Utility-Baed Flow Control for Sequential Imagery over Wirele Networ Tomer Kihoni, Sara Callaway, and Mar Byer Abtract Wirele enor networ provide a unique et of characteritic that mae them uitable for building urveillance, including their mall ize and unobtruivene and their capabilitie for rapid deployment, decentralized monitoring and control. Bandwidth utilization contrain the ucce of a wirele urveillance ytem, epecially when the networ erve to tranmit video imagery. Creating an effective information management policy will help to obtain an allocation of bandwidth that i bet from the end uer perpective. Thi paper preent a decentralized pricing-baed flow control method to allocate bandwidth acro multiple competing camera that are imultaneouly attempting to mae ue of a hared, low bit-rate wirele channel. At preent, thi flow control model i being integrated into a building urveillance ytem in which off-the-helf camera detect intruder and then tart treaming data according to uerpecified preference for quality, where variou component of the flow control ytem monitor bandwidth utilization and adjut the quality etting of image tream to maimize enduer utility. I. INTRODUCTION Wirele enor networ have provided the ability to monitor, collect, and analyze large volume of data in a variety of environment and application. However, wirele networ are ubject to limited reource, uch a bandwidth. Overloading the networ with information can caue networ error. A networ uage increae, a need arie to develop an efficient information management policy to allocate bandwidth to competing application. Information management i the deciion maing proce about the ue of networ reource. Succeful creation and implementation of an information management policy will provide benefit to all wirele enor networ application by allowing the enor to intelligently meet the need of it uer. Flow control, or the adjutment of tranmiion rate for individual traffic ource in a networ, form a core foundation for an efficient information management policy. Manucript received April 9, 007. Thi wor wa upported by the Sytem Technology Integration Lab of the Department of Sytem and Information Engineering at the Univerity of Virginia, Charlotteville, VA. T. Kihoni i with the Sytem and Information Engineering Department, Univerity of Virginia, Charlotteville, VA 90 USA (email: t@alumni.virginia.edu). S. Callaway i with the Sytem and Information Engineering Department, Univerity of Virginia, Charlotteville, VA 90 USA (email: ec9b@alumni.virginia.edu). M. Byer i with the Sytem and Information Engineering Department, Univerity of Virginia, Charlotteville, VA 90 USA (email: mab6nu@alumni.virginia.edu). Inefficient flow control can lead to ource greedily uing bandwidth, overloading the ytem and cauing networ congetion. A congeted networ delay or loe information ent to the uer, which decreae the overall effectivene of the networ. Converely, an ideal flow control algorithm would allocate bandwidth o a to give higher priority (more bandwidth) to ource that the uer deem important. Thi paper develop a pricing-baed flow control method for a wirele urveillance ytem where variou area of a building vary in ignificance to the uer. The algorithm preented in thi paper adapted to a pecific building urveillance cenario. In thi ytem, a uer require image tream to mae accurate deciion regarding the action to tae againt intruder. Additionally, multiple tream limit the cognitive ability of the uer, which increae the difficulty the uer incur in maing an effective deciion. Therefore, the uer require accurate information that ha the neceary level of fidelity to mae deciion. II. LITERATURE REVIEW Price-Baed Flow Control (Economic Model) The method of flow control ued in thi paper revolve around a pricing model that i imilar to recent wor in the networing literature on optimal flow control. For eample, in [1], Kelly et al. eamine the tability and fairne of two interacting rate control algorithm, decompoing the problem into contituent primal and dual part olved independently by the networ and all uer, repectively. The networ algorithm erve to et a price for bandwidth utilization baed on networ congetion, while uer adjut tranmit rate o a to maimize net utility (utility for bit rate offet by cot). The bai for fairne derive from an earlier paper [], alo due to Kelly, which provide two common eample of fairne: ma-min fair and proportionally fair. The ma-min fairne criterion require that maller flow receive their requeted flow rate before larger flow and aert that increaing the rate of flow for larger ource at the epene of the maller flow i unfair. To conider flow rate proportionally fair, the rate mut be feaible and for any other feaible rate, the aggregate of proportional change i zero or negative []. In [3], Low and Lapley epand on [1]-[] by creating ynchronou and aynchronou verion of both algorithm. A ynchronou algorithm arie when update to flow ource and networ lin occur at the ame time interval. Their eperimental reult how that their prototype behaved a epected and that the algorithm trac the theoretical bandwidth allocation optimum [3].
Kelly and Low provide a olid framewor for flow control uing a pricing cheme. Their thorough etup of a utility model provide the motivation to following a imilar pricing cheme. However, the pecific application dicued in thi paper concern that of tranmitting treaming image to the uer in a deciion critical cenario. Thi directly tranlate to a different et of objective a compared to Kelly and Low. Wherea in their wor they ought to efficiently allocate tranmiion rate while minimizing the ue of bandwidth, thi model require maimizing bandwidth in order to provide a much information a poible to the deciionmaing uer. Thi preent an intereting problem that require further dicuion and analyi a done in the remainder of thi paper. III. GENERAL UTILITY MODEL In thi ection we dicu the algorithm ued to allocate bandwidth to camera in the networ. Thi model preent a pricing model imilar to that of Kelly and Low in [1]-[3], with two new feature: multiple uer and dynamic utility. A. Utility Model Thi ection introduce the optimization problem and it tructure. A in [1]-[3], we can conider a networ that contain a unidirectional lin with a capacity c. Connected to the networ are K = (1,...,K) uer. The networ include a et S = {1,...,S} of camera that have the following characteritic: A ubet of uer L = ( K ) that receive tream from camera. The utility function U : R + R, which we aume to be trictly concave and increaing. Specifically, U (, t ) = w ( t ) u ( ) (1) w ( t ) = r d ( t ) () r = a L i L i : tranmiion rate of camera u ( ) : trictly concave and increaing utility function baed on the tranmiion rate r : aggregate tatic priority of a camera tream weighted by uer L a : uer tatic priority i : uer weighting level ( R ) + d(t): dynamic priority trictly decreaing in time Minimum and maimum tranmiion rate, m 0 and M <, repectively. The camera randomly activate for a random equence of time. Additionally, R = ( S ) repreent the ubet of camera that are active. Thi lead to the objective of the (3) problem, chooing tranmiion rate ( S ) to =, o a ma U (, t ) ().t. c (5) R = According to [3], a unique maimizer of aggregate utility eit becaue the objective function i trictly concave and the feaible et of olution i compact. Solving the optimization problem in a centralized fahion require coordination among all active camera and i impractical in real networ, leading to our propoal of a decentralized pricing-baed flow control algorithm. B. Pricing Algorithm for Flow Control In thi ubection, we preent two cae for flow control price. The firt cae introduce a ingle uer cenario that eclude the dynamic priority and build off the recurive model found in [1]-[3]. The econd cae provide a general model where quic repone are required due to the incluion of the dynamic priority. 1) Single Uer with Static Priority and Logarithmic Utility Function In thi cae, a ingle uer provide a tatic priority level to the camera. Additionally, all thee camera ehibit a logarithmic curve, which reduce (1) to U ( ) = a log( ) (6) Thi mimic the utility function by Kelly in [] and a imilar recurive method (7) olve the optimization problem. p ( t ) = p( t ) + γ c + 1 (7) R The camera allocate bandwidth baed on a greedy approach by maimizing their benefit (or utility minu cot). (We will dicu the allocation cheme more in depth in the general cae.) a () t = (8) p () t In fact, ince the optimization problem require full networ utilization at c, thi pecialized cae implifie to = a a R Additionally, the allocation in (9) provide proportional fairne and derive all the attribute decribed by Kelly in []. (9)
) General Model of Pricing The idea from [1]-[3] and the pecialized cae do not depend on a dynamically changing variable. In the cae of a dynamically changing utility function, the ytem require quic repone, omething not guaranteed in a recurive model. The general cae we preent here provide a method to allocate bandwidth in a dynamic environment. In order for the networ operator to upply bandwidth properly, the lin mut monitor the aggregate utilization of bandwidth and calculate a price per unit of bandwidth conumed. We attribute to the lin a diutility aociated with bandwidth conumption, where the form of the diutility function reflect the effect of congetion at full networ utilization. Similar to a traffic jam where an additional car bring flow to a halt, an additional bit at full networ utilization may caue lo of information. In thi paper, we aume that, an eponential price curve accurately reflect the networ operator increaing diutility aociated congetion. Specifically, the lin determine the price uing a looup method (10), a oppoed to the numerical method in (7). P = b = b ( b) e (10) R Therefore, the two tep neceary for the lin to upply bandwidth involve (i) monitoring the networ utilization and (ii) calculating the price baed on the diutility at that utilization. Acting a a conumer, each active camera ee to conume the networ bandwidth, and it doe o on behalf of the end uer aociated with the camera. Uing it default behavior, each camera will attempt to maimize it private net benefit (1) according to the utility function et by it aociated uer. The point of maimization indicate the optimal bit rate for the camera. Each camera ue a imilar logarithmic utility function differentiated by priority etting a een in (1). The priority etting allow the uer to give bandwidth preference to certain camera. C ( ) = p (11) B ( ) = U (, t ) C ( ) (1) = arg ma( B ( )) (13) C: cot B: benefit p: price Change in bandwidth conumption propagate bac to the lin (networ operator) which then calculate a new price reflective of the new level of utilization. Thi upplierconumer relationhip iterate and produce a damped ocillation around an equilibrium price. Since the pacet ent by the camera contain time enitive information, a method to damp further the ocillation may be needed, a in the following price averaging cheme: p( t 1) + pc ( t) p( t) = (1) p c : current lin calculated price A drawbac of the price-etting formula of (10) i that the price and end uer utilitie are micalibrated, reulting in equilibrium bandwidth allocation that fail to conume all available bandwidth. To fully maimize aggregate utility (a in ()), an additional tep i required to enure that networ reource are fully utilized. The mechanim we adopt in thi paper i to end a proportional multiplier (a computed by the networ operator according to (15) below) to all camera when the bandwidth price have reached relative equilibrium (price vary by le than 5%). c m = (15) b When the equilibrium bandwidth allocation would otherwie ettle on a networ operating at le than full capacity, the normalization factor erve to increae the bandwidth allocation for each camera, while maintaining the proportional bandwidth allocation between the camera. Thi model decribe the general method to reach equilibrium in allocation. However, it remain to be een whether the allocation olve the optimization problem in (5). We plan to further invetigate thi model to tet for optimality, which will provide the focu of future paper. C. Simulation Proof of Concept In thi ubection, we preent imulation reult that provide a proof of concept for our ditributed pricing algorithm above. Rather than preenting a complete numerical evaluation of our approach, we illutrate the effect of the pricing cheme in a imple cenario. (We plan to document a more complete numerical evaluation of the cheme in a future publication.) The imulation reult below were computed in Matlab uing a ynchronou implementation of the price/bandwidth update of (10), (13), (1), and (15). To et the tage for the numerical reult, conider ingle wirele lin with 11 Mbp total capacity, of which 8 Mbp are to be allocated (o that the target bandwidth i 8 Mbp). We aume that there are five camera, which are either active or aleep in each of 10 round, where each round i 0 unit of time in length. Each price/bandwidth update conume one unit of time, o that there are up to 0 price update per round. We aume that all camera have the ame utility function, with different priority etting (a = 1,, or 3) and a dynamic priority decreaing each ubequent iteration. Fig 1 and how the effect of the imulation with and without the price averaging cheme.
Price ($) 3.5 3.5 1.5 Pricing Simulation with Price Averaging Total Bit Rate (Right Ai) Price (Left Ai) 1 0 50 100 150 00 0 Time Fig. 1: Pricing Simulation with Price Averaging. Active camera indeed by round = (1,, 3, 5,, 3,, 1, 0, 1) Price ($) 3.5 3.5 Pricing Simulation without Price Averaging Total Bit Rate (Right Ai) Price (Left Ai) 1 10 8 6 1 10 8 6 Total Bit Rate (Mb/) Total Bit Rate (Mb/) and a imulated environment to imulate the affect of monitoring multiple building [6]. The information management policy, implemented on the node of the networ (erver and camera), i ued to control information flow by adjuting the quality control of the image tream. Thi ection erve to decribe both the hardware and oftware apect of our pricing-baed flow control algorithm within the overall ytem. A. Implementation Architecture and Strategy 1) Hardware A laptop with the Linu laptop erve a a central erver. Thi erver provide all intance of the web-baed interface acce to available image tream o that they can be een by end uer. (While the web interface can be made available to any number of uer, it actually encode only one uer et of preference for video from individual camera.) The ame laptop alo erve to implement the lin functionality of our flow control algorithm. Specifically, the laptop monitor the bandwidth utilization in the networ and compute the prevailing price-per-unit-bandwidth according to (10). Camera (Logitech Quiccam Pro 000 [7]) are deployed throughout the building for urveillance and are attached to Stargate computer that run the change detection and image treaming application of []. (Stargate are mall wirele device which communicate uing 80.11 a their wirele protocol [8].) The Stargate alo implement the bandwidth update functionality of our flow control algorithm, receiving price periodically from the laptop and adjuting image treaming parameter. 1.5 1 0 50 100 150 00 0 Time Fig. : Pricing Simulation without Price Averaging. Active camera indeed by round = (1,, 3, 5,, 3,, 1, 0, 1) A een in Fig. 1 and, the effect of the price averaging i ignificant. Thi imulated reult how that uer will have a relatively contant tream of fidelity uing thi cheme. IV. IMPLEMENTATION OF THE FLOW CONTROL ALGORITHM The reearch of thi paper wa conducted in the contet of a larger project on the development of ytem that ue wirele camera to provide facility video urveillance. An image-proceing application from [] provided the method to tream a equence of encoded image from camera. The encoding cheme ue a change detection algorithm to tream only the change relative to a bacground image, which then recontruct itelf on the interface uing the ame bacground image. Other component of the larger project are the development of a web-baed interface for etting up the urveillance networ and viewing motion detection [5], Fig. 3. Sytem Architecture ) Software A hown in Fig. 3, all communication will go through the central erver, which reide on the Linu laptop dicued previouly. When camera detect motion, they end the reulting image to the erver creating input traffic on the erver. When uer connect to the erver to view image tream, output traffic i created. The laptop alo upport the lin functionality of our pricing-baed flow control algorithm, and central to thi meaurement capability for aeing bandwidth utilization. At preent we are uing a Linu utility nown a iptable [9] to monitor the bandwidth utilization in the networ. Iptable i generally ued to create firewall on a machine [9]. In thi cae, it i ued to create a firewall that doe not actually bloc any networ traffic. Intead, rule about pacet being ent through the networ card are ued to monitor the networ uage [9]. Thee rule count the traffic in three categorie: input, output, and forward. Thi traffic
i monitored for a period of time. Thi period i an input that fit the need of the networ being monitored. A more volatile networ will require a longer interval to mooth out the obervation. The total monitored bandwidth i divided by the number of econd in the time interval to determine the average bit/ during the period. After each meaurement, the counter are reet to zero and the period retarted. The lin functionality implemented on the laptop ue thi meaured bandwidth utilization to compute price (per unit of bandwidth), which are then ent to the flow control algorithm running on the Stargate computer (to which the camera are attached). The bandwidth updating functionality of the flow control algorithm reide on the Stargate, where lin price and utility function are reconciled and new video encoding parameter are computed. When encoding parameter change, they are written to a configuration file for the encoding oftware, which alo reide on the Stargate, and the ytem end a meage to the encoder telling it to read the new configuration file. B. Component Detail Our pricing-baed flow control algorithm required eparate routine for the lin and camera, implemented in Java uing tandard librarie. Java portability and objectoriented tructure allow for traightforward algorithm implementation. 1) Lin Routine We refer to the price-etting mechanim that reide on the erver a the lin routine or imply the lin. Implementing the lin required developing a multithreaded program. Fig. preent the functional relationhip between the thread and the variou Java method we developed. ) Camera Routine We refer to the bandwidth updating mechanim of our flow control algorithm, which reide on the Stargate, a the camera routine or imply the camera. Similar to the lin routine, the camera routine implement a multithreaded application a een in Fig. 5. Fig. 5: Camera Routine a) Video Encoding Routine The output of the camera routing i a new et of encode parameter for the encoder developed in [], with the goal being to et encoding parameter to achieve the optimal bandwidth allocation computed by our flow control algorithm. The relevant encoding parameter include (i) frame per econd and (ii) change detection threhold. The frame per econd parameter greatly influence a camera bandwidth and erve a a coare bandwidth adjutment mechanim. The change detection threhold, which decribe the enitivity the program to change in the image, allow for more precie control of the bandwidth utilization. The combination of thee two parameter offer u a method to match the calculated tranmiion rate. b) Camera/Lin Initialization To initialize the ytem, each camera mut notify the lin that it ha entered it networ. Thi i the very firt tep done by every camera. Upon regitration of each camera, the lin aign and tranmit a multicat addre to the camera. After completing thee initial tep, each intance of the camera routine proceed a required by the flow control algorithm: litening for price and adjuting it imageproceing etting. Fig. : Lin Routine The function mared in the grey bo repreent concurrent thread. To tranmit the price from the lin the camera, a multicat UDP pacet i ent. Additionally, the lin liten for the uer et priority and relay the level to all camera in it networ. Although not dicued in thi paper, the lin routine i actually implemented in a way that allow it to be ued in the cae of a multilin environment. C. Statu At the time of thi writing, the eparate component of the larger facility video urveillance project are being combined into a ingle woring ytem. A demontration of the final ytem i cheduled to occur on April 16, 007 at the Nuclear Reactor Building at the Univerity of Virginia. V. CONCLUSIONS The model preented in thi paper provide a method to efficiently allocate bandwidth to camera with varying level
of importance to the uer. The pricing model ue a multiround proce to allocate bandwidth to the camera. In each round a lin monitor current bandwidth utilization and calculate a new price-per-unit-bandwidth. The camera receive thi price and elect the bandwidth that maimize it own benefit (net utility). Achieving thi new bandwidth allocation require adjutment to two video encoding parameter, frame-per-econd and change detection threhold, a pecified by the video encoding format of []. A encoding parameter are adjuted, the lin routine continue to monitor the bandwidth utilization by mean of the Linu utility, iptable. Thi allow the lin to calculate a new price for the net round, and the proce retart. The algorithm produce, in a decentralized fahion, an equilibrium bandwidth allocation for the camera tream maimizing aggregate utility. Invetigating the optimality of thi equilibrium provide the focu of future reearch. An effective information management policy amplifie the utility gained by ytem uer. The technique of thi paper alo apply to other application, uch a home networ bandwidth management. The importance of creating an effective utility-maimizing flow control algorithm increae with the addition of more and more Internetconnected conumer product. For eample, uer may download movie trailer to their televiion, play online video game, or watch treaming media. An effective information management policy will efficiently allocate bandwidth to all the uer according to prioritie (utilitie) et by the networ operator. REFERENCES [1] F. Kelly, A. Maulloo, & D. Tan, Rate control for communication networ: hadow price, proportional fairne and tability. Journal of the Operational Reearch Society, 9(5), 37-5, 1998. [] F. Kelly, Charging and rate control for elatic traffic. European Tranaction on Telecommunication, 8(1), 33-37 1997. [3] S. Low, & D. Lapely, Optimization flow control. I. Baic algorithm and convergence. Networing, IEEE/ACM Tranaction on, 7(1), 861-87, 1999. [] M. Farrell, Peronal Communication, 007. [5] F. Thompon, A. Vawani, B. Walthall, M. DeVore, Computer viion information ytem for wirele urveillance, 007. [6] D. Chao, K. Datt, J. Rehi, G. Learmonth, Human-in-the-loop imulation tetbed for wirele enor networing for infratructure urveillance, 007. [7] Logitech Homepage. www.logitech.com [8] Crobow: Wirele Senor Technology: Homepage. www.bow.com [9] G. Beeman, Bandwidth monitoring with iptable. Linu.com. http://www.linu.com/article.pl?id=05/1/15/1773