A Real-e Adapve Traffc Monorng Approach for Muleda Conen Delvery n Wreless Envronen * Boonl Adpa and DongSong Zhang Inforaon Syses Deparen Unversy of Maryland, Balore Couny Balore, MD, U.S.A. bdpa1@ubc.edu, zhangd@ubc.edu Absrac - Nework raffc onorng s crcal for wreless applcaons o ensure he QoS whle sreang uleda conen. The hgh rae of raffc onorng s lkely o oban he hgh accuracy of he nework raffc. However, usng a hgh onorng rae has a radeoff due o he exra bandwdh consupon. In a wreless nework envronen wh led bandwdh, raffc load generaed by a hgh onorng rae ay cause he nework congeson. Therefore, boh accuracy and bandwdh consupon should be aken no consderaon whle deernng an approprae raffc onorng rae. In hs paper, an adapve real-e raffc onorng approach ha can dynacally adjuss he onorng rae based on real-e nework bandwdh consupon and nework congeson saus s proposed. In he serous raffc condon, he syse uses a sall raffc onorng rae n order o avod ncreasng nework raffc load. If he nework raffc s seady, he rae s reduced. When he raffc s flucuaed, he rae s ncreased o deec varaons n he nework. Keywords: Nework raffc onorng, uleda syse, wreless nework, oble envronen, Traffc deecon, adapve syse. 1 Inroducon Wh he laes advances of wreless councaon echnology, a varey of applcaons for delverng uleda nforaon o handheld devces, such as oble phones and personal dgal asssans (PDAs), have eerged o ee user deands. For exaple, a busnessan who s vsng a cusoer ay vew a shor vdeo clp abou a new produc on hs PDA. A sockbroker keeps checkng he changes n he sock arke hrough hs cell phones whle ravelng. These scenaros deonsrae dfferen applcaons of uleda conen delvery n oble envronen. However, here are any challenges for developng such applcaons. For nsance, uleda applcaons requre suffcen nework bandwdh n order o ransfer connuous uleda sreas durng a perod of e. A sgnfcan varaon of bandwdh capacy n he nework durng ranssson can nfluence he qualy of uleda conen delvery and presenaon. In addon, he unrelably and nsably of wreless neworks cause varous probles such as fadng sgnal, raffc flucuaon and nework congeson, whch can consderably degrade he qualy of uleda ransssons. Therefore, s crcal o explore effecve engneerng echanss for ransng uleda conen va a wreless nework. Nework onorng s a process o collec varous nework nforaon ha can represen he nework condon, such as raffc load, node falure, packe error raes, loss raes, bandwdh, delay, and user saus [8]. If here are changes n he nework, applcaons can ulze hs nforaon o nae approprae acons o adap o he changes. For exaple, an applcaon reduces and expands bandwdh for ongong ransssons based on he changes n he nework. Snce ransng uleda conen s sensve o he varaon of nework raffc as descrbed above, nework onorng plays an poran role n ensurng he qualy of nforaon delvery and sooh dsplay of uleda conen on oble devces. In hs paper, he er raffc onorng refers o collecng raffc daa n a wreless nework. The (raffc) onorng rae s defned as he frequency of obanng raffc saples fro he nework. The an challenge n raffc onorng s o oban accurae raffc nforaon. If he colleced raffc nforaon canno accuraely reflec he curren nework suaon, an applcaon canno ake approprae adjusens o he conen delvery. Ths can cause any probles n nework ranssson and degrade qualy of servce (QoS). Users ay experence a long delay whle vewng he conen ransed fro he applcaon server on handheld devces. In addon, accurae raffc nforaon provdes a nuber of benefs. For exaple, oble applcaons can use o srea an approprae verson of uleda conen (e.g., vdeos wh hgh/low copresson raes) o users., I s also benefcal o any nework conrol asks such as adsson conrol, QoS anageen, faul deecon, and load balancng, whch ulze raffc nforaon o adjus he syse envronen [6,8]. * 0-7803-795-7/03/$17.00 003 IEEE.
Typcally, an applcaon usng a very hgh onorng rae s ore lkely o capure he os recen nework raffc daa han usng a low onorng rae, bu ncreases bandwdh consupon by saplng he nework ore frequenly for collecng raffc daa [,6].. In a wreless envronen wh led bandwdh, hese probles ay cause negave effecs. For exaple, under heavy raffc condons, a hgh onorng rae ay worsen he suaon and force soe connecons o be ernaed. Therefore, approprae raffc onorng raes should be dynacally deerned by facors ncludng bandwdh consupon and nework condon. In hs paper, we propose an adapve real-e raffc onorng approach ha dynacally adjuss he onorng rae based on wo facors: raffc conssency and nework condon. We defne he raffc conssency as he slary beween recen nework raffc daa, and nework condon as he congeson suaon (congesed or no congesed). In a prooype syse, he colleced raffc daa wll be sored n a nework raffc profle resdng n he applcaon server. Monorng raes n he lgh raffc condon wll be hgher han ones n he heavy raffc condon. In addon, he adjusen o onorng rae also depends on he varaon of nework raffc. If colleced raffc daa has relavely lle flucuaon for a perod of e, he rae wll be reduced. On he oher hand, f here are sgnfcan and frequen changes n he nework raffc, he onorng rae wll be ncreased n order o deec frequen changes n he nework. We beleve ha hs adapve approach can faclae he syse o oban he accurae raffc nforaon ore effecvely whle reducng he exra nework raffc load generaed by he onorng process. The reander of hs paper s organzed as follows. In secon, we brefly nroduce he concep of raffc onorng and exsng approaches. Then, we presen a real-e adapve raffc onorng approach n deals n secon 3. In secon 4, we descrbe he syse pleenaon. Fnally, we dscuss he fuure research and conclude he paper n he secon 5. Traffc onorng Traffc onorng s a process of collecng he nework nforaon n fve funconal areas,.e., faulanageen, accounnganageen, confguraonanageen, perforance anageenand secury anageen [8]. In hs paper, we prarly focus on nework perforance anageen. Accordng o [7], nework perforance anageen s a process of gaherng nforaon on he sae of nework (e.g. hroughpu and avalably of bandwdh capacy) n order o enable a syse o ake approprae acons o enhance he nework perforance. In he process of raffc onorng, a nework onor (usually a sofware coponen ravelng hrough he nework o gaher raffc daa such as packe rae and avalable bandwdh) s he an eleen for collecng raffc daa. Exensve sudes have been conduced on nework raffc onorng. In wred neworks, he exsng ehods are prarly operaed n a cenralzed anner based on SNMP and CMIP 1 [1,8]. A cenral anager node s responsble for anagng all ransssons n he nework. As a resul, all colleced raffc daa has o be ransed o a cenral daabase. Ths process consues a lo of bandwdh [1,8]. To allevae hs proble, Tha e al. [10] propose RM-based and IRM-based schees for raffc onorng. Boh schees collec raffc nforaon fro all nodes va nework onors. The onors keep rackng of raffc n he nework and perodcally send relaed nforaon o local daabases of ndvdual sub-neworks, so hese wo approaches work n a dsrbued fashon and consue less bandwdh han SNMP and CMIP. In wreless neworks, Manv [8] proposes anoher echans for raffc onorng by usng oble agens o collec raffc nforaon. The benef of usng oble agen ehod ncludes allevang he copuaon load a nework nodes. However, saplng raffc a a hgh frequency (hgh onorng rae) produces hgh bandwdh consupon. In addon, he huge volues of raffc nforaon cause a long processng e. All of he above exsng approaches share an dencal feaure ha hey all onor nework raffc usng a pre-defned sac onorng rae regardless of he flucuaon of nework raffc. Recenly, soe adapve nework raffc approaches have been nroduced. Fu e al. [6] propose an algorh ha collecs raffc nforaon adapvely by changng he range of onorng rae based on he bursness of raffc sae. I apples a e-seres odel, called Auo- Regressve Movng Average (ARMA), o specfy an approprae onorng rae based on sascal nferences fro he hsorcal raffc daa. Traffc daa s enered no he ARMA odel o deerne a suable onorng rae for a parcular range of raffc values. If he presen raffc daa reans seady coparng o he recen raffc daa, he onorng rae keeps unchanged. On he oher hand, n case he raffc s changed abruply, he syse recursvely ulzes he ARMA odel o generae a new rae of onorng. Slar o [6], Cho e al. [3] propose anoher echans for generang adapve onorng raes usng a dfferen aheacal odel (Auo- Regressve odel). Boh ehods odfy he onorng rae based on raffc daa and are proved ha hey can produce sasfacory resuls of raffc onorng process 1 Sple Nework Manageen Proocol (SNMP) and Coon Manageen Inforaon Proocol (CMIP) Relevan Monor (RM) and Iproved Relevan Monor (IRM)
n er of accuracy and nework cos. However, hey do no nclude he facor of nework saus whle odfyng he onorng rae. 3 An Adapve Real-e Traffc Monorng Approach In a wreless envronen, s desrable o save he bandwdh capacy as uch as possble, especally durng a heavy raffc suaon. We exend he approaches proposed by [3] and [6] by consderng wo ore varables,.e., raffc conssency and nework congeson saus. The raffc conssency s easured by he slary beween colleced raffc daa (bandwdh). If values of raffc daa are close enough, we consder he nework raffc s n a seady sae (conssen). As a resul, s reasonable o reduce he onorng rae n order o save nework bandwdh. When he nework s congesed, reducng onorng rae can avod ncreasng raffc o he nework. Collecng Traffc Info. Idenfyng Nework Saus If nework s congesed Deernng Monorng Rae Checkng Traffc Conssency Fgure 1. Adapve Real-e Traffc Monorng Process As shown n Fgure 1, afer collecng raffc daa wh a pre-defned onorng rae a he nal sep, he daa s forwarded o he second sep, called Idenfyng nework saus, n order o deerne he congeson saus of he nework. If he nework s serously congesed, he process drecly jups o he Deernng Monorng Rae sep. In hs sudy, a nework s defned n he congeson suaon when he percenage of s hroughpu ulzaon s hgher han a predefned hreshold value. However, f he nework s no congesed, he process connues o he Checkng raffc conssency sep o exane he seadness of he nework raffc. If he raffc values have been conssen whn a specfc e nerval, he nework s consdered n he seady sae. Oherwse, s consdered n changng sae. The process hen oves o he Deernng Monorng Rae sep. Accordng o he resuls of prevous seps, here are hree possble cases for assgnng a new onorng rae. Frsly, f he nework s congesed, he lowes onorng rae wll be assgned. Secondly, f he nework raffc s seady, he sep decreases he onorng rae. Thrdly, f he nework raffc s changng, he sep apples he cenral l heore of rando saples [,3] and Auo-Regressve (AR) odel [3,4] o calculae a new onorng rae based on he gahered raffc nforaon. The process repeas hese seps eravely, as shown n he Fgure 1. In he followng sub-secons, we descrbe every sep n hs adapve raffc onorng approach n deals. 3.1 Collecng Traffc Inforaon Ths sep capures he followng raffc nforaon [5]: raffc load he oal sze of packes flowng hrough he nework whn a e nerval, bandwdh (hroughpu) capacy of a nework easured by he nuber of byes of daa ransferred per second, and delay he duraon beween he e ha he source sends he daa and he e he desnaon receves. 3. Idenfyng Nework Saus Beng aware of nework suaon n wreless neworks s one of he exensons n our approach over he prevous adapve echanss [3,6]. In a congesed nework, s unnecessary o use a hgh onorng rae because worsens nework raffc. In addon, he process gh parly ncrease nework raffc load ha ay lead o eher ernae ongong ransssons or deny new callng requess due o he heavy raffc. Accordng o [9], here s no sandard creron for defnng nework congeson. A nuber of defnons have been nroduced. Fro a user s perspecve, congeson occurs n he nework when s/he experences unusual evens such as a longer delay or freezng of vdeo playng. Fro a echncal perspecve, he occurrence of congeson n he nework s denfed when he rans delay s greaer han a hreshold or he effecve hroughpu s less han a specfc value. We adap he laer defnon o deerne he nework congeson saus. In hs sudy, a nework s defned o be under congeson f he percenage of oal hroughpu fro all curren ransssons (copared o he axu hroughpu capacy of he nework) s larger han a specfc hreshold value (e.g. 80%). The hroughpu ulzaon of he nework can be defned as follows: oal_ hroughpu hroughpu _ ulzaon (%) = 100 (1) Max_ nework_ hroughpu If he nework s congesed, he process wll drecly ove o he Deernng Monorng Rae sep. If he nework s no congesed, he echans wll proceed o he Checkng Traffc Conssency sep as descrbed prevously.
3.3 Checkng Traffc Conssency The purpose of hs sep s o check he conssency of nework raffc, so ha he syse can decrease he onorng rae f he nework raffc reans seady for a perod of e. To deec he seadness or varaon of he nework raffc, we adop a sple sascal odel, naely varance ( σ ) []. Assung a syse keeps os recen hroughpu values, we can calculae varance a he e as follows. ( X µ σ () ) = = 1 where (-1, -,, -) denoes he nuber of hroughpu records, X (0 ) s he hroughpu value a h record, and µ s he ean of hroughpu values. In [7], he er σ +1 /σ s used o denfy he level of shor-er nework raffc varaon. Snce varance s a easure of how daa s spreaded ou, a sall nuber of varance ples a sall varaon fro he ean of a group of daa. Therefore, f he rao of wo varances σ +1 /σ s less han (or hgher han) 1, ndcaes ha he varance of new raffc hroughpu s reduced (or ncreased). As shown n he forulas (3) and (4), f we defne a hreshold value (TH σ ), we can generae he followng rules: f he rao of σ +1 /σ s whn he range of [1 - TH σ, 1+ TH σ ], he nework raffc s consdered seady. If he rao s ou of he above range, we consder here s a sgnfcan varaon of raffc n he nework. 1 - TH σ 1 σ + 1+ TH σ seady sae (3) σ 1 σ + < 1 - TH σ or σ + 1 > 1+ TH σ changng sae (4) σ σ 3.4 Deernng he Monorng Rae Ths sep as o deerne an approprae raffc onorng rae based on hree ypes of nforaon: nework congeson saus, raffc conssency and colleced nework raffc values. Fro hs nforaon, we can assgn onorng raes based on hree sauses of he nework raffc: congesed, seady, and changng. A new onorng rae s deerned based on he followng gudelnes. (1) f nework = congesed, hen new onorng rae = a predefned salles onorng rae () else f nework = seady sae, hen new onorng rae = reduced onorng rae (3) else f nework = changng sae, hen new onorng rae = newly calculaed onorng rae We descrbe he ehods o defne a new onorng rae n each parcular case n he followng sall secons. (1) Defnng he Salles onorng Rae If he nework s congesed, a pre-defned lowes onorng rae wll be used as he new onorng rae. () Reducng Prevous Monorng Rae If he nework raffc has been n a seady condon whn he e nerval, he curren onorng rae wll be reduced n order o save nework bandwdh and decrease processng e. We propose o apply he followng sple lnear equaon o gradually ncrease he lapse beween wo onorng probes: T = Told a, new new * T old T <= TH probe (5) where T new and T old (boh defned n second) are he new and prevous duraons beween wo adjacen onorng probes (See Fgure ). The value of he coeffcen a ( a > 1 ) can be changed a dfferen e based on how long he seady sae has lased. However, T new has o be less han or equal o a hreshold value, TH probe, because he onorng rae canno be oo sall. Monorng Probes T new Te nerval (sec.) Fgure. Reducng Monorng Rae by Increasng he Duraon Beween Two Adjacen Monorng Probes (3) Calculang a New Monorng Rae If he nework s n a changng sae, a new onorng rae s calculaed by usng cenral l heore and Auo-Regressve (AR) e seres odel descrbed as follows.
Accordng o [3], he requred nu nuber of raffc saples s calculaed usng he followng forula: 1 φ (1 η / ) σ n = ε µ = z where φ() s he cuulave dsrbuon funcon, µ and σ represens populaon eans and sandard devaon 1 of he packe sze respecvely, φ (1 η / ) and z p = ε σ S =. Values of η and ε (0< η <1, 0< ε <1) are µ anually creaed. The saller nubers of boh varables (η,ε) can lead o a ore accurae n. S s he squared coeffcen of varance (SCV) of he packe sze dsrbuon n a e nerval (e.g. 30, 60, 90 seconds). Assung ha packes are ransed hrough he nework durng a specfc e nerval and he sze of he h packe s X, we can derve he oal sze of packes (V): X = 1 V = However, he value of S has o be solved wh he acual nuber of packes durng he presen e nerval, so a e seres odel, Auo-Regressve (AR), s appled o predc he S paraeer based on he prevous packe nforaon. The predced SCV, S, and nuber of p packes,, are calculaed as follows [3]: S = = u = 1 u = 1 a S b where u refers o he eory sze. I s suggesed ha when u 5, (8) and (9) can yeld good predcng resuls [3]. Therefore, he necessary nuber of packe saples are calculaed as follows: S (6) (7) (8) (9) =. (10) n z p S Afer he nuber of saples s aaned, we can now calculae he onorng rae (R): R = n / Inerval (11) Moble Devces where n s he predced necessary nuber of saples, and nerval s he e duraon. 4 Syse Ipleenaon and Evaluaon Ths secon descrbes a wreless applcaon syse ha s currenly under developen. The syse s developed usng JME and negraes he proposed adapve nework onorng approach. A hs oen, he syse, as shown n he Fgure 3, consss of eulaed oble devces, an oracle daabase server o sore he raffc nforaon, and a web server. Voce, Tex, Iage, Web page, ec. Capure nework raffc nforaon Sore raffc nforaon n Daabase reques Web Server Applcaon Rereve profle DB Traffc Profle Fgure 3. The Archecure of he Wreless Applcaon Syse A nework-onorng ool called WnPcap [11] s ulzed n he syse o oban he raffc of he wreless nework. WnPcap can capure nework hroughpu and sze of packes. The capured daa s sored n a raffc profle daabase. The WnPcap keeps onorng raffc a wo levels: he frs level s he enre nework level, where he WnPcap- keeps deecng hroughpu currenly used by all ransssons n he nework. The second level s he ranssson level. Every acve ranssson n he nework s onored fro he begnnng o he end. The hroughpus a he second level are aouns of bandwdh occuped by ndvdual ransssons and her conen ransferrng saus (percenage of copleon). Currenly, we anly focus on he raffc deecon a he enre nework level. The syse sores a nuber of os recen raffc daa records for perforng he adapve raffc onorng echans and generang new onorng raes. An exaple of he raffc profle sored n he oracle daabase s shown n Table1.
Te_Inerval Toal_Used_bandwdh Throughpu_ (Mbps) Ulzaon (%) 1 5.5 45% 10 3.83 33% Table 1: Traffc Profle a Nework Level We plan o assess he effecveness of our approach by coparng he perforance and resource effcency [3] beween he sac onorng approach (usng consan onorng rae o deec nework raffc) and our adapve onorng approach. 5 Concluson The naure of wreless neworks and exra requreens of real-e uleda applcaons have brough nuerous neresng ssues o researchers and praconers. In order o effecvely delver uleda conen o oble users n he unsable wreless envronen, onorng nework raffc and akng dynac adjusen s crcal. In hs paper, we presen an adapve real-e raffc onorng approach ha ncludes he facors of raffc conssency and nework congeson suaon o dynacally change he onorng rae. If he raffc reans seady for a perod of e, he onorng rae should be reduced n order o decrease he raffc load and processng e. In addon, wheher he nework s congesed or no can also nfluence he onorng rae. The onorng rae should be reduced o a pre-defned lowes rae n order o avod ncreasng raffc whle he nework s congesed. Alhough we beleve ha he proposed approach can generae ore approprae onorng raes whle reducng chances o cause nework congeson, has soe poenal laons. For exaple, requres exra copuaon. However, we beleve ha benefs of hs approach ouweghs he exra copuaon cos because can help lessen he probably of nework congeson and prove he QoS of uleda conen delvery and presenaon. One of he neresng ssues n he fuure sudy s o deerne approprae values of hresholds n he proposed adapve approach hrough a seres of experens. I wll be also poran o evaluae and refne he approach hrough real-e applcaons. [] Donald A. Berry and Bernard W. Lndgren, Sascs heory and Mehods, nd ed., Duxbury Press, ITP, 1996. [3] B.-K. Cho, J. Park, Z.-L. Zhang, Adapve Rando Saplng for Load Change Deecon, Techncal Repor, Dep. of Copuer Scence & Engneerng, Unversy of Mnnesoa, 001. [4] Jonahan D. Cryer, Te Seres Analyss, Duxbury Press, Boson, MA, 1986. [5] Francos Fluckger, Undersandng Neworked Muleda, Prence Hall, Englewood Clff, NJ, 1995. [6] Z. Fu and N. Venkaasubraanan, Adapve Paraeer Collecon n Dynac Dsrbued Envronens, The 1 s Inernaonal Conference on Dsrbued Copung Syses, pp. 469-478, 001. [7] S. Hayes, Analyzng Nework Perforance Manageen, IEEE Councaon Magazne, vol. 31, ssue 5, pp. 5-59, May 1993. [8] S. S. Manv and P. Venkaara, A Mehod of Nework Monorng by Moble Agens, Proc. of he Inernaonal Conference Councaon, Conrol and Sgnal Processng n he nex llenu (CCSP- 000), pp. 1-5, July 000. [9] E. Monero, G. Quadros and F. Boavda, A Schee for he Quanfcaon of Congeson n Councaon Servces and Syses, Proc. of he 3 rd Inernaonal Workshop on Servces n Dsrbued and Neworked Envronen, pp. 5-60, June 1996. [10] C.K. Tha, Y. Jang and C.C. Ko, Monorng QoS dsrbuon n uleda neworks, Inernaonal Journal of Nework Manageen, vol.10, pp. 75-90, 000. [11] WnPcap: he Free Packe Capure Archecure for Wndows, developed by he Unversy of Calforna, Lawrence Berkley Laboraory [Onlne]. Avalable: hp://wnpcap.polo./, June 003. References [1] Larry L. Ball, Muleda Nework Inegraon & Manageen, McGraw-Hll Seres on Copuer Councaons, New York, pp. 8-83, 1996.