A Near-Optimal Distributed QoS Constrained Routing Algorithm for Multichannel Wireless Sensor Networks

Transcription

1 Sensors 2013, 13, ; doi: /s Artice OPEN ACCESS sensors ISSN A Near-Optima Distributed QoS Constrained Routing Agorithm for Mutichanne Wireess Sensor Networks Frank Yeong-Sung Lin 1, Chiu-Han Hsiao 1, *, Hong-Hsu Yen 2 and Yu-Jen Hsieh Department of Information Management, Nationa Taiwan University, No. 1 Section 4, Roosevet Road, Taipei City 106, Taiwan; E-Mai: ysin@im.ntu.edu.tw Department of Information Management, Shih Hsin University, No. 1, Lane 17, Mu-Cha Road, Section 1, Taipei City 116, Taiwan; E-Mais: hhyen@cc.shu.edu.tw (H.-H.Y.); d @ntu.edu.tw (Y.-J.H.) * Author to whom correspondence shoud be addressed; E-Mai: chiuhanhsiao@gmai.com; Te.: ; Fax: Received: 28 October 2013; in revised form: 20 November 2013 / Accepted: 22 November 2013 / Pubished: 2 December 2013 Abstract: One of the important appications in Wireess Sensor Networks (WSNs) is video surveiance that incudes the tasks of video data processing and transmission. Processing and transmission of image and video data in WSNs has attracted a ot of attention in recent years. This is known as Wireess Visua Sensor Networks (WVSNs). WVSNs are distributed inteigent systems for coecting image or video data with unique performance, compexity, and quaity of service chaenges. WVSNs consist of a arge number of battery-powered and resource constrained camera nodes. End-to-end deay is a very important Quaity of Service (QoS) metric for video surveiance appication in WVSNs. How to meet the stringent deay QoS in resource constrained WVSNs is a chaenging issue that requires nove distributed and coaborative routing strategies. This paper proposes a Near-Optima Distributed QoS Constrained (NODQC) routing agorithm to achieve an end-to-end route with ower deay and higher throughput. A Lagrangian Reaxation (LR)-based routing metric that considers the system perspective and user perspective is proposed to determine the near-optima routing paths that satisfy end-to-end deay constraints with high system throughput. The empirica resuts show that the NODQC routing agorithm outperforms others in terms of higher system throughput with ower average end-to-end deay and deay jitter. In this paper, for the first time, the agorithm shows how to meet the deay QoS and at the same time how to achieve higher system throughput in stringenty resource constrained WVSNs.

2 Sensors 2013, Keywords: Wireess Sensor Networks; Wireess Visua Sensor Networks; system perspective; user perspective; QoS; Near-Optima Distributed QoS Constrained (NODQC) routing agorithm; Lagrangian Reaxation 1. Introduction In recent years, Wireess Visua Sensor Networks (WVSNs) have emerged as an interesting fied. Popuar appications are environmenta monitoring, seismic detection, miitary surveiance, medica monitoring, video surveiance, or the Internet of Things (IoT), etc. [1]. IoT integrates a part of the Future Internet and extends network capabiities to Machine-to-Machine (M2M) communications by a dynamic network infrastructure with sef-organization techniques. As compared to a WSN, the camera equipped sensor nodes in WVSNs can capture, process and transmit the rea time visua data (images, video). Since the visua data is much arger and compicated than scaar data, deay Quaity of Service (QoS) is a chaenging issue resource for constrained WVSNs. WVSNs attract more attention on how the camera sensor nodes can cooperativey pass their data more efficienty with minimum deay through the network on a arge scae or under reaistic physica environmenta conditions [2,3]. The architecture of WVSNs is shown in Figure 1. Figure 1. Wireess Visua Sensor Networks. Sink Sensor nodes Cameras In WVSNs for video surveiance, the deay is a very important QoS metric, and the deay is highy dependent on the routing strategies. However, deay routing in WVSN is more difficut than in WSN due to the fact visua data is much arger and compicated than scaar data. Furthermore, battery-imited powered and resource (channe-, CPU-) constrained camera nodes make this deay routing probem more chaenging in WVSNs. [4]. Basicay, for rea-time video surveiance appications, meeting the end-to-end deay QoS is the most important criteria. However, due to the imited resources of WVSNs, deay QoS routing decisions shoud aso consider overa resource utiization so that the future video surveiance appications coud be aso satisfied. Hence, besides meeting the end-to-end deay QoS for each user, minimizing the average deay in WVSNs is aso another crucia performance metric. In this paper, for the first time, we consider the deay QoS routing

3 Sensors 2013, strategy in WVSNs to meet the end-to-end deay QoS for each appication and at the same time minimize the average deay. 2. Literature Survey Deay QoS routing is a chaenging issue in WVSNs. Traditiona approaches for wireess sensor networks are not efficient or even feasibe in a WVSN due to three reasons: first of a, visua data are much bigger and compicated than scaar data, and processing and transmitting visua data wi consume much more system resources than scaar data. Secondy, WVSNs need efficient coaborative image processing and coding techniques that can expoit correation in data coected by adjacent camera sensor nodes. Video coding/compression that has ow compexity, produces a ow output bandwidth, toerates osses, and consumes as itte power as possibe is required. Third, it needs to reiaby send the reevant visua data from the camera sensor nodes to aggregation nodes or the sink node in an energy-efficient way [1,2]. This issue is reated to the transmission technique that consists of the routing assignment to provide energy efficient and stabe routes that meet the end-to-end QoS guarantees [2 4]. In wireess visua sensor networks for rea time video surveiance, sensor nodes need to capture and forward the packets to the sinks within an acceptabe deay under the imited resource constraints, incuding embedded vision processing, data communication, and battery energy issues. It is a chaenging issue since it needs to meet the end-to-end deay for each appication and at the same time optimize the system resource utiization by minimizing the average system deay so that more rea-time appications can be granted access in the near future. Existing research on wireess sensor networks address the ink scheduing techniques [5], channe assignments [6 8], routing agorithms [9,10], or resource aocations [11 13]. Some of the we-known routing protocos, such as Ad hoc On-demand Distance Vector (AODV), Dynamic Source Routing (DSR), Optimized Link State Routing (OLSR), and Destination Sequenced Distance Vector (DSDV) are the protocos appied to Ad hoc networks, such as MANETs, WMNs, VANETS, etc., that consist of hosts interconnected by routers without a fixed infrastructure that can be arranged dynamicay [14]. Even though ike in the MANETs, there is no centraized AP or base station in WVSN, sensor nodes are aso the reay nodes to forward the data to the sink node. However, without addressing the deay QoS, these approaches are not appicabe to rea-time appications. In the foowing, we survey existing approaches on shortest path routing, QoS routing, and routing agorithms. To the best of our knowedge, there is no research addressing the end-to-end deay from the user perspective and at the same time optimizing the system resource utiization by minimizing the average system deay Shortest Path Routing The routing agorithm is determined by the shortest path and the routing decision for the O-D pair. When a new route is estabished and a traffic sti foows the pre-existing path, this phenomenon is caed session routing because a path remains in force for the entire user session. The routing metric (i.e., cost) is used to define the path ength and can be measured in terms of hops, the mean deay, and even the extent of fow deviation. The mean deay can be used to find the fastest path for routing,

4 Sensors 2013, and fow deviation can be appied to system optimization routing where each route has equa fow deviation QoS Routing Quaity-of-Service (QoS) is a major issue that can be divided into severa factors incuding the reiabiity, mean deay, deay jitter, and bandwidth of packet transmissions [15]. Controing QoS is a chaenge faced by the routing of mutihop wireess sensor networks with interference [16]. In mutichanne wireess sensor networks, transmitted packets may coide in an interference environment if the nodes and neighbors within transmission range use the same channe. This adversey affects system performance and the QoS. Different types of appications have different QoS standards. As such, the QoS routing probems differ by the QoS types requirements of end users and therefore require carefu contro of channe assignment within an interference environment. By using different and non-overapping channes, each node can freey communicate with its neighboring nodes (i.e., the nodes within transmission range) without interference. As noted in [15], using mutipe channes instead of a singe channe in mutihop wireess sensor networks has been shown to not ony reduce packet contention probabiity in non-overapping channes, but to aso dramaticay improve the network throughput. A channe assignment strategy is necessary to efficienty use mutipe channes with commodity hardware; besides, the channe assignment can be changed with significant changes in traffic oad or network topoogy [7,8] Routing Protocos Routing protocos are categorized into tabe driven and on-demand, based on route cacuation. Researches indicate that a Mobie Ad hoc NETwork (MANET) has severa characteristics: (1) dynamic topoogies; (2) bandwidth-constrained inks; (3) energy constrained operation; and (4) imited physica security. Therefore the routing protocos for wired networks cannot be directy used for wireess networks [14,17,18]. The characteristics (1), (2) and (3) are simiar to WVSNs, which is the driving force to design a protoco to appy to this case. Proactive (tabe-driven) protocos are based on periodic exchange of contro messages and maintaining routing tabes. These protocos maintain compete information about the network topoogy ocay. On the other hand, reactive (on-demand) protocos try to discover a route ony on-demand, when it is necessary. These protocos usuay take more time to find a route compared to proactive protocos. The foowing sections briefy introduce three we known protocos appied to MANET: Optimized Link State Routing (OLSR), Ad hoc On-demand Distance Vector (AODV), and Destination Sequenced Distance Vector (DSDV). The deay, throughput, contro overhead and packet deivery ratio are the four common measures used for the comparison of the performance of the above protocos [18] Optimized Link State Routing (OLSR) (Tabe-Driven) OLSR is an optimization of a pure ink state agorithm for mobie Ad hoc networks that used the concept of Muti-point Reays (MPR) for forwarding contro traffic. It is a proactive ink-state routing protoco, which uses heo and topoogy contro (TC) messages to discover and then disseminate ink

5 Sensors 2013, state information. OLSR is intended for diffusion into the entire network. The MPR set is seected such that it covers a nodes that are two hops away. The forwarding path for TC messages is not shared among a nodes, but vary depending on the source, and ony a subset of nodes source ink state information, and not a inks of a node are advertised but ony those that represent MPR seections. Being a proactive protoco, routes to a destinations within the network are known and maintained before use. Having the routes avaiabe within the standard routing tabe can be usefu for some systems and network appications as there is no route discovery deay associated with finding a new route [19] Ad hoc On-demand Distance Vector (AODV) (On-Demand) AODV is a combination of on-demand and distance vector hop-to-hop routing methodoogy [7]. A connection broadcasts a request when a node needs to know a route to a specific destination. The route request is forwarded by nodes, and records a reverse route for itsef to the destination. When the request reaches a node with a route to the destination it creates a backwards message which contains the number of hops that are required to reach the destination. A nodes that participate in forwarding this repy to the source node create a forward route to the destination. This route created from each node from source to destination as a hop-by-hop state and not the entire route as in source routing [18]. The main advantage of this protoco is having routes estabished on demand and that destination sequence numbers are appied to find the atest route to the destination. The connection setup deay is ower. One disadvantage of this protoco is that intermediate nodes can ead to inconsistent routes if the source sequence number is very od and the intermediate nodes have a higher but not the atest destination sequence number, thereby having stae entries. Aso, mutipe route repy packets in response to a singe route request packet can ead to heavy contro overhead. Another disadvantage of AODV is unnecessary bandwidth consumption due to periodic beaconing [18] Destination Sequenced Distance Vector (DSDV) (Tabe-Driven) DSDV is a hop-by-hop distance vector routing protoco. It is a tabe-driven routing scheme for Ad hoc mobie networks based on the Beman-Ford agorithm. A routing tabe isting the next hop is maintained in each node for the reachabe destination. Number of hops to reach destination and the sequence number are assigned by destination node. Each entry in the routing tabe contains a sequence number used to distinguish oder routes from newer ones and thus avoid oop formation. The nodes periodicay transmit their routing tabes to their immediate neighbors. Routing information is distributed between nodes by sending fu dumps infrequenty and smaer incrementa updates more frequenty, so the updating process is both time- and event-driven. DSDV requires a reguar update of its routing tabes, which uses up battery power and a sma amount of bandwidth even when the network is ide. Whenever the topoogy of the network changes, a new sequence number is necessary before the network re-converges. DSDV is not suitabe for highy dynamic networks [20]. Tabe 1 summarizes a comparison of these routing protocos.

6 Sensors 2013, Tabe 1. Routing protoco comparison. Routing Protoco NODQC AODV OLSR DSDV Type Reactive/On-demand Reactive/On-demand Proactive/Tabe driven Proactive/Tabe driven Optimization with Optimization: MutiPoint Quick adaptation under construct shortest paths Reays (MPRs) dynamic ink Adds two things to within an acceptabe Forward broadcast conditions distance-vector routing deay. messages during the Lower transmission Sequence number; Lower average end-to-end fooding process Short atency avoid oops deay and deay jitter Ony partia ink state is Description Consume ess network Damping; hod Higher system throughput distributed bandwidth (ess advertisements for with QoS satisfaction in Optima routes (in terms broadcast) changes of short arge networks of number of hops) Loop-free property duration Scaabe to arge Suitabe for arge and networks dense networks Distributed Yes Yes Yes Yes Unidirectiona Link Support Yes No Yes No Periodic Broadcast Yes Yes Yes Yes QoS Support Yes No Yes No Arc weight of ink can be determined and prioritized Faster decision making Suitabe for arge and dense networks Construct shortest paths within an acceptabe Advantages deay. Lower average end-to-end deay and deay jitter Higher system throughput with QoS satisfaction in arge networks Low compexity with more efficiency and effectiveness Constrains must be reaxed Disadvantages Proprietary design Appications oriented Simpicity Lower connection Lower route request setup time than DSR atency, but higher Sequence numbers overhead used for route Incrementa dumps and freshness and oop Suitabe for arge and setting time used to prevention dense networks reduce contro Maintains ony active Independent from other overhead routes protocos Perform best in Uses sequence network with ow to numbers to determine moderate mobiity, few route age to prevent nodes and many data usage of stae routes sessions Count-to-infinity probem Link break detection Lack of security Not efficient for arge adds overhead No support for muticast Ad hoc networks Sti have possibe Overhead: maintaining Routing information is atency before data routes to a nodes is often broadcasted transmission can begin unnecessary periodicay and incrementay

7 Sensors 2013, Motivation and Paper Organization For rea-time WVSN appications, the routing strategies shoud meet the end-to-end deay requirements and at the same time optimize the system resources so that more rea-time appications coud be admitted in the near future. To the best of our knowedge, there is no existing iterature addressing these two issues at the same time. In this paper, for the first time, we propose WSN routing strategies that meet the end-to-end deay requirement and ate the same time optimize the system resources by minimizing the average deay. The rest of this paper is organized as foows: in Section 3, we propose the routing mathematica mode to satisfy the end-to-end deay and minimize the average deay in the WSN. In Section 4, we present the soution approaches for our mode and deveop a heuristic to get a prima feasibe soution. In Section 5, computationa experiments are performed to verify the soution quaity of our approach. Finay, we concude this paper and outine future research in Sections 6 and Probem Formuation 3.1. Mathematica Modeing The environment considered here is mutichanne WSNs. This paper addresses the probem of determining a good QoS with the average cross-network packet deay whie taking user perspective and system perspective into consideration. The foowing are detais of the probem identification: Assumptions: 1. The channe assignment for each sensor node is fixed for a ong period. 2. Each sensor node is stationary. 3. Each sensor node is equipped with mutipe wireess interfaces, each of which operates on an individua and non-overapping channe. 4. Each sensor node can simutaneousy communicate with its neighbors within transmission range without interference from the use of different channes for each ink. 5. A virtua node is added as the destination node to ony connect to the sink via wired-ine. 6. A fows are transmitted to this virtua node via the sink. Given: 1. The set of inks. 2. The set of sensor nodes. 3. The ink capacity of each ink. 4. The number of interference inks of each ink. 5. The traffic requirement for each O-D pair. Objective: To minimize the average cross-network packet deay of the WSNs. Subject to: 1. QoS constraints. 2. Path constraints. 3. Capacity constraints. 4. Fow constraints. To determine: The minimum the average cross-network packet deay of the WSNs take system perspective and user perspective into consideration for each O-D pair.

8 Sensors 2013, We propose a channe assignment heuristic and routing agorithm to maximize the transmission rate for each node and enhance system performance by WVSN panning. A simiar formuation that aimed to minimize the average cross-network packet deay subject to end-to-end deay constraints for users was proposed by Yen and Lin [21]. This paper s formuation can be extended to use queue modes. Once the channe is assigned to the ink, two nodes wi use the assigned channe to communicate with each other unti the channe assignment is changed (i.e., static channe assignment). Every ink is assumed to be fairy used at the scheduing phase, meaning the average time data transmission over each ink is equa. In mutichanne WVSNs, ink capacity degrades because other inks use the same channe in the interference range. Therefore, we divide the ink capacity by the number of interference C I () inks in the foowing formuation (i.e., C ) [6]. The foowing notations ist the given parameters and the decision variabes of our formuation, as iustrated in Tabes 2 and 3. Objective Function (IP): min D ( g ) g L Subject to: D ( g ) yw D L pp w pp w 0 g w (IP) ww (IP 1) x p 1 ww (IP 2) x p y ppw ww p C I( ) p w x p g w ww, L (IP 3) L (IP 4) L (IP 5) x p 0 or 1 ppw, ww (IP 6) y w 0 or 1 ww, L. (IP 7) The objective function is iustrated as (IP) to minimize the mean deay on ink. The expression of the form D (g ) is a monotonicay increasing and convex function with respect to g, where g is the aggregate fow on ink measured in packets per second [6 9]. Tabe 2. Given parameters. Notation Description W The set of Origin-Destination (O-D) pairs in the WSNs, where w W. P w The set of directed paths from the origin to the destination of O-D pair w, where p P w. L The set of communication inks in the WSNs, where L. I() The number of interference inks of ink. n The indicator function which is 1 if ink is on path p and 0 otherwise. C (packets/s) The ink capacity of ink. w (packets/s) The given traffic input of O-D pair w. D (g ) The mean deay on ink, which is a monotonicay increasing and convex function of aggregate fow g. D w The maximum aowabe end-to-end QoS for O-D pair w.

9 Sensors 2013, Tabe 3. Decision variabes. Notation Description x p 1 if path p is used to transmit the packets for O-D pair w and 0 otherwise. y w 1 if ink is on the path p adopted by O-D pair w and 0 otherwise. g (packets/s) The estimate of the aggregate fow on ink Expanation of the Objective Function Objective function (IP) is actuay the summation of average number of packets on each ink, i.e., the queue ength, which is obtained by the product of ink mean deay and aggregate fow on the ink. The expression of the form D (g ) is a monotonicay increasing and convex function with respect to g, where g is the aggregate fow on ink measured in packets per second [15,21 23]. Through Litte s Law (i.e., N = T), the objective function is proportiona to the average cross-network packet deay. Thus, for a given traffic input (i.e., w ), minimizing the average ww number of packets in the network is equivaent to minimizing the average cross-network packet deay [24,25] Expanation of the Constraints QoS constraints: Constraint (IP 1) confines that the end-to-end deay shoud be no arger than the maximum aowabe end-to-end QoS requirement. Path constraints: Constraint (IP 2) confines that a the traffic required by each O-D pair is transmitted over exacty one candidate path. Constraint (IP 3) confines that once path p is seected and ink is on the path, y w must be equa to 1. Capacity constraints: Constraint (IP 4) confines the boundaries of aggregate fow on ink. Fow constraints: Constraint (IP 5) confines the aggregate fow on ink shoud not exceed the ink capacity. Integer constraints: Constraint (IP 6) and (IP 7) are the integer constraints of decision variabes Lagrangian Reaxation The Lagrangian Reaxation Method was introduced in the 1970s to sove arge-scae mathematica programming probems and was foowed by a arge amount of subsequent research [26,27]. Due to its versatie practicaity, the Lagrangian Reaxation Method has become a widey used too for deaing with optimization probems such as integer programming probems and even non-inear programming probems. The main idea of the Lagrangian Reaxation method is to pu apart the mode by reaxing (i.e., removing) compicated constraints in the prima optimization probem. The next procedure coud be modified by the objective function corresponding to the associated Lagrangian mutipiers of the

10 Sensors 2013, reaxed constraints. The prima optimization probem can be transformed into a Lagrangian Reaxation form. The Lagrangian Reaxation probem is separated into severa independent sub-probems for each decision variabe or other rues by appying the decomposition method. For sub-probems, we can design some heuristics or agorithms to appy and find the optima vaue. Figure 2. Concept of Lagrangian Reaxation Method. Figure 3. Lagrangian Reaxation Procedures. Initiaization Sove Lagrangean Dua Probem Get Prima Soution Update Bounds Adjust Mutipier Check Termination F If ( Z* - LB ) / min ( LB, Z* ) < ε or k reaches Iteration Counter Limit or LB = Z*? T STOP

11 Sensors 2013, Through the adoption of Lagrangian Reaxation, a compicated programming probem can be viewed as a sma set of easy-to-sove probems with side constraints. The method simpifies the origina probem by decomposing it into severa independent sub-probems, each with their own constraints and each of which can be further soved by other we-known agorithms. For exampe, if the probem is a minimization probem, the optima vaue of the reaxed constraints is aways a ower bound on the optima vaue of origina probem under the reaxed conditions. The ower bound can be improved by adjusting the set of mutipiers iteration by iteration to reduce the gap of the soution between the prima probem and the Lagrangian Reaxation. This procedure is aso caed the Lagrangian Dua probem. Figures 2 and 3 iustrate the underying idea of the Lagrangian Reaxation Method, respectivey. The Lagrangian Reaxation of the prima probem is deveoped to be a ower bound of the optima vaue for the origina minimization probem because some constraints of the origina probem are reaxed. Therefore, it can be used as the boundary to design heuristic agorithms to get the prima feasibe soution. The Subgradient Optimization Method can be utiized to minimize the gap between the prima probem and the Lagrangian Reaxation probem; this can be done through deriving the tightest ower bound by adjusting the mutipiers for each iteration and then updating them to improve resuts. 4. Soution Approach The soution approach to the probem formuation is based on Lagrangian Reaxation. Constraints (IP 1), (IP 3) and (IP 5) are reaxed and mutipied by nonnegative Lagrangian mutipiers, 1 w, 2, and w, 3 respectivey. They are added to the objective functions as foows. The constraints are reaxed in such a way that the corresponding Lagrangian mutipiers 1 w, 2, and w, 3 can be reaxed to objective function. This approach is not guaranteed to have a feasibe soution due to accessing a big vaue of w into the network which wi resut in unsatisfactory QoS requirement to the O-D pair w. ZLR min (,, ) w w D ( g ) g L 1 w( ( g w ww L 2 ( x p p w L pp www D ) y D ) ( 3 w ww L pp w Subject to: x p w g ) p y x p 1 pp w ww C 0 g I( ) w ) (LR) (IP 2) L (IP 4) x p 0 or 1 ppw, ww (IP 6) y w 0 or 1 ww, L. (IP 7) We use the Lagrangian Reaxation Method to reax constraints in the formuation and decompose (LR) into two independent sub-probems. In the first Sub-Probem (SP 1), the nonnegative weight is cacuated as ( 3 w + 2 w ) on each ink. The actua meaning of the mutipiers 3 w and 2 are the ink

12 Sensors 2013, mean deay and the derivative of queue ength, respectivey; both can be derived from the soution of the probem (LR). By the probem (LR), the mutipier 3 w is reated to the reaxation of ink seection for each O-D pair. When one unit on the decision variabe of ink seection is added (i.e., y w ), it generates one unit of the corresponding mean deay as formuated in QoS constraint, but when the probem (LR) is soved, the decision variabe y w is either 1 or 0 not ony one unit, thus, the mutipier 3 w is equivaent to the mean deay on the chosen ink. Besides, the ink seection is based on the sum of ink mean deay D (g ) of each O-D pair which is weighted by 1 w, that is, how much the end-to-end deay wi impact the objective function (IP). Moreover, in the probem (SP 2.1 ), the first step for soving it is aso reated to the mutipier 3 w and chooses the inks for each O-D pair according to y w(g * ) = 1 wd (g ) 3 w = 0. It means that whenever a ink is seected for an O-D pair, the ink wi produce corresponding ink mean deay 3 w and affect the end-to-end deay constraint which is weighted by mutipier 1 w in the probem (LR). On the other hand, the mutipier 2 is reated to the reaxation of aggregate fow over each ink. The physica meaning of 2 woud be that when 2 can be reaxed on the ink fow, it wi affect the objective function (IP), which means 2 woud be the sensitivity of the average number of packets over the ink with respect to the aggregate fow. A brief iustration is shown in Tabe 4, 3 w impies the ink mean deay of each ink, and 2 indicates the derivative of queue ength. The term w in the arc weight form represents the weighting factor between 3 w and 2. Thus, this metric consists of the ink mean deay and the derivative of queue ength, considers both system perspective and user perspective for our distributed routing protoco, and uses the traffic requirement of each O-D pair as the weighting factor to combine these two parameters. Finay, The LR probem can be decomposed into independent and sovabe optimization sub-probems which are deveoped in a sufficient way one by one. The physica meanings of mutipiers are aso deveoped and mentioned above to hep us to determine reationships of the ink mean deay and derivative of queue ength. We can get these important parameters to derive the routing assignments in system perspective and user perspective. The detai soving steps are iustrated in the Appendix. Tabe 4. Arc weight of each ink. arc weight = ( 3 w + 2 w ) = ink mean deay derivative of queue ength required traffic D = D ( g ) ( g ) g + w g = D ( g ) + D ( g ) w D ( g ) g g In foowing section, the computationa experiments are constructed and impemented to anayze the quaity of the soution approach to verify the routing agorithms correcty. The experiments are designed for anayzing the performance of WVSN, if it can be satisfied with ower average end-to-end deay and deay jitter to transmit video data by different agorithms.

13 Sensors 2013, Computationa Experiments and Resuts 5.1. Experiment Environment Each session foows one path for routing its required traffic emuated as a rea time surveiance streaming data and is not aowed to change during the hoding time. The session arrivas foowed a Poisson process, and the hoding time of each session is set to be an exponentia distribution with average 10 s. The experiment with average session arriva rates of 0.25, 0.5, 1 and 2, and the corresponding packet arriva rates are 40, 20, 10 and 5. Figure 4 is shown as the experiment environment which QoS requirements are set in 3 3, 5 5, 7 7 and 9 9 squares in 3 ms, 4 ms, 5 ms and 6 ms, respectivey. Figure 4. Experimenta environment (5 5). Tabe 5. Experimenta session types. Parameters Vaue Average Hoding Time 10 Average Session Arriva Rate Average Number of Active Sessions Packet Arriva Rate Average Traffic Input 100 (packets/s) As Tabe 5 shows, there is average of 100 packets per second in the network of each topoogy. The size of each transmitting UDP packet is 1,000 bytes at each packet arriva rate. Aside from the contro messages of distribution routing protoco, messages such as HELLO messages for sensing the neighbors and TC message for broadcasting topoogy information are sent at a fixed period of 5 s. The estimate the parameters of our routing metric (shown in Tabe 6), we record each packet deay and packet inter-arriva time and then cacuate mean deay and aggregate fow with corresponding queue ength (i.e., average number of packets) of each ink for every second. The number of data fit for power regression is set to be 10, that is, the data used to compute the regression function are the newest 10 records and the interva of each is 1 s. Additionay, the vaue of K in the routing agorithm is set as 5 (i.e., at most the five shortest or fastest paths for each O-D pair) and the threshod for admission contro heuristic agorithm is set to be 1.3.

14 Sensors 2013, Tabe 6. Parameters. Parameters Vaue Transmit and Receive Antenna Gain 1.0 Transmit and Receive Antenna Height 1.5 (m) Reception Threshod e 10 Carrier Sensing Threshod e 11 Transmission Range 250 (m) Interference Range 550 (m) Distance between Each Node 200 (m) UDP Packet Size 1,000 (bytes) Sending Interva of HELLO Message 2 (s) Sending Interva of TC Message 5 (s) Recording Interva of Fitting Data 1 (s) Number of Fitting Data 10 K If the received signa strength exceeds the reception threshod, the packet can be successfuy received. If received signa strength exceeds the carrier-sensing threshod, the packet transmission can be sensed. However, the packet cannot be decoded uness signa strength surpasses the reception threshod. Both the transmission range and interference range are cacuated by the two-ray ground refection mode according to the reception threshod and the carrier-sensing threshod, respectivey. This paper focuses on the effectiveness of our routing metric and reated parameters at the on-ine stage, and its reduction of system-wide impact between each coming session with QoS provisioning. Thus, we use a singe channe in the experimenta environment to decoupe the effect of the channe assignment agorithm and evauate the performance of our routing agorithm Performance Evauation In this Section, we present experimenta resuts to demonstrate the effectiveness of our routing agorithm, NODQC. We aso evauate NODQC with other different session types of agorithms under the same average traffic oading in the network, and then compare the performance in terms of average end-to-end deay, deay jitter and system throughput with QoS satisfaction. Deay jitter is defined as the variance in the foowing tabes and figures. Performance evauation is defined in Tabe 7, a 7 7 square is simuated as the network topoogy with 660 s and measure the packets at ast 600 s. For comparison with other routing agorithms, we use the session types whose average session arriva rate and packet arriva rate are equa to 0.25 and 40, respectivey, and use different network sizes to show the performance of NODQC, OLSR, AODV, and DSDV routing agorithms. The time of our experiment is set at 660 s and the measurement time is the ast 600 s. Increasing the number of nodes wi aso increase the number of broadcast packets and route tabe updating. From Tabes 8 and 9 and Figure 5, the vaue of average end-to-end deay is shown to be increased and the system throughput with QoS is shown to be decreased with the average session arriva rate. This is because nodes exchange information every 5 s, but the sessions enter the network at a short time interva. Most sessions are consequenty routed to sub-optima paths due to a ack of

15 System Throughput with QoS (Kbps) Average End-toEnd Deay (ms) Sensors 2013, updated information for the routing metric. Evauation resuts indicate that our routing agorithm performs better in the ower average session arriva rate but higher packet arriva rate under the same system oading of average 100 packets per second. Tabe 7. Two-Ray Ground Refection Mode. Transmission Range (or Interference Range) Tabe 8. Evauation with different session rates (average end-to-end deay). Average End-to-End Deay (ms) Average Session Rate Packet Arriva Rate Average End-to-End Deay Tabe 9. Evauation with different session rates (system throughput with QoS). System Throughput with QoS (Kbps) Average Session Rate Packet Arriva Rate System Throughput with QoS Figure 5. Evauation with different session arriva rates. Throughput with Qos (Kbps) Average End-to-End Deay (ms) 800 Average Session Arriva Rate Packet Arriva Rate 2 If the performance evauation is taken in the scaabiity scenario, Average end-to-end deay (AE2ED) shows that it has no effect significant difference in smaer number of nodes obviousy, but deay increases as the number of nodes increase. Tabes 10 12, Figures 6 8 show the resuts of performances between different routing agorithms in each network size. In the sma-scae network (e.g., 3 3 and 5 5 squares), the broadcasting of routing protoco contro messages for exchanging information wi causes arge deays to packet transmission, especiay for the destination node. Additionay, the average path ength of each session is shorter in smaer networks, in which the superiority of our routing metric is not obvious.

16 Average End-to-End Deay (ms) Sensors 2013, Tabe 10. Experiment resuts of routing agorithms (average end-to-end deay). Average End-to-End Deay (ms) Routing Agorithms NODQC OLSR AODV DSDV Tabe 11. Experiment resuts of routing agorithms (deay jitter). Deay Jitter (ms 2 ) Routing Agorithms NODQC OLSR AODV DSDV Tabe 12. Experiment resuts of routing agorithms (system throughput with QoS). System Throughput with QoS (Kbps) Routing Agorithms NODQC OLSR AODV DSDV Figure 6. Experiment resuts of routing agorithms (average end-to-end deay). AODV DSDV OLSR NODQC Number of Nodes As networks grow in size, the average path of each session engthens with path seection becoming increasingy important. AE2ED taken in the scaabiity scenario shows that deay is much ess for NODQC as compared to others. As routes break, nodes have to discover new routes which ead to onger E2ED (packets are buffered at the source during route discovery). According to the arc weight on each ink is the combination of end-to-end deay from the user perspective and average deay from the system perspective, NODQC can choose other ess or non-congested paths for the sessions and baance network-wide the traffic oading. This makes the routing strategy more fexibe for making

17 System Throughput with QoS (Kbps) Deay Jitter (ms^2). Sensors 2013, routing decisions for new routing constructions. As far as deay is concerned, NODQC performs better than OLSR, AODV, and DSDV with arge numbers of nodes. Figure 7. Experiment resuts of routing agorithms (deay jitter). AODV DSDV OLSR NODQC Number of Nodes Figure 8. Experiment resuts of routing agorithms (system throughput with QoS). 800 NODQC OLSR AODV DSDV Number of Nodes Hence for rea time traffic, NODQC is preferred over OLSR, AODV, and DSDV. This is significant for the fact that, as the variations of packet deay becomes more predictabe, the routing mechanisms can factor in that deay to determine whether a packet is ost or not. Furthermore, the NODQC aso takes QoS provisioning into account for the coming sessions, thereby causing system throughput with QoS satisfaction to be greater than others in arger networks. By taking into consideration both the perspective of the system as we as we as users, the NODQC has proven itsef to have ower average end-to-end deay and deay jitter and higher system throughput with QoS satisfaction than other routing agorithms in arge-scae networks. By everaging the Lagrangian Reaxation method, the arc weight on each ink can be appied by our routing metrics. It can empoy the ink-state routing protoco to construct the shortest or shorter paths within an acceptabe deay. Within arger networks, routing strategies can be payed to improve system

18 Sensors 2013, performance with ower average end-to-end deay and deay jitter than aternative agorithms (OLSR, AODV, DSDV). NODQC outperforms them in terms of system throughput with QoS satisfaction. 6. Future Work One of the most important parts of this paper is the routing metric and reevant parameters. The Lagrangian Reaxation formuations are appied to the arc weight and to infer the actua meaning of the corresponding mutipiers. In addition, Lagrangian mutipiers and Karush-Kuhn-Tucker Conditions can be used to succincty describe the arc weight form. This simpifies the process of inference and has the same consequence with Lagrangian Reaxation. Through the soution of the Lagrangian Reaxation formuation, various QoS requirements can be impemented for different purposes or services. The end-to-end deay is presented as a function in our formuation to be the QoS consideration. Distinct QoS metrics (e.g., deay jitter or packet oss rate) can be appied to the Lagrangian Reaxation formuation, and the corresponding mutipier of the routing metric, ink mean deay wi be repaced by the newy defined QoS metric. 7. Concusions In wireess visua sensor networks for rea time video surveiance, sensor nodes need to be forwarded the packets to the sinks within an acceptabe deay under imited resource constraints, incuding embedded vision processing, data communication, battery energy issues. It is a chaenging issue since it needs to meet the end-to-end deay for each appication and at the same time optimize the system resource utiization by minimizing the average system deay so that more rea-time appications coud be granted in the future. By everaging the Lagrangian Reaxation method, the arc weight on each ink is the combination of end-to-end deay from the user perspective and average deay from the system perspective. Thereafter, our routing metric empoys the ink-state routing protoco to construct shortest paths within an acceptabe deay. Within arger networks, the superiority of the Near-Optima Distributed QoS Constrained (NODQC) routing agorithm is more visibe as path seection pays a more important roe. Experimenta resuts show that, and especiay so in arge-scae networks, the NODQC not ony has ower average end-to-end deay and deay jitter than aternative agorithms (OLSR, AODV, DSDV) but aso outperforms them in terms of system throughput with QoS satisfaction. In WVSNs for video surveiance, video coding/compression that has ow compexity, produces a ow output bandwidth, toerates oss, and consumes as itte power as possibe is required. Our routing strategies can be appied we in a arge scae network with more efficiency and effectiveness. Acknowedgements This work was supported by the Nationa Science Counci, Taiwan (grant No. NSC E ). Confict of Interest The authors decare no confict of interest.

19 Sensors 2013, Appendix Probem Soving Steps The soution approach to the probem formuation is based on Lagrangian Reaxation. Constraints (IP 1), (IP 3) and (IP 5) are reaxed and mutipied by nonnegative Lagrangian mutipiers, respectivey. They are added to the objective functions as foows: ZLR min (,, ) w w D ( g ) g L 1 w( ( g w ww L 2 ( x p p w L pp www D ) y D ) ( 3 w ww L pp w Subject to: x p w g ) p y x p 1 pp w ww C 0 g I( ) w x 0 or 1 ppw ww p ) (LR) (IP 2) L (IP 4), (IP 6) y w 0 or 1 ww, L. (IP 7) The constraints are reaxed in such a way that the corresponding Lagrangian mutipiers 1 w, 2, and 3 w are non-negative. This approach is not guaranteed to have a feasibe soution due to accessing a big vaue of w into the network which wi resut in unsatisfactory QoS requirement to the O-D pair w. However, LR can be decomposed into the foowing two independent and sovabe optimization sub-probems in a sufficient way. Step 1: Sub-probem 1 (reated to x p ) Objective Function: Z 2 3 (, ) Sub1 w min ww ppw L Subject to: pp w 3 2 ( ) x w x p 1 ww w x 0 or 1 ppw ww p p p (SP 1) (IP 2), (IP 6) y w 0 or 1 ww, L. (IP 7) The probem can be further decomposed into W independent shortest path probems. Each shortest path probem can be easiy soved by Dijkstra s Agorithm with nonnegative arc weights. ( 3 w + 2 w ) is the exact arc weight determined from our distributed routing protoco.

20 Sensors 2013, Step 2: Sub-probem 2 (reated to y w and g ) Objective Function: Z (,, ) Sub2 w w min[ L Subject to: C 0 g I( ) ( D ( g ) g y D ( g ) g y ) D ] w ww w w ww w w ww w (SP 2) L (IP 4) y w 0 or 1 ww, L. (IP 7) In objective function (SP 2), the ast term 1 D w w is not changed to determine the optima ww soution. It can be disregarded first and then added back to the objective vaue. Therefore, (SP 2) can be reformuated and further decomposed into L independent sub-probems. The objective function (SP 2) can be determined to function (SP 2.1) for each ink. 1 min[ D ( g ) g wy w W Subject to: C 0 g I( ) w D 2 3 ( g ) g y ] w W w w (SP 2.1) L (IP 4) y w 0 or 1 ww, L. (IP 7) In probem (SP 2.1), the first term D (g )g is a monotonicay increasing and nonnegative function. If the minimum vaue of the other terms in (SP 2.1) can be determined, the first term D (g )g can be eiminated in the first step. (SP 2.1) can be expressed as (SP 2.1 ) for each ink, 1 min[ w w W Subject to: C 0 g I( ) y w D 2 3 ( g ) g y ] w W w w (SP 2.1 ) L (IP 4) y w 0 or 1 ww, L. (IP 7) Tabe A1. Agorithm for soving probem (SP 2.1 ). Step 1: Sove y * w(g ) = 1 wd (g ) 3 w = 0 for each O-D pair w, and define the resuts as a set of break points of g. Step 2: Sort these break points and denote as g 1, g 2,,g n. There are at most W break points. i i1 Step 3: At each interva g g g, y w(g * ) is 1 if 1 wd (g ) 3 w 0 and is 0 otherwise. 1 * i 3 * i i i1 Step 4: Denote wy w( g ) as a and wyw( g ) as b, within the interva g g g, the probem ww ww (SP 2.1 ) can be expressed as: a D (g ) 2 g b. Thus, the oca minimum is either the boundary point, g 1 i1 * * D ( g ) or g, or at point g, where g is the soution to a 2 0. g Step 5: By examine the W + 1 intervas, the goba minimum point can be found by comparing these oca minimum points.

21 Sensors 2013, Figure A1. Typica graph of probem (SP 2.1 ). Finay, this probem can be soved by the agorithm deveoped by Cheng and Lin in [28]. The soution steps and graphics of the agorithm are briefy described in Tabe A1 and Figure A1. Step 3: The Dua Probem and the Subgradient Method According to the Weak Lagrangian Duaity Theorem [29], for the 1 w, 2, 3 w 0, the objective vaue of the Lagrangian Reaxation probem Z LR ( 1 w, 2, w) 3 is a ower bound of the prima probem Z IP. Based in probem (LR), the foowing dua probem Z D is constructed to cacuate the tightest ower bound. Dua Probem (D): Z D = max Z D ( 1 w, 2, 3 w) Subject to: 1 w, 2, 3 w 0. (D) We use the Subgradient Method is used to sove that [26,27] to sove the dua probem (D). First, et the vector S be a subgradient of Z D ( 1 w, 2, w). 3 In iteration k of the Subgradient Optimization Procedure, the mutipier vector k = ( 1 w, 2, w) 3 k k k k is updated by 1 t S. The step size t k is h k k ( Z IP Z D ( )) h determined by t 2. Z IP is the prima objective function vaue (an upper bound on Z IP ) and is constant where 0 2. Step 4: Getting Prima Feasibe Soutions S k A prima feasibe soution to (IP) by definition must aso be a soution to (LR) and satisfy a constraints. Otherwise, it must be modified to be feasibe to (IP) by a getting prima feasibe soutions heuristic. According to the computationa experiments in [21], the heuristic can be better soutions that consider the end-to-end deay constraints. That is, if the end-to-end deay constraints are not satisfied, more arc weights aong those paths vioate the end-to-end deay constraints in which case routing assignments must then be recacuated. Note that, this approach requires a few minutes to obtain the feasibe soution or even the near-optima soution. However, this paper focuses on the dynamic network environment and the derivation of the arc weight. Therefore, this arc weight wi be our routing metric used by the distributed routing protoco for dynamic routing.

22 Sensors 2013, Step 5: Routing Metric The cost of passing through the network can be used as a metric. By virtue of the routing agorithm, each router chooses the path with the smaest (shortest) sum of the metrics. Different routing protocos define the metric distincty. The metric can be distance, number of hops, deay, and so on. Here, the metric is defined as a combination of the ink mean deay and the derivative of queue ength for each ink. Step 6: Estimation of Routing Metric Parameters Perturbation Anaysis In [25], Towsey et a. proposed Perturbation Anaysis (PA), a usefu technique for estimating the so-caed margina deay defined in [24]. More specificay, this method is used to estimate the derivative of queue ength. As previousy mentioned, the derivative of queue ength is composed of ink mean deay, aggregate fow, and the partia derivative of ink mean deay. The PA approach can aso estimate the vaue of ink mean deay, a step that precedes the cacuation of the derivative of queue ength. One can use the PA technique to determine the ink mean deay and the derivative of queue ength, both major components of the distributed routing protoco. Time Stamping for Packets Other approaches can aso estimate critica information of each ink in a conceptuay straightforward way. During packet reception, packet deay time is defined as the difference between sending time and receiving time. For each adjacent ink mean deay, it the sum of each packet deay divided by the number of received packets corresponding to each neighbor node. Approximation Function For the derivative of queue ength, queue ength and the corresponding aggregate fow over the ink can be recorded at each time interva. Figure A2 is iustrated a monotonicay increasing and convex function such as the power regression function (i.e., y = ax b, where a 0 and b 1) can be used to approximate the recording resuts, and the newest record has the most impact to this approximation. Therefore, the derivative of queue ength can be obtained by cacuating the derivative of the function with respect to the corresponding aggregate fow on the ink. Figure A2. Estimation of the derivative of queue ength.

23 Sensors 2013, Step 7: Distributed Routing Agorithm The routing protoco is proposed and based on the we-known ink-state routing protoco [16]. Link-state routing protoco assigns a cost (i.e., metric) to each route. The ink mean deay is provided to be a metric by the present routing protoco. The derivative of queue ength is added to the information exchanged by each node for reaizing the traffic fow status and estimation. The underying idea behind this type of routing is that each node shares the state of its neighborhood to every other node in the network. By exchanging oca information with a other nodes, each node wi have a ink states in its own database [16]. In other words, every node has access to the entire network topoogy incuding the weight information of each ink. Link state packets for a nodes in Figure A3 are shown in Figure A4. Figure A3. Simpe networks with weight information of each ink. Figure A4. Link state packets for a nodes in simpe networks. The underying idea behind this routing protoco is that ony a few changes in present ink-state routing protoco appy to our routing agorithm. Each node must do the foowing steps in Tabe A2. Whenever a route from source to any destination in the network is required, the source node computes the arc weight for each ink by the required traffic and pertinent information of each ink in the ink state database.