols. The metric used for comparison is the percentage of successful packet reception (PRR) at the base station for all the nodes in the topology. Nodes are deployed randomly in a 100m x 100m target area. Nodes after forming a connected topology report their sensed data back to a centrally located base station (at position 50,50) after every 3 seconds. Note that as the default NS2 implementation does not support multi-channels, we modied the MAC layer to discard packets received from the physical layer on all channels except the one in use at that node. We assume that external interference from WiFi is present on all the three commonly used WiFi channels (1, 6 and 11). We also assume that all the nodes in the topology are inuenced by same amount of interference from WiFi transmissions. Thus the value of Rthresh is assumed as -75 dBm for all the nodes. ZigBee transmissions with RSSI values lower than -75 dBm experience packet drops probability @

10% for each 2 dBm decrease in RSSI values below -75 dBm with maximum packet drop probability of 50% for all RSSI values lower than -85 dBm. This packet drop policy has been implemented at the modied MAC layer in NS2. We compared the performance of RMMTC with centralized ILP assignment and two other distributed schemes. For centralized ILP topology assignment, we used MATLAB to discover links with lowest interference values (Ie ) to form multi-channel routing subtrees. The topology is then provided to NS2 with parent information as static routes to evaluate its network performance. The rst distributed scheme included in our simulation study is the hop count based topology formation (HopCount) based on distance vector routing [12]. In this protocol, topology is setup and routes are established based on the received hop count values in the route broadcast messages. The same topology is used for two simulation scenarios No WiFi and then subjected to WiFi interference (HopCount with WiFi). The second distributed scheme included in our comparison is the RSSI based topology formation which, is part of the RMMTC protocol but only uses a single channel (channel 13). Links in RSSI based scheme are only selected if the RSSI values are above the Rthresh value. Note that this RSSI based assignment takes care of external interference from WiFi but does not take any measures to combat co-channel interference and channel contentions. Each set of experiment was repeated ve times for different number of nodes.
100

%age Packet Reception Rate

90 80 70 60 50 40 HopCount no WiFi HopCount with WiFi RSSI with WiFi RMMTC with WiFi Centralized ILP with WiFi 80 100 120 Number of Nodes 140

Fig. 7.

Network Performance of RMMTC Protocol

Topology formed by HopCount, RMMTC and centralized ILP solution are shown in Figure 6. Simulation results in Figure 7 show that for HopCount PRR falls substantially in the presence of WiFi interference. For 80 nodes, PRR is reduced to about 69% from an average of about 98% when no WiFi is present. Similarly, for 140 nodes topologies, PRR reduces to 55% from about 94%. RSSI based topology

165

improves the packet reception rate by selecting links with better RSSI values. However, PRR is still lower than the HopCount - No WiFi scenario. The PRR is further improved by RMMTC by forming different routing trees based at the base station operating on non-interfering channels with RSSI based links selection. For 140 nodes topology, the RMMTC protocol reduces the percentage of dropped packets from 45% to less than 5%, a nine-fold reduction. Comparing the performance of RMMTC with centralized ILP assignment, RMMTC performs close to the ILP, especially for dense topologies. In dense topologies, it is easier to nd better RSSI based links and to assign different channels to routing trees due to presence of more neighbors.
Empirical CDF (No of Lost Packets)
1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 20 40 60 HopCount No WiFi HopCount With WiFi RSSI based With WiFi RMMTC With WiFi Centralized ILP With WiFi 80 100 120 140

Experimental Study: Comparison of Multi Hop Protocols

100 90 Min Max

%age Packet Reception Rate

80 70 60 50 40 30 20 10 HopCount MintRoute RSSI RMMTC

Fig. 10.

Comparison of Multi-Hop Protocols

Fig. 8.

Empirical CDF for 120 Nodes Topology

Figure 8 displays the empirical CDF for number of lost packets in a 120 node topology for the schemes under comparison. Result shows that for HopCount - with WiFi almost 70% of the nodes lost upto 32% of their packets (46 packets out of 140). RSSI-with WiFi improves the distribution (70% of the nodes now lost less than 10% of their packets). RMMTC has network performance closest to the centralized ILP assignment as shown by the CDF. For RMMTC, 98% of the nodes lost less than 9% of their packets. B. Experimental Evaluation

Fig. 9.

12 Nodes Test Bed Topology

We conducted experiments to verify the operation and performance of the RMMTC protocol in a real world setting. We used standard off-the-shelf Crossbow MicaZ sensor nodes for the experiments. These experiments were performed indoors in an environment where multiple WiFi Access Points (APs) were active on the commonly used WiFi channels (channels 1, 6 and 11) We also created WiFi interference through a laptop running a FTP client on WiFi channel 1. There was another WiFi AP active on channel 1. Channel 6 and 11, on the other

hand, had 1 and 3 active APs respectively at the time the experiments were conducted. We set up a 12 node indoor topology in a 2m wide L shaped corridor where nodes are placed on inverted foam glasses raised about 10 cm from the ground (See Figure 9). The node transmission power level was set at -15 dBm to ensure that multi-hop communication is required between the nodes. Nodes send a packet every 0.5 sec to the base station. These packets are carried over multiple hops to the base station where the received packet is logged with the sender ID and its corresponding sequence number. For comparison purposes, we implemented three of the protocols, used in the simulation study, in TinyOS (namely; HopCount, RSSI and RMMTC). In addition, we also included the widely used MintRoute multi-hop routing protocol [13] (implementation available in TinyOS 1.x distribution). In this protocol, routes are setup and maintained depending on the link quality estimations by snooping the received signal strength and packet losses etc. We compared the individual as well as cumulative Packet Reception Rate (PRR) achieved for all the nodes using different multi-hop topology control schemes. The default channel used for all protocols is ZigBee channel 13 (chosen to be under the WiFi inuence of FTP client operating on Channel 1) except for RMMTC for which the default control channel is 26 and channel 13 and 21 are used for data transfer phase. Default MAC for MicaZ was used for all the experiments. Rthresh was set at -82 dBm. Figure 10 shows the maximum and minimum PRR for the compared schemes. The overall reception rate for the HopCount protocol uctuates between a maximum of about 68% to minimum value of about 47%. This can be attributed to the fact that no consideration has been given to the potential interferences during the topology construction. MintRoute protocol and RSSI based protocol both show improvement in the overall packet reception rate with values ranging between 79% and 66% for MintRoute and 84% and 62% for RSSI respectively. This performance improvement is due to the fact that both these protocols consider the link performance parameters while selecting the link for topology construction/maintenance. This shows that even simple consideration of RSSI values during topology construction can considerably improve the network performance. In addition to reducing the WiFi interference by selecting higher RSSI values, RMMTC creates multiple collision domains (2 in this case) to considerably reduce channel contention and improve the average

166

HopCount 100 90 100 90

MintRoute 100 90

RSSI

%age Packet Reception Rate

%age Packet Reception Rate

80 70 60 50 40 30 20 10 1 2 3 4 5 6 7 Nodes 8 9 10 11 12

80 70 60 50 40 30 20 10 1 2 3 4 5 6 7 Nodes 8 9 10 11 12 13

%age Packet Reception Rate

80 70 60 50 40 30 20 10 1 2 3 4 5 6 7 Nodes 8 9 10 11 12 13

(a) Fig. 11.

100 90

(b)

(c)

Successful Packet Reception Rate for HopCount, MintRoute and RSSI based Protocol
RMMTC (Channel 21) 100 90

RMMTC (Channel 13)

%age Packet Reception Rate

80 70 60 50 40 30 20 10 2 4 5 7 Nodes 9 11 12

%age Packet Reception Rate

80 70 60 50 40 30 20 10 1 3 6 Nodes 8 10

Fig. 12.

Packet Reception Rate for RMMTC [3] R. Musaoiu-E. and A. Terzis, Minimizing the Effect of WiFi Interference in 802.15.4 Wireless Sensor Networks, 2008: Int. J. Sensor Networks, Vol. 3, No. 1, pages 4354. [4] K. Srinivasan, P. Dutta, A. Tavakoli, and P. Levis, An Empirical Study of Low-power Wireless, 2010: ACM TOSN, pages 1-49. [5] Y. Wu, J. A. Stankovic, T. He, J. Lu and S. Lin, Realistic and Efcient Multi-Channel Communications in Wireless Sensor Networks, 2008: Infocom, pages 1193-1201. [6] Crossbow Technology Inc., Avoiding RF interference between WiFi and Zigbee, 2004: Available at: http : //www.xbow.com/products/P roduct pdf f iles/ W ireless pdf /ZigBeeandW iF iInterf erence.pdf . [7] G. Zhou, C. Huang, T. Yan, T. He and J. A. Stankovic, MMSN: Multi Frequency Media Access Control for Wireless Sensor Networks, 2006: IEEE INFOCOM. [8] K. R. Chowdhury, N. Nandiraju, D. Cavalcanti, and D. P. Agrawal, CMAC: A Multi-channel Energy Efcient MAC for Wireless Sensor Networks,2006: IEEE WCNC. [9] M.D. Jovanovic and G.L. Djordjevic, TFMAC: Multi-channel MAC Protcocol for Wireless Sensor Network, 2007:TELSIKS07. [10] Y. Kim, H. Shin and H. Cha. Y-MAC: An Energy-efcient Multichannel MAC Protocol for Dense Wireless Sensor Networks. 2008: IPSN08. [11] N. Ahmed, S. Kanhere and S. Jha, Multi-Channel Interference in Wireless Sensor Networks, 2009: IPSN, pages 367-368. [12] C. E. Perkins and P. Bhagwat, Highly Dynamic Destination Sequenced Distance Vector Routing (DSDV) for Mobile Computers, 1994: ACM Computer Communication Review, pages 234-244. [13] A. Woo, T. Tang and D. Culler, Taming the Underlying Challenges of Reliable Multi-hop Routing in Sensor Networks, 2003: SenSys, pages 14-27. [14] K. Srinivasan, and P. Levis, RSSI is under appreciated, 2006: EmNets06. [15] J. Hauer, V. Handziski and A. Wolisz, Experimental Study of the Impact of WLAN Interference on IEEE 802.15.4 Body Area Networks, 2009: EWSN, pages 17-32. [16] W. Xu, W. Trappe and Y. Zhang, Channel Surng: Defending Wireless Sensor Networks from Interference, 2007:ACM IPSN07, pages 499508. [17] C. A. Boano, Z. He, Y. Li, T. Voigt, M. Zuniga and A. Willing, Controllable Radio Interference for Experimental and Testing Purposes in Wireless Sensor Networks, 2009: SenseApp, pages 865-872.

packet delivery rate to above 85%. Figures 11 and 12 show the individual average PRR for all the nodes in the topology. As expected, nodes located near the base station shows better PRR than those further away. Note that routing subtree working on channel 13 shows more packet losses as compared with that on channel 21, due to the WiFi interference from the FTP le transfer. This experimental study has highlighted that the effect of WiFi interference can be mitigated by selecting short, multi-hop links with adequate RSSI values as compared to long distance transmissions. Similarly, reducing the interference by incorporating multiple collision domains based on multiple channels considerably improves the performance. VI. C ONCLUSION This work shows that the use of multiple channels and links with adequate RSSI values can mitigate the effect of both external and internal interferences. We formulated an ILP model for describing the multiple channel assignment to overcome the effects of interferences. We proposed a fully distributed multi-channel topology control protocol that is shown to work close to the centralized optimal ILP solution. The performance evaluation is done by carrying out discrete event simulations and veried by deployment in a real test bed. The results show that our proposed protocol can successfully achieve acceptable network performance, more for dense networks, when exposed to interferences. R EFERENCES
[1] S.Y. Shin, H. S., Park, S., Choi, and W. H., Kwon, Packet Error Rate Analysis of IEEE 802.15.4 under IEEE 802.11b Interference, 2005: Wired/Wireless Internet Communications (WWIC05). [2] O.D. Incel, S. Dulman, P. Jansen, and S. Mullender, Multi-Channel Interference Measurements for Wireless Sensor Networks, 2006:LCN, pages 694-701.

167