An Archtecture for Vrtual Organzaton (VO-Based Effectve Peerng of Content Delvery Networks Al-Mukaddm Khan Pathan, James Broberg, Krs Bubendorfer*, Kyong Hoon Km, Rajkumar Buyya Grd Computng and Dstrbuted Systems (GRIDS Lab Department of Computer Scence and Software Engneerng The Unversty of Melbourne, Australa {apathan, brobergj, jysh, raj}@csse.unmelb.edu.au ABSTRACT The propretary nature of exstng Content Delvery Networks (CDNs means they are closed and do not naturally cooperate, resultng n slands of CDNs. Fndng ways for dstnct CDNs to coordnate and cooperate wth other CDNs s necessary to acheve better overall servce, as perceved by end-users, at lower cost. In ths paper, we present an archtecture to support peerng arrangements among CDN provders, based on a Vrtual Organzaton (VO model. Our approach promotes peerng among provders, reduces expendture, whle upholdng user perceved performance. Ths s acheved through proper polcy management of negotated Servce Level Agreements (SLAs among peers. In addton, scalablty and resource sharng among CDNs s mproved through effectve peerng, thus evolvng past the current landscape where slands of CDNs exst. We also show analytcally that sgnfcant performance mprovement can be acheved through the peerng of CDNs. Categores and Subject Descrptors C.2.1 [Computer-Communcaton Networks]: Network Archtecture and Desgn; C.2.2 [Computer-Communcaton Networks]: Network Protocols; C.2.4 [Computer- Communcaton Networks]: Dstrbuted Systems; H.3.4 [Informaton Storage and Retreval]: Systems and Software Dstrbuted Systems; H.3.5 [Informaton Storage and Retreval]: Onlne Informaton servce Web-based Servces General Terms Management, Measurement, Performance, Desgn, Economcs. Keywords Archtecture, Content Delvery Networks, Vrtual Organzaton, Peerng 1. INTRODUCTION Content Delvery Networks (CDNs provde servces that mprove network performance by maxmzng avalable bandwdth, mprovng accessblty and mantanng correctness through content replcaton. Thus, they offer fast and relable applcatons and servces by dstrbutng content to edge servers located close to end-users [2]. Permsson to make dgtal or hard copes of all or part of ths work for personal or classroom use s granted wthout fee provded that copes are not made or dstrbuted for proft or commercal advantage and that copes bear ths notce and the full ctaton on the frst page. To copy otherwse, or republsh, to post on servers or to redstrbute to lsts, requres pror specfc permsson and/or a fee. UPGRADE-CN 07, June 26, 2007, Monterey, Calforna, USA. Copyrght 2007 ACM 978-1-59593-718-6/07/0006...$5.00. *School of Mathematcs Statstcs and Computer Scence Vctora Unversty of Wellngton Wellngton 6140, New Zealand. krs@mcs.vuw.ac.nz Exstng commercal CDNs are propretary n nature. They are owned and operated by ndvdual companes. Each of them has created ts own closed delvery network, whch s expensve to setup and mantan. Runnng a global CDN s even more costly, requrng an enormous amount of captal and labor. In addton, content provders typcally subscrbe to one CDN and thus can not use the resources of multple CDNs at the same tme. Furthermore, commercal CDNs make specfc commtments to ther customers by sgnng Servce Level Agreements (SLAs [6]. An SLA s a contract between the servce provder and the customer to descrbe provder s commtment and to specfy penaltes f those commtments are not met. So, f a partcular CDN s unable to provde qualty of servce to the end-user requests, t may result n SLA volaton and end up costng the provder. One approach for reducng expenses and avodng adverse busness mpact s to establsh peerng arrangements among CDN provders [13]. Such peerng arrangements are appealng. However, the challenges n adoptng such an arrangement nclude desgnng a system that vrtualzes multple provders and offloads end-user requests from the prmary provder to peers based on cost, performance and load. In partcular we dentfy the followng key ssues: When to peer? The crcumstances under whch a peerng arrangement should be trggered. The ntatng condton must consder expected and unexpected load ncreases. How to peer? The strategy taken to form a peerng arrangement between multple CDNs. Such a strategy must specfy the nteractons among enttes and allow for dvergent polces among peerng CDNs. Whom to peer wth? The decson makng mechansm used for choosng CDNs to peer wth. It ncludes predctng performance of the peers, workng around ssues of separate admnstraton and lmted nformaton sharng among peerng CDNs. How to manage and enforce polces? How polces are managed accordng to the negotated SLAs. It ncludes deployng necessary polces and admnsterng them n an effectve way. In ths paper, we present a novel archtecture of a Vrtual Organzaton (VO [10] based model for formng peerng CDNs. In our archtecture, a CDN serves end-user requests as long as the load can be handled by tself. If the load exceeds ts capacty, the excess end-user requests are offloaded to the Web servers of the peers. The VO-based peerng system endeavors to cut expenses, mprovng localty whle also upholdng user perceved performance. The man contrbutons of ths paper are:
an archtecture for an open and decentralzed system that supports the effectve peerng of CDNs wthn a scalable VObased model, and a polcy-based framework for SLA negotaton among peerng CDNs to ensure that requests are effectvely served, meetng user QoS requrements. a prelmnary analytcal model based on the fundamentals of queung theory to demonstrate the performance gan through the proposed VO-based peerng of CDNs. The rest of the paper s structured as follows: Secton 2 presents the related work. Secton 3 presents our archtecture for peerng CDNs wth a broad vew of the VO creaton steps and VO formaton scenaros. It also provdes a descrpton on the archtectural components and features. Secton 4 descrbes the polcy management n peerng CDNs envronment, wth a focus on negotated SLAs and defned polcy levels. Secton 5 outlnes analytcal model for peerng CDNs. Fnally, Secton 6 concludes the paper wth a bref summary of expected contrbutons and future drectons. 2. RELATED WORK Peerng of content delvery networks s ganng popularty among researchers of the scentfc communty. Several projects are beng conducted for fndng ways to peer CDNs for better overall performance. In ths secton, we outlne some of these efforts. The nternet draft by IETF proposes a Content Dstrbuton Internetworkng (CDI Model [11], whch allows CDNs to have a means of afflatng ther delvery and dstrbuton nfrastructure wth other CDNs who have content to dstrbute. The CDI Internet draft assume a federaton of CDNs but t s not clear how ths federaton s bult and by whch relatonshps t s characterzed. A protocol archtecture for CDI s presented n [12]. The man drawback of ths protocol s beng a pont-to-pont protocol, f one end-pont s down the connecton remans nterrupted untl that end-pont s restored. A peerng system for content delvery workloads n a federated, mult-provder nfrastructure has been presented n [13], but the peerng strategy, resource provsonng and performance guarantees among partnerng CDNs are unexplored n ths work. CDN brokerng [14] allows one CDN to ntellgently redrect endusers dynamcally to other CDNs n that doman. The drawback s that, the routng mechansm used s propretary n nature and mght not be sutable for a generc CDI archtecture. From a usersde perspectve, Cooperatve Networkng (CoopNet [8] addresses the flash crowd problem through the cooperaton of end-hosts. The man problem of the user-sde mechansms s that they are not transparent to end-users, whch are lkely to restrct ther wdespread deployment. Other systems such as Coral [17], CoDeeN [18], Globule [15], and DotSlash [16] address the ssue of collaboratve content delvery. However, they do not vrtualze multple provders for cooperatve management and delvery of content n a peerng envronment. 3. VO-BASED PEERING CDNS A CDN s expected to provde the necessary dstrbuted computng and network nfrastructure to ensure SLAs are met wth ts customers. In order to meet such SLAs and to manage ts resources properly, t could be necessary to cooperate wth other CDNs. In our archtecture, cooperaton among the peerng CDNs s acheved through a VO. Formaton of a VO s ntated by a CDN, whch realzes that t wll not be able to meet ts SLAs wth the customers. The ntator s called a prmary CDN; whle other CDNs who share ther resources n a VO are called peerng CDNs. Users nteract transparently wth the VO by requestng content from Web servers of the prmary CDN. A content request may ntate further VO actvtes that the end-users are unaware of (e.g. nter-cdn request-routng, content replcaton and delvery n a peerng arrangement. Thus, the partcpatng enttes act as a sngle conceptual unt n the executon of common goal(s. Fgure 1: A formed VO A VO s composed of explct members who are the prmary and any peerng CDNs who cooperate for resource sharng, and mplct members who are content provders and end-users. Implct members are transparent to a VO but they share the beneft from t. Fgure 1 shows the example of VO-based peerng CDNs. End-user requests for content are forwarded to the prmary CDN whch s holdng the content from the content provder. The requested content s served ether drectly by the prmary CDN or by any peerng CDNs wthn a VO. Let us consder content c1 and c3 n Fgure 1. Snce c1 and c3 resde n the Web servers wthn VO1 and VO2 respectvely, requests of c1 and c3 are served accordngly from VO1 and VO2. In case of content c2, the prmary CDN drectly delvers the requested content. Termnology Web server (WS Medator Servce regstry (SR Peerng Agent (PA Polcy repostory (PR P WS P M P VO Table 1: Lst of commonly used terms Descrpton A Contaner of content A polcy-drven entty, authortatve for polcy negotaton and management Dscovers and stores resource and polcy nformaton n local doman A resource dscovery module n the peerng CDNs envronment A storage of Web server, medator and VO polces A set of Web server-specfc rules for content storage and management A set of medator-specfc rules for nteracton and negotaton A set of VO-specfc rules for creaton and growth of the VO
Fgure 2: Archtecture of a system to assst the creaton of peerng CDNs 3.1 System archtecture The archtecture of our VO-based peerng CDNs s shown n Fgure 2. The termnologes used to descrbe the system archtecture are lsted n Table 1. In the VO-based peerng CDNs model, a CDN endeavors to balance ts servce requrements aganst the hgh costs of deployng customer- dedcated, overprovsoned resources. Thus, to cut expenses and avod the potental peak load threat of volatng SLAs wth the customers, CDNs wll be able to leverage computng and network nfrastructure from other CDNs through peerng. The negotaton among CDNs for resource peerng allows a peerng CDN to agree to allocate the requred amount of ts local resources (Web servers, bandwdth etc. to provde content and servces on behalf of the prmary CDN. A peerng arrangement among CDNs that provsonng and sharng of computng resources must also provde settlement and exchange of generated revenue. The prmary CDN ultmately controls the resources t has acqured whch are delegated rghts for the peerng CDNs physcal resources. The physcal resources could consst of resources from multple peerng CDNs dstrbuted over dfferent geographcal locatons. The prmary CDN determnes what proporton of the Web traffc (.e. end-user requests s redrected to the Web servers of the peerng CDNs, whch content s replcated there, how the replcaton decsons are taken, and whch replcaton polces are beng used. In our archtecture, Web Servers (WS wthn a CDN are the actual holders of content. Each Web server has ts own polces, defned as a set of server-specfc rules, P WS, for the storage and management of content. The Servce Regstry (SR helps n dscoverng local resources wthn a CDN by provdng resource and access related nformaton. The Peerng Agent (PA, Medator and Polcy Repostory (PR collectvely act as a gateway for a gven CDN, and all three assst n creaton of a new VO. The PA of a CDN acts the role of a resource dscovery module. It acts as the frst pont of contact for other CDNs when they are ntatng a peerng agreement, and a condut through whch a CDN can tself dscover potentally useful resources avalable from other CDNs. The medator s responsble for negotaton among CDNs and management of operatons wthn a VO. The medator has ts own polces (defned as a set of medator-specfc rules, P M and also manages the polces (defned as a set of VO-specfc rules, P VO necessary for negotaton and creaton of Vrtual Organzaton(s. The PR vrtualzes all of the ndvdual polces from wthn the VO, ncludng P WS, P M, and P VO, and wll ultmately nclude polces from peerng CDNs. 3.2 Lfecycle of a VO A VO may vary n terms of purpose, scope, sze, and duraton. Hence, VOs are of two types: short-term on-demand VOs and long-term VOs wth establshed SLAs. A short-term VO s formed for lmted duraton, based on current user request patterns to prevent the generaton of hotspots [9]. Such a peerng arrangement should be automated to react wthn a tght tme frame as t s unlkely that a human drected negotaton would occur quckly enough to satsfy the evolved nche. A short-term VO s formed on-demand and the polcy for such VO formaton s establshed dynamcally to handle the stuaton, one such negotaton mechansm s brefly descrbed n secton 3.4. Shortterm VOs are phased out when the workload returns to normal. On the other hand, a long-term VO s formed for an event whch wll be known n advance. In a long-term VO, CDNs collaborate for longer perod of tme and such a VO remans for the duraton of the event. In ths stuaton we would expect negotaton to nclude a human-drected agent to ensure that any resultng decsons comply wth partcpatng companes strategc goals.
$! " # Fgure 3: The formaton of a Vrtual Organzaton (VO Fgure 4: A formed VO Relevant scenaros for short and long-term VO creaton have been llustrated n [1]. Fgure 3 llustrates the VO creaton steps, whle Fgure 4 shows a VO after t s formed. The followngs are typcal steps for a VO creaton: Step 1. A (prmary CDN provder realzes that t cannot handle a part of the workload on ts Web server(s. A VO ntalzaton request s sent to the medator. Step 2. The prmary CDN constructs a shell VO, wth a medator nstance, a servce regstry nstance, and a polcy regstry. The medator nstance obtans the resource and access nformaton from the SR, whlst SLAs and other polces from the PR. Step 3. The shell VO represents the potental for peerng of resources. Hence, t needs to be expanded to nclude addtonal resources from other CDNs. The medator nstance on the prmary CDN s behalf generates ts servce requrements based on the current crcumstance and SLA requrements of ts customer(s. Step 4. The medator nstance passes the servce requrements to the local Peerng Agent (PA. If there are any preexstng peerng arrangements (for a long term scenaro then these wll be returned at ths pont. Otherwse, t carres out short term negotatons wth the PA dentfed peerng targets (Secton 3.4. Step 5 When the prmary CDN acqures suffcent resources from ts peers to meet ts SLA wth the customer, the new VO becomes operatonal. If no CDN s nterested n such peerng, VO creaton through re-negotaton s resumed from Step 3 wth reconsdered servce requrements. An exstng VO may need to ether dsband or re-arrange tself f any of the followng condtons hold: (a the crcumstances under whch the VO was formed no longer hold; (b peerng s no longer benefcal for the partcpatng CDNs; (c an exstng VO needs to be expanded further n order to deal wth addtonal load; or (d partcpatng CDNs are not meetng ther agreed upon contrbutons. 3.3 Archtectural components In ths secton, we provde the descrpton of the archtectural components along wth ther responsbltes: Web server Web Servers are responsble for storng content and delverng them relably. We separate a server s structure nto two layers: overlay and core. In the overlay layer, a server comprses a Web-servce host (for example, Apache or Tomcat, a polcy agent, and a SLA-allocator. The Web servces host ensures the delvery of content to end-users based on the negotated polces. The polcy agent s responsble (n conjuncton wth the medator for determnng whch resources can be delegated and under what condtons (polces delegaton s permtted. The SLA-allocator performs the provsonng and reservaton of Web server s resources (e.g. CPU, bandwdth, storage etc. to satsfy both local and delegated SLAs, and ensures that the terms of the SLAs are enforced. The Web server s core conssts of hgh performance computng systems such as symmetrc multprocessors, cluster systems, or other enterprse systems (such as desktop grds. The Web servers underlyng algorthms perform on-demand cachng, content selecton, and routng between servers. Ths requres each Web server to express ts own polces for storage and management of content. Medator The (resource medator s a polcy-drven entty n a VO-based peerng CDNs envronment. The rason d'etre for an nstance of the medator wthn a VO s to ensure that the partcpatng enttes are able to adapt to changng crcumstances (aglty and are able to acheve ther objectves n a dynamc and uncertan envronment (reslence. Once a VO s establshed, the medator controls what porton of the Web traffc (.e. end-user requests s redrected to the Web servers of the peerng CDNs, whch content s replcated there, how the replcaton decson s taken, and whch replcaton polces are beng used. When performng automated peerng the medator wll also drect any decson makng durng peerng negotatons (durng VO
creaton, polcy management, and schedulng. A medator holds the ntal polces for VO creaton and obtans addtonal composte polces as a result of successful peerng negotatons. A medator works n conjuncton wth ts local Peerng Agent (PA to dscover external resources and to negotate wth other CDNs. An example of a medator led negotaton s gven n Secton 3.4. Servce Regstry (SR The SR encapsulates the resource and servce nformaton for each CDN. It helps n dscoverng local resources through enablng the Web servers of CDN provders to regster and publsh ther resource, servce and polcy detals. In the face of traffc surges, the servce regstry s accessed by the medator n shell VO creaton to supply any necessary local resource nformaton. When a shell VO s grown to form a new VO, an nstance of the servce regstry s created that encapsulates all local and delegated external CDN resources. Polcy Repostory (PR The PR vrtualzes all of the polces wthn the VO. It ncludes the Web server-specfc polces, medator polces, VO creaton polces along wth any polces for resources delegated to the VO as a result of a peerng arrangement. These polces form a set of rules to admnster, manage, and control access to VO resources. They provde a way to manage the components n the face of complex technologes. Peerng Agent (PA The PA s a CDN specfc entty that exsts pror to a VO creaton. It s ndependent of any VO. It acts as a polcy-drven resource dscovery module for VO creaton and s used as a condut by the medator to exchange polcy and resource nformaton wth other CDNs. It s used by a prmary CDN to dscover the peerng CDNs (external resources, as well as to let them know about the local polces and servce requrements pror to commencement of the actual negotaton by the medator. 3.4 Short-term resource negotaton In order to respond to hotspots that may result n a CDN falng to meet ts QoS oblgatons, we propose that tme-crtcal agreements for a short-term VO should be automatcally negotated. We expect any such agreements to hold for a lmted duraton and only nvolve an artfcally restrcted set of CDN resources. Even so, there are serous ssues nvolvng such agreements ncludng trust,.e. who governs the allocaton decsons and are they trustworthy; and the potental commercal senstvty of nformaton about the current state and costs of a CDN. Dvulgng commercally senstve nformaton (e.g. resources, access and polces as a bass for negotatng a peerng arrangement would be, n general, commercally unacceptable. Even wth the lmtatons placed on resources that can be automatcally delegated, there must be checks and balances to ensure that any delegaton s made properly. Otherwse, t would also be unlkely that a CDN would agree to have an external party (e.g. medator of the prmary CDN make allocatons of ther resources and bnd them to negotated SLAs. One soluton to these problems s to utlze a cryptographcally secure aucton [22], whch hdes both the valuatons that CDNs place on ther resources and who s partcpatng n the aucton. In ths case all the medators would between themselves act as a secure dstrbuted auctoneer, n whch the cryptographc protocol tself guarantees a trustworthy outcome of the aucton. These auctons have been shown to be tractable n practce and are therefore an deal bass for automaton of peerng agreements. A negotaton would start wth the VO beng formed and the medator determnng the shortfall n resources. The medator then ssues ts local PA a call for bds and wthn ths ncludes the SLAs that t requres. The PA dstrbutes the request to other PAs of the peers and each of them then passes the request to ts CDN medator. All medators that wsh to bd then regster wth the requestng medator and a subset of the medators (actng as the dstrbuted auctoneers are selected va a cut and shuffle [23]. Note that the requestng medator does not act as an auctoneer. The aucton s then held securely and only the fnal resource allocatons are revealed. An auctoneer starts an aucton not for sellng an tem (.e. allocaton, but for buyng t. Buyers (peerng CDNs bd wth the prce they are wllng to sell the allocaton of ther Web servers. One bdder can not see the bd of other bdders. An auctoneer gathers bds from the bdders and selects the lowest bddng agent(s as the wnner and the wnner s pad second-lowest bddng prce. In other words, a reverse Vckrey aucton s used. As mentoned earler, we assume that an aucton s held usng a cryptographcally secure aucton [22] protocol to hde all aucton related senstve nformaton. Through ths approach, a mendacous behavor from a provder s restrcted. Thus, overprovsonng of resources by harnessng data through VO membershp, or modfyng and falsfyng of content by some rogue CDN provders s not allowed. Here below, we also summarze the steps for the aucton to be held wthn a VO: Step 1. The medator of the prmary CDN (buyer realzes the need of addtonal resources to replcate content. It nternally determnes the maxmum payable amount (expressed by Payoff Value. The medator then ssues ts local PA a call for bds and wthn ths ncludes the SLAs that t requres (Aucton Polcy. Step 2. The PA dstrbutes the request to other PAs of the peers and each of them then passes the request to ts CDN medator. Step 3. All medators that wsh to bd then regster wth the requestng medator and a subset of the medators are chosen to act as dstrbuted auctoneers. Step 4. All bdders (peerng CDNs use a Bddng Functon to determne the bddng amount. Step 5. An auctoneer collects bds and selects the lowest bddng buyer(s as wnner and a wnner s pad by the amount of second-lowest bd. Step 6. An aucton takes place successfully when wnners are chosen accordng to the Aucton polcy of the prmary CDN. At ths pont, t can be assumed that the prmary CDN has acqured suffcent resources from ts peers to meet ts SLA wth customers. Therefore, the prmary CDN replcates ts content to the wnnng CDNs Web servers. If no wnner s selected through the aucton, re-negotaton through aucton takes place, startng from Step 1. 3.5 Archtectural features The operaton of a CDN s drven by sem-autonomous logc that ensures content s served relably through content replcaton, request-routng and redrecton whlst mantanng constant awareness of the health (e.g. load nformaton of partcpants. Request assgnment and redrecton can be performed n a CDN at multple levels at the DNS, at the gateways to local clusters and also (redrecton between servers n a cluster [19][20]. Commercal CDNs predomnantly rely on DNS level end-user
assgnment combned wth a rudmentary request assgnment polcy (such as weghted round robn, or least-loaded-frst whch updates the DNS records to pont to the most approprate replca server [7]. In the proposed framework, end-users can be assgned va DNS (by the peerng agents of partcpatng CDNs updatng ther DNS records regularly and also va redrecton at the CDN gateway when approprate. Content replcaton occurs from orgn servers to other servers wthn a CDN. Exstng CDN provders (e.g. Akama, Mrror Image use a non-cooperatve pull-based approach, where requests are drected (va DNS to ther closest replca server [7]. If the fle requested s not held there, the replca server pulls the content from the orgn server. Co-operatve push-based technques have been proposed that pushes content onto partcpatng mrror servers usng a greedy-global heurstc algorthm [21]. In ths approach, requests are drected to the closest mrror server, or f there s no sutable mrror nearby, t s drected to the orgn server. In the context of our peerng framework, ths replcaton extends to partcpatng servers from other CDNs n a gven VO, subject to the avalable resources t contrbutes to the VO. Ths s defned by the polces (P VO agreed upon at VO creaton tme. Load nformaton can be measured and dssemnated wthn ndvdual CDNs and amongst other CDNs. A load ndex can provde a measure of utlzaton of a sngle resource on a computer system. Alternatvely, t can be a combned measure of multple resources lke CPU load, memory utlzaton, dsk pagng and actve processes. Such load nformaton needs to be dssemnated amongst all partcpatng CDNs n a tmely and effcent manner to maxmze ts utlty. Such ndces wll also be crucal to dentfy stuatons where formng a VO-based peerng s approprate (e.g. when servers or entre CDNs are overloaded or when CDNs resources are underutlzed and could be offered to other CDN provders. Wthn the VO-based framework, we antcpate a herarchcal approach, where current bandwdth and resource usage of web servers n a CDN s reported to the CDN gateway (.e. medator, PA and polcy repostory as a sngle conceptual entty n a perodc or threshold-based manner. The gateways of partcpatng CDNs then communcate aggregated load nformaton descrbng the load of ther consttuent servers. 4. SLA NEGOTIATION AND POLICY MANAGEMENT When CDNs peer accordng to a VO-based model, the partcpants sgn SLAs wth dfferent performance objectves. Once the SLAs have been agreed upon, the partcpants n a VO work n order to satsfy the negotated SLAs. The SLA components nclude: Descrpton of servce requrements, a specfcaton of the resource and servce requrements of the prmary CDN. Ths descrpton ncludes the storage requrements, the requred rate of transfer, the prmary CDN s preference to gan resources at a partcular regon, and the expected duraton of recevng servce. Admnstraton for VO actvtes, whch specfes the role of the medator as an authortatve entty n the VO. Renegotaton for problem resoluton, whch llustrates the steps to be undertaken n face of any problem n provdng necessary servces. Consequences of SLA volaton, whch outlnes the possble results of SLA volaton n whch servce expectatons are not met. The consequences of SLA volaton may range from mposng penalty on the partcpants through rembursement of part of the revenues lost due to the loss of servce, to termnaton of peerng relatonshp, and to dsbandng and/or rearrangng the partcpants to form a new VO. SLA bypassng condtons, whch detals the condtons under whch the SLAs are not applcable. Such stuatons nclude the damage of physcal resources due to natural dsaster, theft etc. 4.1 Polcy management to support SLAs The proper operatons of a VO-based peerng CDNs archtecture seek for the consstent performance and avalablty of a large number of wdely dstrbuted system enttes, specfed n a Servce Level Agreement (SLA. A polcy-based framework can smplfy the complextes nvolved n the operaton and management of a large content dstrbuton network [5]. Wthn our VO-based peerng CDNs archtecture we apply the standard polcy framework defned by the IETF/DMTF [4]. We also defne three levels of polces, namely Web server polcy, medator polcy and VO polcy. These three polcy levels are detaled n Secton 4.2. Fgure 5: Basc polcy framework We defne a polcy as a descrptve statement that allows an entty to properly admnster, manage, and execute ts actvtes. Polces n the context of peerng CDNs are statements that specfy how the Web servers should deal wth dfferent types of traffc, what type of content can be moved out to a CDN node, what resources can be shared between the VO partcpants, what measures are to be taken to ensure qualty of servce based on negotated SLAs, and what type of programs/data must be executed at the orgn servers. Thus these polces endeavor to provde a way to manage multple enttes deployng complex technologes wthn the archtecture of the VO-based peerng CDNs. We llustrate the standard polcy framework n Fgure 5. In the standard polcy framework, the admn doman refers to an entty whch admnsters, manage and control access to resources wthn the system boundary. An admnstrator uses the polcy management tools to defne the polces to be enforced n the system. The polcy enforcement ponts (PEPs are logcal enttes wthn the system boundary, whch are responsble for takng
Table 2: Polcy mappng Polcy framework Peerng CDNs Component Specfed polces Descrpton Component System Peerng CDNs All polces n the system The dstrbuted computng and network nfrastructure for peerng CDNs Admn doman Formed VO Negotated VO polces An admnstratve entty for resource management and access control Polcy management tool Admnstrator dependent An admnstrator dependent tool to generate polces Polcy repostory Polcy repostory Web server, VO and Storage of polces n the system medator polces Polcy Enforcement Web Servces host, Polcy Web server polces A logcal entty whch ensures proper enforcement of Ponts (PEPs Agent, SLA-based allocator PDPs Medator Medator polces, VO polces polces An authortatve entty for retrevng polces from the repostory acton to enforce the defned polces. The polcy repostory stores polces generated by the admnstrators usng the polcy management tools. The polcy decson pont (PDP s responsble for retrevng polces from the polcy repostory, for nterpretng them (based on polcy condton, and for decdng on whch set of polces are to be enforced (.e. polcy rules by the PEPs. The model (and the correspondng enttes for peerng CDNs can be mapped to the basc polcy framework. We show ths mappng n Table 2. The polcy repostory (PR from Fgure 2 vrtualzes the Web server, medator and VO polces. These polces are generated by the polcy management tool used by the VO admnstrator. The dstrbuton network and the Web server components (.e. Web Servces host, Polcy Agent, SLA-based Allocator are the nstances of PEPs, whch enforce the peerng CDN polces stored n the repostory. The medator s the nstance of the PDPs, whch specfes the set of polces to be negotated at the tme of collaboratng wth other CDNs, and passes them to the peerng agent at the tme of negotaton. The polcy management tool s admnstrator dependent and t s not shown n Fgure 2. 4.2 Polcy levels In ths secton we delneate dfferent levels of polcy defntons that are present wthn the doman of peerng CDNs. We propose followng a general mplementaton of polces whch can be specfed accordng to ther granularty level. Beng nfluenced by [5], we also argue for storng the specfcaton of the polcy rules (for all three levels of polces n the polcy repostory for mantanng nteroperablty. We also ntate smplfed management by storng the defnton of polces n the polcy repostory, and by defnng abstractons that provde a machne dependent specfcaton of polces. We propose that the polcy specfcaton contaned n the polcy repostory should be defned n terms of the technology to whch a polcy would apply, rather than n terms of the confguraton parameters of any specfc resource. Here below, we descrbe the dstnct polcy levels: Web server polces specfy, how a server performs consstent content cachng; how the polcy agent module operates based on negotated SLAs; and how the SLA-based allocator module ensures provsonng of resources to satsfy negotated SLAs. Medator polces specfy, how the medator nteracts wth the PA to pass nformaton on servce requrements; how the medator takes over the admnstraton of delegated resources once they are acqured; how the medator effectvely manages VO actvtes to cope wth changng crcumstances; and how the medator coordnates wth other enttes to assst n resource allocaton. VO polces nclude, the polces necessary for ntatng VO creaton; the polces need to be admnstered n face of SLA volaton by VO partcpants; and the polces to dynamcally dsband or rearrange a VO. 5. ANALYTICAL MODELING In ths secton, we develop a smple analytcal model based on the fundamentals of queung theory to demonstrate the performance gan through the peerng of CDNs. λ(t Fgure 6: Prmary CDN modeled as an M/G/1 queue Let us model a CDN as an M/G/1 queue as shown n Fgure 6. Table 3 shows the parameters and expressons that are used n the analytcal modelng. An M/G/1 queue conssts of a FIFO buffer wth requests arrvng randomly accordng to a Posson process at rate λ and a processor, called a server, whch retreves requests from the queue for servcng. We assume that the total processng of the Web servers of a CDN s accumulated through the server and the servce tme s a general dstrbuton. User requests are servced at a frst-come-frst-serve (FCFS order. We use the term task as a generalzaton of a content request arrval for servce. We denote the processng requrements of an arrval as task sze. Here, we wll use the terms task and request arrval nterchangeably. Such a content request can be a clent requestng
Table 3: Parameter and expressons for the analytcal model Parameter Expresson Mean arrval rate λ = ( 1/ T (requests/second Mean arrval tme Mean servce rate T µ Load ρ = ( λ µ k P. D. F of servce dstrbuton 1 ( k p k x p Task sze varaton Smallest possble task sze Largest possble task sze k p Expected watng tme E (W Mean servce tme E (X an ndvdual fle or object, a Web page (contanng multple objects, the results of executon of a scrpt (e.g. CGI, PHP or any dgtal content. We abstract all the request streams comng to the Web servers of the prmary CDN as a sngle request stream. Clent requests arrve at a conceptual entty, called dspatcher, followng a Posson process wth the mean arrval rate λ. All requests n ts queue are served on a FCFS bass wth mean servce rate µ. It has been observed that the workloads n Internet are heavy-taled n nature [24][25], characterzed by the functon, Pr{ X > x} ~ x, where 0 2. In a CDN, clents request for content of varyng szes (rangng from small to large. Based on sze of the content requested, the processng requrements (.e. task sze also vary. Thus, the task sze on a gven CDN s servce capacty follows a Bounded Pareto dstrbuton. The probablty densty functon for the Bounded Pareto B ( k, p, s k 1 f ( x = x, 1 ( k p where represents the task sze varaton, k s the smallest possble task sze, and p s the largest possble task ( k x p. By varyng the value of, we can observe dstrbutons that exhbt moderate varablty ( 2 to hgh varablty ( 1. We start wth the dervaton of the expectaton of watng tme E (W. W s the tme a user has to wat for servce. E N s the ( q number of watng customers and X s the mean servce tme. By Lttle s law, the mean queue length E N can be expressed ( q n terms of the watng tme. Therefore, N q = λ W and j load on the server, ρ = λx. Let E ( X be the j-th moment of the servce dstrbuton of the tasks. We have, j j p (( k p ( k p j ( j (1 ( k p X = k ln( p k (1 ( k p 1 f ( x = x, f j f j = Expected Watng Tme W Expected Watng Tme W 100000000 10000000 1000000 100000 10000 1000000 100000 10000 No-Peerng 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.99 Load on Server (ro Hence, usng P-K formula, we obtan the expected watng tme n 2 the prmary CDN s queue, E ( W = λ X / 2(1 ρ. Now we want to measure the expected watng tme wth respect to varyng server load and task szes. 5.1 Performance gan through peerng Now, let us assume that we have N peerng CDNs n the system. All the partcpants share ther resources to deal wth flash crowds. In face of sudden surge n demand, the load on the prmary CDN becomes, ρ = ρ 1 P and the redrected ( redrect Peerng of CDNs Fgure 7: Effectveness of peerng among CDNs No-Redrecton 1000 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.95 Redrecton Probablty Request Redrecton Fgure 8: Impact of request-redrecton on expected watng tme load s dstrbuted equally among the peerng CDNs. Fgure 7 shows the effectveness of peerng among CDNs n terms of expected watng tme for a two CDNs case. In Fgure 7, we can see a reasonable mprovement n expected watng tme through peerng. Wthout peerng when the system load approaches to 1.0, the user perceved response tme for servce by the prmary CDN tends to nfnty. In ths case, the prmary CDN peers wth other CDNs for coordnated and cooperatve delvery of content. Hence, as the prmary CDN becomes overloaded, t redrects some of ts request to the peers. In ths way, even n the face of hgh demand, user perceved performance remans satsfactory. Fgure 8 shows the mpact of request-redrecton on the expected watng tme wth hgh system load ( ρ = 0.95 wth task varablty, = 1.5. In Fgure 8, we show that as the redrecton probablty
ncreases, approachng to 1, expected watng tme on the prmary CDN decreases substantally. 5.2 Three CDN peerng scenaro In ths secton, we examne the performance gan through peerng n three CDN peerng arrangement, as llustrated n Fgure 9. Each of the partcpatng CDNs s modeled as an M/G/1 queue. As shown n Fgure 9, when the prmary CDN s unable to serve all the ncomng content requests, t redrects a fracton of the request to the peerng CDN, whch n return serve those requests. Note that, the redrected requests are equally dstrbuted between the peerng CDNs. Under ths assumpton, we can measure the expected watng tme for end-user requests for each of the three partcpatng enttes n the peerng arrangement. λ ( T 1 λ ( T 2 λ 3( T Fgure 9: A peerng scenaro wth three CDNs Fgure 10 shows the mpact of redrectng the requests on expectng watng tme for the three CDNs peerng scenaro and Table 4 lsts the notatons used for ths scenaro. Intally, the prmary CDN and peer 1 have moderate system load ( ρ = 0.5, whereas peer 2 s less loaded ( ρ = 0.3. As the load on the prmary CDN ncreases (approachng to 1.0, t offloads content requests to the peers, wth redrecton probablty, P redrect = 0. 5. For ths scenaro we assume that content requests are comng as dfferent request streams to the partcpatng CDNs. We also assume that each of the partcpatng enttes has the same servce rate. The ntal and new load on CDN s measured by, ρ = λ X and ρ = λ X respectvely. λ s the ntal arrval rate, whereas, λ s the new arrval rate at CDN. We calculate λ as, (1 Predrect λ f = 1 λ = Predrect λ + λ1( f > 1 N We measure the ntal and new expected watng tme of CDN as, λ X W = and 2(1 ρ 2 2 λ X W =, respectvely. 2(1 ρ From Fgure 10, we fnd that, the expected watng tme of the prmary CDN decreases as the requests are redrected to the peerng CDNs. Peerng CDNs, on the other hand, show an ncrease n expected watng tme as t receves extra requests from the prmary CDN. Table 4: Lst of notatons n three CDNs peerng scenaro Notaton Descrpton N Number of CDNs, N { 1,2,..., N} j j-th moment of CDN s servce dstrbuton E ( X ρ Intal load on CDN, ρ = λ E X ( ( X ρ New load on CDN, ρ = λe λ Intal arrval rate at CDN λ New arrval rate at CDN, (1 Predrect λ f = 1 λ = Predrect λ + λ1( f > 1 N E ( W Expected watng tme (ntal at CDN, 2 λ X W = 2(1 ρ E ( W Expected watng tme (new at CDN, 2 λ X W = 2(1 ρ P Probablty to redrect content requests from redrect CDN Expected Watng Tme W 10000000 1000000 100000 10000 Prmary CDN: No-Peerng Prmary CDN: After Peerng Peer 1: After Peerng Peer 2: After Peerng 0.5 0.6 0.7 0.8 0.9 0.95 0.99 Load on Prmary CDN Fgure 10: Three CDNs scenaro - performance gan through peerng 6. CONCLUSION AND FUTURE WORK In ths paper, we present an open and scalable system to assst the creaton of open content delvery networks. In our archtecture, when the load on the prmary CDN exceeds ts capacty, t peers wth other CDNs, and the excess end-user requests are offloaded to the Web servers of the peerng CDNs. Our contrbuton les n desgnng an archtecture for the VO-based peerng approach that endeavors to reduce setup and mantenance cost of network nfrastructures, whle also respectng end-user performance requrements through proper polcy management for negotated SLAs. It also promotes extended scalablty and resource sharng wth other CDNs through cooperaton and coordnaton. We also constructed a prelmnary analytcal model for peerng CDNs demonstratng the performance gan through peerng. Such results hghlght the utlty of peerng among the exstng CDNs. We
antcpate that, wth more advanced modelng the proposed framework wll motvate and drect our research n fndng best practce technques n measurng and dssemnatng load nformaton, performng request assgnment and redrecton, and enablng content replcaton for CDNs partcpatng n VO-based peerng. No pror work done n the content nternetworkng doman has consdered VO-based peerng among CDNs. Many of them make strong assumptons on the characterstcs of applcatons wthout vrtualzng multple provders for cooperatve management and delvery of content n a peerng envronment. Moreover, none of these systems have explored the ssue of polcy management. Our future work ncludes usng market models n ths context n order to encourage resource sharng and peerng among dfferent CDNs at global level. Ths approach s nspred by the successful utlzaton of economc concepts n management of autonomous resources n global grds [3]. The use of economc concepts n ths context would provde a sold bass for ratonal agents to decde whether to jon n peerng arrangements. The use of economc models may be the bass for a dynamc replcaton mechansm that makes replcaton decsons to utlze surrogates n areas whch exhbt the potental to generate hotspots. Our ntal work on ths ssue can be found n [1]. We expect that a VO polcy drven model for formng CDNs wll be a tmely contrbuton to the ongong content-networkng trend. For more nformaton, please vst the project Web ste at www.grdbus.org/cdn. REFERENCES [1] Pathan, A. M. K. and Buyya, R. Economy-based content replcaton for peerng CDNs. TCSC Doctoral Symposum, In Proc. of the 7 th IEEE Internatonal Symposum on Cluster Computng and the Grd (CCGrd 2007, Brazl, May, 2007. [2] Buyya, R., Pathan, A. M. K., Broberg, J., and Tar, Z. A case for peerng of content delvery networks. IEEE Dstrbuted Systems Onlne, 7(10, USA, Oct. 2006. [3] Buyya, R., Abrahamson, D., and Venugopal, S. The Grd economy. Proc. of the IEEE, 93(3, pp. 698-714, 2005. [4] Westernen, A., Schnzlen, J., Strassner, J., Scherlng, M., Qunn, B., Herzog, S, Huynh, A., Carlson, M., Perry, J., and Waldbusser, S. Termnology for polcy based management. IETF RFC 3198, Nov. 2001. [5] Verma, D. C., Calo, S., and Amr, K. Polcy-based management of content dstrbuton networks. IEEE Network, pp. 34-39, March/Aprl 2002. [6] Bouman, J., Trenekens, J., and Zwan, M. Specfcaton of servce level agreements, clarfyng concepts on the bass of practcal research. In Proc. of the Software Technology and Engneerng Practce Conference, pp. 169, 1999. [7] Dlley, J., Maggs, B., Parkh, J., Prokop, H., Staraman R., and Wehl, B. Globally dstrbuted content delvery. IEEE Internet Computng, pp. 50-58, Sept./Oct. 2002. [8] Padmanabhan, V. N. and Srpandkulcha, K. The case for cooperatve networkng. In Proc. of Internatonal Peer- To-Peer Workshop (IPTPS02, 2002. [9] Adler, S. The SlashDot Effect: An analyss of three Internet publcatons. Lnux Gazette, Vol. 38, 1999. [10] Norman, T. J., Preece, A., Chalmers, S., Jennngs, N. R., Luck, M., Dang, V. D., Nguyen, T. D., Deora, V., Shao, J., Gray, W. A., and Fddan, N. J. Agent-based formaton of vrtual organzatons. Knowledge-Based Systems, 17 (2-4, pp. 103-111, 2004. [11] Day, M., Can, B., Tomlnson, G., and Rzewsk, P. A model for content nternetworkng. IETF RFC 3466, Feb. 2003. [12] Turrn, E. An archtecture for content dstrbuton nternetworkng. Techncal Report UBLCS-2004-2, Unversty of Bologna, Italy, March 2004. [13] Amn, L., Shakh, A., and Schulzrnne, H. Effectve peerng for mult-provder content delvery servces. In Proc. of 23 rd Annual IEEE Conference on Computer Communcattons( INFOCOM 04, pp. 850-861, 2004. [14] Blrs, A., Cranor, C., Dougls, F., Rabnovch, M., Sbal, S., Spatscheck, O., and Sturm, W. CDN brokerng. Computer Communcatons, 25(4, pp. 393-402, 2002. [15] Perre, G. and Steen, M. Globule: A platform for selfreplcatng Web documents. In Proc. of the 6th Internatonal Conference on Protocols for Multmeda Systems (PROMS 01, Enschede, The Netherlands, pp. 1-11, 2001. [16] Zhao, W. and Schulzrnne, H. DotSlash: A selfconfgurng and scalable rescue system for handlng Web hotspots effectvely. In Proc. of the Internatonal Workshop on Web Cachng and Content Dstrbuton (WCW, Bejng, Chna, pp. 1-18, 2004. [17] Freedman, M. J., Freudenthal, E., and Mazères, D. Democratzng content publcaton wth Coral. In Proc. of the 1 st Symposum on Networked System Desgn and Implementaton (NSDI 04, San Francsco, CA, pp. 239-252, 2004. [18] Wang, L., Park, K. S., Pang, R., Pa, V. S., and Peterson, L. Relablty and securty n the CoDeeN content dstrbuton network. In Proc. of the USENIX 2004 Annual Techncal Conference, 2004. [19] Colajann, M., Yu, P. S., and Das, D. M. Analyss of task assgnment polces n scalable dstrbuted Web-server systems. IEEE Transactons on Parallel and Dstrbuted Systems, 9(6, pp. 585-600, June 1998. [20] Cardelln, V., Colajann, M., and Yu, P. S. Request redrecton algorthms for dstrbuted Web systems. IEEE Transactons on Parallel and Dstrbuted Systems, 14(4, pp. 355-368, Aprl 2003. [21] Cardelln, V., Colajann, M., and Yu, P. S. Effcent state estmators for load control polces n scalable Web server clusters. In Proc. of the 22 nd Annual Internatonal Computer Software and Applcatons Conference, 1998. [22] Bubendorfer, K. and Thomson, W. Resource management usng untrusted auctoneers n a Grd economy. In Proc. of the 2 nd IEEE Internatonal Conference on escence and Grd Computng, Dec. 2006. [23] Ogston, E. and Vasslads, S. A peer-to-peer agent aucton. In Proc. of the 1 st Internatonal Jont Conference on Autonomous Agents and Mult-Agent Systems, 2002. [24] Crovella, M. E., and Bestavros, A. Self-smlarty n World Wde Web traffc: Evdence and possble causes. IEEE/ACM Transactons on Networkng, 5(6, pp. 835-846, 1997. [25] Crovella, M. E., Taqqu, M. S., and Bestavros, A. Heavy- Taled Probablty Dstrbutons n the World Wde Web. A Practcal Gude To Heavy Tals, Brkhauser Boston Inc., Cambrdge, MA, USA, pp. 3-26, 1998.