Eonomy-based Content Replaton for Peerng Content Delvery Networs Al-Muaddm Khan Pathan 1 and Raumar Buyya Grd Computng and Dstrbuted Systems (GRIDS) Laboratory Department of Computer Sene and Software Engneerng The Unversty of Melbourne, Parvlle, VIC 3010, Australa {apathan,ra}@sse.unmelb.edu.au Abstrat Exstng Content Delvery Networs (CDNs) by nature are losed delvery networs whh do not ooperate wth other CDNs and n prate, slands of CDNs are formed. The urrent logal separaton between ontents and serves n ths ontext results n two ontent networng domans. In addton to that, meetng the Qualty of Serve requrements of lents aordng to negotated Serve Level Agreements s rual for a CDN. Present trends n ontent networs and ontent networng apabltes gve rse to the nterest n nteronnetng these networs. Hene, n ths paper, we present an open, salable, and Serve- Orented Arhteture (SOA)-based system that assst the reaton of open Content and Serve Delvery Networs (CSDNs), whh sale and support sharng of resoures through peerng wth other CSDNs. To enourage resoure sharng and peerng arrangements between dfferent CDN provders at global level, we propose usng maret-based models by ntrodung an eonomy-based strategy for ontent replaton. 1. Introduton Content Delvery Networs (CDNs) [1], whh evolved frst n 1998, replate ontent over several mrrored Web servers, strategally plaed at varous loatons around the globe to deal wth flash rowds and to enhane response tme. In a typal ontent delvery envronment Web server lusters are loated at the edge of the networ to whh the lents are onneted. A ontent provder an sgn up wth a CDN provder for serve and have ts ontent plaed on the ontent servers. The ontent s replated ether ondemand when lents request for t, or t an be replated beforehand, by pushng the ontent to the surrogate servers. A lent s served wth the ontent 1 The author s n the frst year of hs anddature as postgraduate researh student and Assoate Professor Buyya s the supervsor. from the nearby replated Web server. Thus the lents end up unnowngly ommunatng wth a replated CDN server lose to them and retreves fles from that server. Fgure 1 depts the dfferent ontent/serves served by the CDN to the lents. Fgure 1: Content/serves provded by a CDN 1.1. Motvaton and sope Exstng CDNs are propretary n nature ndvdual ompanes own and operate them. Eah of them omprses of own losed delvery networ, whh s expensve to set up and mantan. They also have lmted salablty. Runnng a global CDN s even more ostly, requrng an enormous amount of aptal and labor. In addton, ontent provders typally subsrbe to one CDN provder and thus an not use the resoures of multple CDNs at the same tme. Suh a losed and non-ooperatve model results n slands of CDNs. Moreover, the logal separaton between ontent and serves (e.g. applaton for vdeo proessng) under the ontent dstrbuton and ontent serves domans s undesrable onsderng the ongong trend n ontent networng. A unfed networ that supports oordnated omposton and delvery of ontent and serves would be muh better. Furthermore, ommeral CDNs mae spef ommtments to ther ustomers by sgnng a Serve Level Agreement (SLA). It desrbes provder s ommtment and spefes penaltes f those
ommtments are not met. So, f a partular provder s unable to provde qualty of serve to the lent requests, t may result n SLA volaton and end up ostng the provder. To ut expenses and to avod adverse busness mpat, peerng wth other provders ould be a soluton to onsder the ssues stated above. In ths paper, we present a model for an open, salable and SOA-based system. Ths system helps to reate open Content and Serve Delvery Networs (CSDNs) [3] that sale well and share resoures wth other CSDNs through peerng, thus evolvng past the urrent landsape where slands of CDNs exst. To enourage peerng among CDNs at global level, we propose usng maret-based models n resoure alloaton and management nspred from ther suessful utlzaton n the management of autonomous resoures, espeally n global grds [4]. Hene, we ntrodue an eonomy-based replaton strategy that nvolves on-demand plaement of outsoured ontent to the surrogates of peerng CDNs. The use of eonom mehansms n ths ontext has the followng benefts: We propose a Vrtual Organzaton (VO) model to peer the CDNs. A VO onssts of real CDN provders, whh are self-nterested and autonomous stae-holders. Hene, an eonom model s sutable to represent ths senaro and to regulate nteratons among the partpants. The problem an be manageable n ths way through analyzng emergent maretplae behavor. A peerng envronment of ontent networs s hghly dynam n nature, where the avalablty of resoures hanges over tme. Thus, an eonom model s approprate to explot the dynamsm of the maret to mae more nformed desons on the fly. An eonom model ould be the bass of a selfregulatng replaton strategy that dynamally adapts to the hanges n the lent request patterns. The rest of the paper s strutured as follows: Seton 2 establshes the sgnfane and relevane of our proposal; Seton 3 addresses the shortomngs of related wor; Seton 4 presents the proposed model; Seton 5 enlghtens the eonomy-based model for ontent replaton; and Seton 6 provdes a summary wth expeted ontrbutons and future dretons. 2. Sgnfane of peerng among CDNs In our approah, a VO s formed through the oordnaton of Web server lusters operated by dfferent CDNs who have ome together to share resoures and to ollaborate on ommon goal(s). A VO n the peerng CDN envronment may vary n terms of purpose, sope, sze, and duraton. VOs n suh an envronment are of two types: short-term on-demand VOs and long-term VOs wth establshed SLAs. A short-term VO s formed for short tme duraton, based on urrent lent request pattern to prevent the generaton of hotspots. Consder the followng senaro as a motvaton for short-term VO formaton. Suppose that the ontent of www.nn.om s hosted by the CDN provder Aama [5]. Aama s Web servers reeve sgnfant lent requests to serve the latest ontent on behalf of www.nn.om. A sudden news outburst, demandng to the lents (e.g. 9/11 ndent n USA), may ause heavy worload on Aama s Pont-of- Presenes (POPs) n a partular regon. As a result a hotspot an be generated. It ould ause Aama s POPs n that regon to be unable to ope wth the stran. Eventually the Web servers wll be totally overwhelmed wth the sudden nrease n traff, and CNN s Web ste wll be temporarly unavalable. Suh sudden spes n Web ontent requests s termed as flash rowd [2] or SlashDot effet [6]. In the peerng CDN envronment, the generaton of hotspots due to flash rowd an be resolved through the formaton of short-term VOs. A short-term VO ntervenes wth sudden spe n requests for partular Web ontent(s), whh results n heavy worload on ertan Web server(s) of a partular CDN. Hene, Web servers of peerng CDNs form a goal-orented onstellaton of dstrbuted sem-autonomous enttes and exess load s dstrbuted to the less loaded Web servers of other CDNs. Suh peerng arrangement should be automated wthn a short tme frame to address the evolvng stuaton. A short-term VO s formed on-demand and the poly for suh VO formaton s establshed dynamally to handle the evolved stuaton. The short-term VO s phased out when the worload returns to normal. On the other hand, a long-term VO s formed for events whh may be nown n advane. A long-term VO remans for the duraton of the event. The formaton of long-term VO omplments the exstene of establshed poles and negotated SLAs among the partpatng enttes. To better understand the formaton of long-term VO, onsder the followng senaro. Suppose that the ICC Cret World Cup 2007 s beng held n the Carbbean, and www.rnfo.om s supposed to provde lve meda overage. As a ontent provder, www.rnfo.om has an exlusve SLA wth the CDN provder, Aama [5]. However, Aama doesn t have a POP n Trndad and Tobago (a Carbbean sland), where most of the ret mathes wll be held. As beng the host of most of the ret mathes, people of ths partular part of Carbbean are expeted to have enormous nterest n the lve overage provded by www.rnfo.om. Sne Aama s expeted to be aware of suh event
well n advane, ts management an tae neessary ntatves to deal wth the evolvng stuaton. In order to provde better serve to the lents, Aama management mght dede to plae ts surrogates n Trndad and Tobago, or they mght use ther other dstant edge servers. Frstly, plang new surrogates ust for one partular event would be ostly and mght not be useful after the event. On the other hand, Aama rss ts reputaton f t an t provde agreed qualty of serve for lent requests, whh ould volate the SLA and stll ause proft reduton. Hene, the soluton for Aama ould nvolve peerng wth other CDN provder(s) to form a VO n order to delver the serve that t ould not provde otherwse. Automaton for long-term VO formaton s not essental sne suh stuaton s nown before-hand. Thus, by ollaboratng wth other CDN provders though the formaton of VO, ontent networs an better satsfy the evolvng needs of ther ustomers and meet ther QoS requrements. 2.1. Researh ssues The ey hallenges that need to be addressed n varous aspets of peerng CDNs are: Load Dstrbuton: How to ensure redued server load, less bandwdth onsumpton (by partular CDN server) and mprove the performane of ontent delvery? Coordnaton of CDNs: What nd of oordnaton mehansms need to be n plae whh ensure effetveness, and allow salablty and growth of ooperatve CDNs? Serve and poly management: How to mae a value-added serve nto an nfrastruture serve that s aessble to the ustomers? What types of Serve Level Agreements (SLA) are to be negotated among CDN partpants? What poles an be generated to support SLA negotaton? How an autonomous poly negotaton happen n tme to form a tme-rtal short-term VO? Prng of ontents and serves: How do CDN provders aheve maxmum proft n a ompettve envronment, yet mantan the equlbrum of supply and demand? 3. Related wor In ths seton, we outlne the efforts for nternetworng of Content Delvery Networs: The IETF RFC doument [7] proposes a Content Dstrbuton Internetworng (CDI) Model that allows the CDNs to have a means of afflatng ther delvery and dstrbuton nfrastruture wth other CDNs who have ontent to dstrbute. An arhteture for Content Dstrbuton Internetworng (CDI) s presented n [8]. It dsusses the desgn, mplementaton and evaluaton of only a protools arhteture for ooperaton among separately admnstered CDNs. A peerng system for ontent delvery worloads n a federated, mult-provder nfrastruture has been presented n [12], but the peerng strategy, resoure provsonng and performane guarantees among partnerng CDNs s unexplored n ths wor. CDN broerng [9] allows one CDN to ntellgently redret lents dynamally to other CDNs n that doman. The drawba s that, mehansm for IDNS s propretary n nature and mght not be sutable for a gener CDI arhteture. Also, dependng solely on DNS-based routng s ll-advsed due to oarse ontrol over requests as a result of ISP and lent ahng of DNS nformaton. Most of the wors mentoned above do not vrtualze multple provders for ooperatve management and delvery of ontent n a peerng envronment. Hene, our ontrbuton les n desgnng an effetve peerng mehansm that endeavors to address the lmtatons of prevous related wor, whle respetng lent performane requrements through proper poly management for negotated SLAs. 4. The model for peerng CDNs VO 1 SMP text Web UserP Poly Repostory1 Web User1 Serve Regstry1 Web Serves Host (e.g., Apahe) SLA negotaton Serve SLA-based Alloator Cluster Web Server 1 ` Coordnated VOR Sheduler Internet Enterprse System Coordnated VO1 Sheduler VOR SMP SMP Web Serves Host (e.g., Apahe) SLA negotaton Serve SLA-based Alloator Cluster Web Server N Serve RegstryR Poly RepostoryR Web Serves Host (e.g., Apahe) SLA negotaton Serve SLA-based Alloator Cluster Web Server 2 Enterprse System Enterprse System Fgure 2: Arhteture of open, salable, SOA-based system to assst the reaton of ooperatve and oordnated Content and Serve Delvery Networs Our proposed system endeavors to solve the problem of non-ooperatve slands of CDNs. It
antpates ensurng qualty of serves based on SLA negotaton, and addresses the problem of logal separaton n ontent networng doman. Arhteture for suh a system s shown n Fgure 2. In the fgure, eah Web server represents a CSDN. In the proposed VO-based model for formng Content and Serve Delvery Networs (CSDNs), the formaton of a VO may be stand alone or may be omposed of a herarhy of regonal, natonal and nternatonal VOs. A VO onssts of Web servers from multple CSDNs, a oordnated VO sheduler, a serve regstry, and a poly repostory. Web servers wthn eah CSDN are apable of delverng serves n order to meet the QoS requrements of the lents. Bref desrpton of VO omponents an be found n [3]. 4.1. Poly management to support SLAs A poly n the ontext of peerng CSDNs would be statements that are agreed upon by the partpants wthn a VO. These statements defne what type of ontents and serves an be moved out to a CSDN node, what resoures an be shared between the VO partpants, what measures are to be taen to ensure Qualty of Serve (QoS) based on negotated SLA, and what type of programs/data must be exeuted at the orgn servers. Wthn our proposed VO-model based CSDN arhteture we apply the standard poly framewor defned by the IETF/DMTF [10]. In the standard poly framewor, the admn doman refers to an entty whh admnsters, manages and ontrols aess to resoures wthn the system boundary. An admnstrator uses the poly management tools to defne the poles to be enfored n the system. The poly enforement ponts (PEPs) are logal enttes wthn the system boundary, whh are responsble for tang aton to enfore the defned poles. The poly repostory stores poles generated by the admnstrators usng the poly management tools. The poly deson pont s responsble for retrevng poles from the poly repostory, for nterpretng them, and for dedng on the set of poles to be enfored by the PEPs. The proposed model for CSDN n Fgure 2 an be mapped to the bas poly framewor. The poly repostory n Fgure 2 s responsble for storng poles generated by the poly management tool used by the VO admnstrator. The dstrbuton networ and the Web server omponents (.e. SLA negotaton serve, SLA based Alloator) are nstanes of the poly enforement ponts (PEPs), whh enfore the CSDN poles stored n the repostory. Eah VO sheduler s an nstane of the poly deson pont (PDP), and determnes the set of poles to be enfored at the tme of peerng among CSDNs. The poly management tool s admnstrator dependent and t s not shown n Fgure 2. A dret beneft of usng suh poly-based ooperatve arhteture s to redue the operatng ost of CSDNs and to meet lents QoS requrements aordng to negotated SLA. 5. The eonom model Standardzed eonom onepts an be deployed for ontent replaton wthn the struture of the VO model for CSDN. The am s to share the networ openng among CDNs to mnmze the networ lateny pereved by lents and to restrt the generaton of hotspots due to the exessve lent requests for some ontent. But suh peerng may result n free-rdng where some CSDN provders use other s resoures free of harge. Suh free-rdng an be avoded through usng an auton model for replatng ontent among CSDN servers, ensurng that partpaton n a VO s due to proft motvaton. The auton model should be able to provde nentves to all partes. In our eonomy-based model, the goal of the auton protool s to selet the heapest sutable Web server n order to replate ontent there. Here we apply the buyer-drven auton mehansm, whh s a type of Vrey auton [11]. Vrey autons are seondpre sealed-bd autons wth low messagng overhead, effeny of alloatons and la of ounter speulaton. They nvolve a sngle negotaton round n whh eah bdder submts a bd to the autoneer. Other bdders annot see the bd. The bddng agent whh maes the hghest bd wns the auton but pays the pre of the seond-hghest bd. In our ase, eah CSDN provder s both a buyer and seller of ts resoures. A oordnated VO sheduler s the autoneer n the peerng CSDN envronment that s responsble for holdng auton wthn the VO. It starts an auton on behalf of a CSDN provder (.e. buyer) for fndng sutable surrogate server(s) n order to perform ontent replaton. A Buyer buys the storage spae of Web server(s) of peerng CSDNs n a partular regon whh nurs exessve lent requests. Sellers are CSDN provders who sell the storage spae of ther Web server to the buyer. SLAbased alloator s the bddng agent that resdes n eah Web server. SLA negotaton module s the ommunaton medator, whh s responsble for establshng and mantanng peerng ommunatons nfrastruture. It also propagates auton messages between autoneer and bddng agents. The autoneer starts an auton not for sellng an tem (.e. alloaton), but for buyng t. Bddng agents bd wth the pre they are wllng to sell the alloaton of ther Web servers. One bdder an not see the bd of
other bdders. Autoneer gathers bds from the bdders and selets the lowest bddng agent(s) as the wnner and the wnner s pad seond-lowest bddng pre. In other words, our eonom model uses a reverse Vrey auton. In the eonom model, we also assume that the auton partpants are trustworthy. Due to ths nature, a mendaous behavor from a provder s not expeted to be usual. Hene, overprovsonng of resoures by harnessng data through VO membershp, or modfyng and falsfyng of ontent by some rogue CSDN provders s not allowed n the system model. Havng nown the nternals of the eonom model, we now dsuss the steps for formaton of VO n our eonom model: 1. A CSDN provder (buyer) realzes the need to replate ontent to the surrogates of peerng CSDNs. The buyer nternally determnes the maxmum payable amount (expressed by Payoff Value) and announes ts Auton Poly. The autoneer starts auton on the buyer s behalf. 2. The bddng agent of seller (other CSDN provders) uses a Bddng Funton to determne the bddng amount. 3. The autoneer ollets bds from the bddng agents and selets the lowest bddng surrogate(s) as wnner and a wnner s pad by the amount of seond-lowest bd. 4. Hene, a VO onsstng of the buyer and seller (.e. wnnng bdders) CSDNs s formed and ontent s replated to the wnners surrogates. 5. Re-negotaton through auton taes plae when ether of the followng holds: (a) A seller vares ts demand after wnnng; (b) Seller fnds that holdng replated ontent s no longer eonomally benefal for t; () A more ompettvely pred CSDN provder (exept wnner(s)) omes up. 5.1. System model Let us assume that N denotes the set of CSDN provders and C s the set of ontents. For a gven ontent C, S s denoted as the sze of the ontent. We also defne δ (, ) as the smlarty funton between two ontents two ontents and, 0 δ (, ) 1. The -th arrvng request to a gven CSDN provder at tme t s r omposed of l r and r, where r s the requestng ontent n C l l l and r the requestng loaton. φ ( r, r ) s the smlarty funton of two requests n terms of dstane, l l 0 φ ( r, r ) 1. We spefy D as the delay threshold wthn whh eah request should be delvered. Auton ntaton: An auton s ntated when a provder an not delver the requested ontent to a lent satsfyng ts QoS requrements. We express the QoS requrements as the user pereved networ delay (.e. response tme). The response tme of r at tme t s RT ( r, t ), where RT ( r, t ) = 0 f r s already served and RT ( r, t ) > 0 f r s not yet served. We antpate measurng the response tme based on networ topology, bandwdth, and queung delay. If there exsts a request r suh that RT ( r, t ) > D, where, a provder (buyer) realzes monetary penalty or loss goodwll for not satsfyng the lent QoS requrements. Hene, t dedes to replate ontent by announng ts serve requrements as Auton Poly A P and nternally alulates the maxmum payable amount (Payoff Value). The Payoff Value P max s alulated as, P max = (the managng ost for r ) + (the expeted proft from r ). The Auton Poly A P onssts of: Storage requrement: The storage spae requred to replate ontent, defned as S (MB or GB). Delay threshold: The tme wthn whh the ontent should be delvered, denoted as D. Preferenes: The buyer s bas for surrogate server(s) n a potental hotspot regon. Bddng funton: The bdders (other provders) bd wth the amount as determned by the Bddng Funton B ( r ) = S( r ) + ER ( r, t, n) + ψ ( AP ), where S ( r ) s the storage ost nurred by seller to replate ontent r, ER ( r, t, n) s the expeted revenue of the ontent r at tme t durng the followng n ontent requests, and ψ ( AP ) s a funton to reflet sellers nterest n the bddng of A P. Expeted revenue depends on the request pattern for next n ontent requests, and on the hstory of request for smlar ontent. Hene, t s defned as: ER( r, t, n) = α + n = + 1 l l δ ( r, r ) φ( r, r ) + (1 α) 1 = 1 l l δ ( r, r ) φ( r, r ) where 0 α 1. Auton termnaton: Autoneer (Coordnated VO sheduler) ollets bds from the bddng agents, and
selets the wnner(s) wth the lowest bd. Afterwards, ontent s replated there. The wnnng bdders are pad the seond-pre bd and a VO s formed onsstng of the buyer and seller CSDNs (wnners) based on a ommon goal to replate ontent, and to serve t to the end-users n an effent manner. The vablty of a VO may hange dependng on the demand of ontent and the partpant s eonom gan. A VO partpant should be able to adapt to a hange n ontext wthn peerng CSDN envronment. Hene, renegotaton should tae plae among the VO partpants to ether dsband or rearrange the VO nto a new organzaton that better fts the prevalng rumstanes. 6. Summary and Future Wor In ths paper, we have presented an open, salable and Serve-Orented Arhteture (SOA)-based system to assst the reaton of open Content and Serve Delvery Networs aordng to a VO model. To enourage resoure sharng and peerng among dfferent CDN provders at global level, we propose the use of maret models. Hene, we ntrodue an eonomy-based ontent replaton strategy based on auton protool for replatng ontent n surrogates of peerng CDNs. The use of eonom onepts n ths ontext provdes a sold bass for ratonal agents n peerng CSDN envronment to dede whether to attend n peerng onstellaton. Use of the eonom model may be the bass of a replaton mehansm that dynamally adapts to the hanges n ontent request pattern, and mae replaton deson to the surrogates of peerng CSDNs n areas whh exhbts the potental to generate Web hotspots. Table 1: Future researh atvtes and tmelne Tmelne Wor Desrptons Mar. 2007 Aug. 2007 (Stage 2) Sept. 2007 Mar. 2008 (Stage 3) Apr. 2008 Sept. 2008 (Stage 4) Implement the arhteture for peerng CSDNs n a realst smulaton envronment to show ts effetveness. Evaluate performane of the proposed eonomy-based ontent replaton strategy. Implement and evaluate an effetve load balanng mehansm for peerng CSDNs n a realst smulaton envronment Develop an effetve request assgnment and redreton poly and evaluate t through smulaton Ot. 2008 Mar. Develop of a prototype system for peerng 2009 (Stage 5) CDNs and deploy t n a real-world test bed n strateg loatons aross the globe to evaluate performane None of the wor done n ontent nternetworng doman has exploted a suessful peerng arrangement through the use of eonom onepts. Hene, realzng the VO model for formng CSDNs, the eonom model for ontent replaton, and deployment of the poly framewor should be a tmely ontrbuton to the ongong ontent-networng trend. Our researh on peerng ontent delvery networs wll follow a multstage proess. We have already ompleted Stage 1 whh nluded the ompleton of a hgh-level framewor for peerng CSDNs, wth the roles, responsbltes, and nteraton patterns of omponents aordng to a poly framewor. Our future atvtes are dvded nto 4 stages, whh are outlned n Table 1. For more nformaton, please vst the proet Webste at: www.grdbus.org/dn Anowledgements We are thanful to James Broberg, Kyong Hoon Km and Krs Bubendorfer of the Unversty of Melbourne for sharng thoughts on ths top, and for dsusson regardng the analytal formulaton of system model. Referenes [1] G. Peng, CDN: Content Dstrbuton Networ, Tehnal Report TR-125, Expermental Computer Systems Lab, Department of Computer Sene, State Unversty of New Yor, Stony Broo, NY 2003. [2] M. Arltt, and T. Jn, A Worload Charaterzaton Study of 1998 World Cup Web Ste, IEEE Networ, May/June 2000, pp. 30-37. [3] R. Buyya, A. M. K. Pathan, J. Broberg, and Z. Tar, A Case for Peerng of Content Delvery Networs, IEEE DSOnlne, 7(10), Los Alamtos, CA, USA, Ot. 2006. [4] R. Buyya, D. Abrahamson, and S. Venugopal, The Grd Eonomy, Pro. of the IEEE, 93(3), 2005, pp. 698-714. [5] Aama Tehnologes, In. www.aama.om, 2006. [6] S. Adler, The SlashDot Effet: An Analyss of Three Internet Publatons, Lnux Gazette, Vol. 38, 1999. [7] M. Day, B. Can, G. Tomlnson, and P. Rzews, A Model for Content Internetworng, RFC 3466, 2003. [8] E. Turrn, An Arhteture for Content Dstrbuton Internetworng, Tehnal Report UBLCS-2004-2, Unversty of Bologna, Italy, Marh 2004. [9] A. Blrs, C. Cranor, F. Dougls, M. Rabnovh, S. Sbal, O. Spatshe, and W. Sturm, CDN broerng, Computer Communatons, 25(4), Marh 2002. [10] A. Westernen, J. Shnzlen, J. Strassner, M. Sherlng, B. Qunn, S. Herzog, A. Huynh, M. Carlson, J. Perry, and S. Waldbusser, Termnology for Poly Based Management, IETF RFC 3198, Nov. 2001. [11] W. Vrey, Counterspeulaton, Autons, and Compettve Sealed Tenders, The Journal of Fnane, 16(1): 8-37, Marh 1961. [12] L. Amn, A. Shah, and H. Shulzrnne, Effetve Peerng for Mult-Provder Content Delvery Serves, In Pro. of IEEE INFOCOM, 2004, pp. 850-861.