Lcast: Software-Defined Inter-Domain Multicast

Transcription

1 Last: Software-Defined Inter-Domain Multiast Florin Coras a,, Jordi Domingo-Pasual a, Fabio Maino b, Dino Farinai b, Albert Cabellos-Apariio a a Universitat Politènia de Catalunya, UPC BarelonaTECH, Barelona, Spain b Ciso Systems, San Jose, CA, USA Abstrat Traditionally, effiient inter-domain data delivery may be implemented either as a network or appliation layer multiast servie. However, while the former has seen little uptake due to prohibitive deployment osts the latter is widely used today, but often without a minimum guaranteed performane. In this paper we present Last, a network-layer single-soure multiast framework designed to merge the robustness and effiieny of IP multiast with the onfigurability and low deployment ost of appliation-layer overlays. The arhiteture involves no end-host hanges and only requires the upgrading of a small set of routers to support the Loator/ID Separation Protool (LISP), an inrementally deployable enhanement to the urrent global routing infrastruture. Content distribution over the Internet s ore is done by means of a router overlay while within domains, end-hosts interfae with Last using onventional multiast protools. The overlay s salability and topologial onfigurability is sustained by logially entralizing group management. We illustrate the versatility of our solution by designing and assessing the salability and performane of three management strategies for low lateny ontent distribution. Our analysis is based on large sale simulations supported by realisti user behavior and Internet-like network topologies. The results show Last s low management overhead and ability to optimize delivery to meet various operational onstraints. Notably, we find that it an deliver traffi with latenies lose to uniast ones, independent of overlay size. Keywords: inter-domain multiast; Loator/ID Separation Protool; software-defined networking 1. Introdution The Internet is gradually beoming the preferred infrastruture for delivering live ontent suh as sports events or news to large user sets. Aording to reent Corresponding author addresses: foras@a.up.edu (Florin Coras), jordi.domingo@a.up.edu (Jordi Domingo-Pasual), fmaino@iso.om (Fabio Maino), dino@iso.om (Dino Farinai), aabello@a.up.edu (Albert Cabellos-Apariio) Preprint submitted to Computer Networks September 11, 2013

2 reports, video streaming is among the largest and the fastest growing bandwidth onsumers [1] and IPTV driven revenues are to rise from less less than USD 9.7B in 2011 to USD 21.3B in 2017 [2]. For suh senarios, where one-to-many ontent delivery to large number of reeivers is required, IP multiast [3, 4] is perhaps the most effiient solution in terms of bandwidth onsumption. However, although often supported within the onfinements of ampus, enterprise or servie provider networks, IP multiast deployments have been generally done disregarding inter-domain onnetivity, thereby resulting in disonneted multiast islands [5, 6, 7]. One fundamental ause for the slowly advaning deployment is the requirement that all routers be upgraded to support the protool. But other, frequently ited, reasons regard management omplexity and the lak of a lear ommerial servie [8]. While the former inurs high operational expenditure the latter leads to situations when multiast implementation over links with uniast eonomial agreements result in loss of revenue. The low uptake of IP multiast has led over the last deade to the development of many appliation-layer multiast (ALM) solutions that build end-host overlays to ensure Internet wide ontent dissemination. They are designed to be very flexible in aommodating globally spread users and adaptable to hanging network onditions. However, the inongruene they introdue between overlay and underlying network topology diminishes their delivery effiieny with respet to that of IP multiast. Furthermore, performane analysis has unovered signifiant limitations of these arhitetures in saling user quality of experiene with the inrease of lient population [9]. Reasons often reported are unavailability of inter-peer bandwidth, hurn or, in some ases, insuffiient upload apaity. Sine they are related to either end-host behavior or upload abilities, these limitations are intrinsi to the overlay s design and, therefore, not avoidable through optimized overlay management. In this paper, we propose a network-layer single-soure multiast framework designed to merge the benefits of IP multiast and ALM while avoiding their respetive deployment and saling issues. Our goal is to enable large sale single-soure streaming by interonneting existing multiast apable domains devoid of end-host software upgrades and transparently for the greater part of existing multiast routers. This approah is omplementary to existing ALM solutions as it aims to offer overlay management ontrol to network operators in exhange for improved reliability and more effiient network resoure use but at the ost of minimal infrastruture support. To ahieve our goal, we exploit a unique window of opportunity offered by the development and deployment of the Loator/ID Separation Protool (LISP) [10], a protool meant to improve the urrent routing infrastruture, whih we use as support for our proposal. In light of the intrinsi dependeny on LISP, we refer to our solution as LISPasting or shortly, Last. Last reates and optimizes a LISP router overlay and transparently interfaes with end-hosts and legay multiast routers by means of existing IP multiast protools [3, 4, 11]. Group management funtions are logially entralized and performed by an overlay oordinator whereby members require no prior 2

3 onfiguration nor need to be manually managed. We stress this as a ruial property sine it irumvents the management omplexity issue that plagues traditional IP multiast deployment and further opens the possibility to dynamially optimize delivery with respet to overlay topology maps. This ability ould be exploited by ontent providers to define their own overlay oordinating algorithms or perform on-line swithing between multiple ones, aording to speifi operational requirements or eonomial agreements. In this sense, Last is akin to Software-Defined Networking (SDN) [12] and, in fat, the two share many of the properties that derive from the implementation of a programmable ontrol plane. From an arhitetural standpoint, Last s ability to aommodate large number of lients is ensured through design, by deoupling the ontrol and forwarding funtions within the overlay. As a result, eah an evolve, be optimized and sale aording to speifi needs. In partiular, data plane salability an be ahieved by onstraining router repliation fators, to avoid performane penalties due to uniast repliation ineffiieny, while ontrol plane salability may be ensured by limiting ommuniation overhead. In both ases the tradeoff is overlay effiieny. To assess ontrol plane salability, the impat of repliation fators on effiieny and overlay onfigurability, we evaluate Last s ability to deliver low lateny ontent in three distint operational setups. Our simulations make use of (i) an Internet-like autonomous system (AS) level topology and (ii) large lient traes that emulate realisti lient behavior, onsisting of 3k ASes and approximately 140k unique IPs obtained through a globally distributed apture of SopCast [13] overlays. First of all, the results show the ontrol plane s ability to sale. Even when ative topology disovery mehanisms are used and lient hurn is high, the load is manageable by a single server ating as overlay oordinator. Seond, repliation fators need not be large for effiient ontent delivery. Finally, the overlay an be easily optimized onsidering various operational onstraints. Notably, if inter-member lateny an be estimated, Last an deliver ontent at lose to uniast latenies, independent of the overlay s size. The rest of the paper is strutured as follows. We disuss the related work in Setion 2 and provide a short overview of LISP in Setion 3. Setion 4 desribes the Last framework and Setion 5 introdues an optimization algorithm and two ways to obtain topology maps that may be used to optimize the overlay. Setion 6 presents our evaluation methodology and in Setion 7 we disuss the results. Finally, we onlude the paper in Setion Related Work As a long standing aademi and ommerial researh hallenge, singlesoure live media streaming benefits from opious amounts of related literature. Consequently, we restrit the disussion to a limited set of solutions and mainly fous on those that bear similarities to Last. Multiast funtionality that enables live media streaming was originally offered as a network-layer servie. But in light of IP multiast s lak of inter- 3

4 domain deployment, network layer solutions have turned into arhitetures that leverage isolated multiast deployments. One notable example is MBone [14], a virtual network designed to onnet multiast islands by means of stati uniast tunnels. Although it was first to support distribution of ontent to users spread in multiple domains, it proved hard to extend sine setting up tunnels involved manual onfiguration. AMT [15] irumvents this limitation by providing mehanisms for automatizing the tunnel setup proess with the help of dediated servers (relays and gateways) plaed in soure and destination domains. However, AMT does not support dynami reonfiguration of the tunnels, i.e., of the inter-domain distribution topology. Therefore, unlike Last, it is not able to adapt to hanging network onditions, lient hurn or to limit repliation overhead. A large set of appliation layer solutions, inluding NICE [16], Narada [17], OMNI [18], ZIGZAG [19] and Sribe [20], have been proposed by aademia in the last deade. Out of them, OMNI [18] is the losest in spirit to our proposal. It requires servie providers to deploy a set of proxying nodes that self organize in an overlay and forward traffi to subsribed lients. The optimizing algorithm employed is a distributed instane of the one we use and the metri onsidered is lateny. In ontrast, Last works at network-layer and is supposed to be deployed, at no additional ost, together with LISP. Moreover, Last s logially entralized ontrol plane allows easy deployment of new optimization algorithms without requiring router hanges. Apart from the aademi solutions, a large array of ommerial ALM arhitetures like SopCast [13], PPLive [21], CoolStreaming [22] or UUSee [23] are widely used for Internet ontent streaming. Being losed soure their arhitetures are not ompletely understood, nevertheless their performane has often been the subjet [24, 25, 26, 27] of aademi srutiny. The results have shown signifiant limitations of these arhitetures in saling user quality of experiene with the inrease of lient population. Last, being an extension of LISP, operates on domain border routers and thus builds an overlay topology that is not diretly exposed to lient hurn. Furthermore, through design it avoids imposing bandwidth strain on overlay members and ould assure ertain performane bounds. Another approah to delivering inter-domain multiast is to onnet islands of multiast enabled end-hosts by means of appliation layer overlays. Two of the solutions to follow this design guideline are Universal Multiast [6] and Island Multiast [7]. They are similar to Last in their use of existing multiast deployments for intra-island ontent delivery and of tunnels to onnet multiast islands. However, they ensure inter-island multiast delivery by building and optimizing overlays onsisting of end-hosts. This gives rise to three fundamental differenes. First, Last does not require hanges to end-hosts as the interdomain router overlay seamlessly interfaes with loal multiast. Seond, Last should use more effiiently the underlying intra-domain network sine paket repliation is always performed in domain border routers and therefore pakets avoid traveling intra-domain prior to being repliated and forwarded to hosts in foreign domains. As a downside to this, when fan-out values are large, the 4

5 routers have a higher proessing overhead however, as shown by our experiments, they need not be large for good performane. Finally, all solutions offer the option to onstrain fan-out values but Last offers ontrol to operators who have a vested interest in the effiieny of the overlay. That is, router out degrees should be generally limited to protet routers from saturating their interfaes and to ensure fairness in distributing the repliation responsibilities. We then believe that providing the ability to onfigure fan-out atually makes Last better suited for operational deployments than its P2P ounterparts. We previously proposed CoreCast [28], a LISP inspired inter-domain streaming arhiteture where soure and lient routers operate aording to a lientserver model. Both CoreCast and Last are based on LISP protool mehanisms but they have quite different approahes to delivering inter-domain traffi. In this sense, CoreCast aims to diminish inter-domain bandwidth use when ompared to P2P live streaming systems while Last aims to provide a salable and easily reonfigurable offloading mehanism for the soure LISP router. Finally, it is worth mentioning that LISP-Multiast [29] integrates IP multiast funtionality into LISP. Naturally, it inherits all the properties of traditional network layer multiast however it also requires ore router support for inter-domain use, thus making it unfeasible for a wide-sale deployment. 3. LISP Bakground LISP [10] is an arhitetural solution to the Internet s salability problem, reently generated by the alarming growth of inter-domain routing tables [30]. To this end, LISP s main goal is to split the semantis of IP addresses with the aim of forming two namespaes that unambiguously identify ore (loators) and edge (identifiers) network objets. Its development is aided by a pilotnetwork [31] and a onsiderable ommunity spanning over 32 ountries, with members pertaining to both aademia and industry. Notably, the protool that has reently undergone IETF standardization [32]. One of LISP s defining features, that also distinguishes it from many of its ompetitors, is its support for an inremental deployment. It is transparent to end-hosts and requires the upgrading of only stub domain border routers. Moreover, to failitate transition from the urrent Internet infrastruture, both loator and identifier namespaes use the existing IP addressing sheme. Therefore, the split does not affet routing within existing stub or transit networks. Nevertheless, as identifiers and loators bear relevane only within their respetive namespaes, a form of onversion, from one to the other, has to be performed at border points between ore and edge networks. LISP enabled border routers make use of a tehnique alled map-and-enap [33] for the translation. Apart from the need for data plane modifiations, map-and-enap also requires the introdution of a new ontrol plane mapping funtion able to provide bindings that link identifiers to loators. Therefore, prior to forwarding a host generated paket (see Figure 1), a LISP router maps the destination address, or what LISP alls an endpoint identifier (EID), to one or more orresponding destination routing loators(s) (RLOC) by means of a mapping obtained from a 5

6 LISP Legay MR Mapping System MS 1 2 Internet RLOCA2 3 RLOCA1 4 RLOCB1 RLOCB2 xtr A EIDSRC EIDDST xtr B Figure 1: Example paket exhange between EID SRC and EID DST with LISP. Pakets travel from EID SRC to xt R A aording to intra-domain routing. xt R A obtains a mapping binding EID DST to RLOC B1 and RLOC B2 from xt R C through the mapping-system (steps 1-3). Then, xt R A hooses RLOC B1 as tunnel destination, enapsulates the pakets and forwards them over the Internet s ore to xt R B (step 4). Finally, xt R B deapsulates and forwards the pakets to EID DST. LISP speifi diretory servie, alled the mapping database system [34, 35]. A priority is assoiated to an RLOC to indiate the preferene of it being used, and a weight indiates how traffi is to be load balaned between RLOCs with the same priority. Mappings are requested on-demand, as opposed to being proatively obtained, and stored in a loal map-ahe suh that they may be reused. To retrieve a mapping, a LISP router direts a Map-Request message arrying the destination EID to a Map-Resolver (MR) part of the mapping system and in return reeives as answer a Map-Reply. One the mapping is obtained, the border router selets one of the EID s RLOCs as tunnel destination, enapsulates the paket with a LISP-UDP-IP header and forwards it to orresponding edge network. At the reeiving router, the paket is deapsulated and forwarded to its intended destination. For a detailed example see Figure 1. In LISP parlane, the soure router, that performs the enapsulation, is alled an Ingress Tunnel Router (ITR) whereas the one performing the deapsulation is named the Egress Tunnel Router (ETR). One that performs both funtions is referred to as an xtr. Additionally, LISP makes use of Re-enapsulating Tunnel Routers (RTRs), that perform re-enapsulation, i.e., deapsulation followed by enapsulation, to enable paket re-routing based on EID. It is also in the ETR s duty to register the EID address spae, one or multiple EID-prefixes it is responsible for, with its assoiated Map-Server (MS) by means of Map-Register messages. At their turn, Map-Servers aggregate and advertise the EID-prefixes in the mapping system to enable EID routing. Then, having this information, the mapping system an ensure the delivery of a Map-Request to the MS that originates the EID-prefix overing the request s destination ad- 6

7 dress. Depending on onfiguration, the MS an either answer or forward to an ETR the Map-Requests it reeives. 4. Last Arhiteture This setion presents our proposal for supporting inter-domain multiast streaming. We start by providing an overview of the arhiteture and then desribe in greater detail group management proedures and signaling Arhiteture Overview Last is a LISP extension that provides a single-soure multiast servie to lients in disjoint multiast islands by means of a router overlay. It ompensates for the lak of an inter-domain multiast infrastruture by performing uniast enapsulated, and if possible also multiast enapsulated, repliation of multiast traffi aross the Internet s ore. The resulting overlay interfaes with existing intra-domain IP multiast protools so it does not require any end-host software upgrades. All member domains must be LISP enabled and may partiipate in the overlay with at least one of their border routers (ETRs). On the data path, the soure domain s border router, an ITR, heads the distribution tree and is the first to perform enapsulated repliation. Subsequently, all downstream overlay members, save for the leaves, repliate the reeived pakets up to a ertain fan-out. Note that sine traffi is unidiretional, from multiast soure to lients, all routers but the soure ITR perform either only deapsulation or deapsulation and re-enapsulation. For brevity we refer to all of them as ETRs, although those that perform both funtions also implement RTR funtionality. See Figure 2 for a depition of an example Last data plane. An important drawbak to uniast enapsulated repliation is that it redues throughput proportionally to the repliation fator, if performed multiple times out the same interfae. As a result, inreasing fan-out an quikly saturate router interfaes and therefore not only deteriorate overlay performane but also ongest other flows sharing the same links. Additionally, sine paket repliation is performed sequentially, the time differene between the instane the first and last replias are forwarded may be onsiderable. So, besides inreasing the delay to obtaining the multiast pakets for diretly onneted downstream members, the resulting lateny may aumulate and distribute unevenly aross the hierarhy, randomly leading to branhes with low performane. Finally, apart from the limitations onerning performane, unbounded fan-out an also lead to unfair and/or eonomially unfeasible situations. Generally, Last and other island multiast solutions substantially redue inter-domain traffi exhange, if ompared to simple uniast delivery or unoptimized P2P overlays (see Setion 7). However, if the distribution tree is not arefully onstruted, member routers serving few lients might be requested to repliate a disproportionate number of times, against their interest and to the advantage of other eonomially benefited peers. To avoid these ineffiienies and in the interest of fairness, we request that Last members have a onstrained fan-out whereby the overlay 7

8 s soure lient Last IP multiast s Internet ITR ETR Figure 2: Example Last data-plane arhiteture. The ITR is the first to repliate the ontent and all downstream ETRs may repliate up to a fixed fan-out value. In the example, fan-out is onstrained to 2. must be organized as a degree-onstrained tree, despite the potential to redue distribution effiieny. Fan-out values ould be fixed for the whole overlay or reported at subsription by eah member. It should be noted that, similarly to end-host overlays, purely repliating routers, RTRs, ould be provisioned in transit domains to ensure improved overlay stability, performane and also onsiderably redue or remove altogether the repliation overhead of ETRs. However, suh a solution also implies a business model, different to the one disussed here, where third party entities manage the RTRs and engage in eonomial agreements with the soure and lient edge-domains. Sine we are not ready to model business relationships, or speulate how suh RTRs ould be deployed, in this paper we limit the analysis to overlays where repliation is performed exlusively by ETRs. For a tehnial disussion on how RTRs ould be onfigured to partiipate in the overlay, we refer the interested reader to an Internet-Draft [36] we published on the subjet. Another limitation to having routers partiipate in an overlay, is that they are generally ineffiient at handling omplex omputation tasks sine they are designed to perform fast forwarding of pakets as opposed to general purpose omputing. To irumvent this drawbak, Last leverages LISP s native separation between data and ontrol plane to ensure the logial entralization of group management funtions in an overlay oordinator. Thereby, at data plane level routers only perform enapsulated repliation while at ontrol plane level, the oordinator must ompute a distribution tree and ensure members are organized aording to it. Besides supporting our original goal of having no management osts for routers, this design also opens the possibility for enhaned overlay onfigurability. In this sense, if the oordinator obtains or is onfigured with a map of the loator underlay, it may proeed to optimize the distribution tree with 8

9 respet to a given metri. Moreover, the arhiteture allows the swithing between multiple optimization funtions or metris, even on-line, to meet hanging operational requirements. Note that we are not the first to propose suh split. This ideas has been previously reommended to aid routing salability [37] and is nowadays entral to urrent SDN researh [38]. Possible implementers of the overlay oordinator may be the soure ITR or the MS. Reusing the previous argument, sine the ITR is a router, we onsider the MS better fit for the funtion. In fat, due to its position in the LISP ontrol plane, the MS is required to proess and provide answers to all Map-Requests originated by routers willing to initiate ommuniation with the multiast soure. Then, as it an reognize and keep trak of all overlay members, the MS should also be the one to deide the attahment point for a joining member or the one to optimize the overlay. Although the design allows for the overlay state and/or management funtions to be distributed, in this artile we are interested in evaluating if the ontrol plane overhead is sustainable by one, off-the-shelf, server ating as MS. Last is ompatible with the urrent LISP speifiation, but apart from the anonial LISP messages we introdued in Setion 3, it additionally requires the signaling messages defined in [39] for onveying joining (Join-Request message) and leaving (Leave-Request message) multiast information. They are not Last speifi and have been designed to simplify the onneting of multiast apable sites with LISP-Multiast Member Subsription In Last enabled domains, end-hosts request single-soured multiast ontent, the way they do with traditional IP multiast, namely, by subsribing to a multiast stream using the Internet Group Management Protool (IGMP) [40, 4]. They learn the hannel identifier (S-EID,G), onsisting of the multiast soure address S-EID and a multiast group address G, used to distinguish between the multiple hannels originated by a soure, either through onfiguration or with the help of appliation-layer protools. Exept in the partiular ase when S-EID is part of the same domain, and therefore the ontent may be delivered loally without Last, the subsription request propagates intra-domain up to one of the domain s border routers, an ETR. If the ETR is already a member of the multiast hannel, it starts repliating the multiast ontent towards the requesting end-host and no further ation is taken. If not, the ETR initiates a two step overlay join proedure whereby it first attahes to the Last overlay serving (S-EID,G) and seondly it advertises its ability to repliate multiast ontent. To omplete the first step, the ETR must initially obtain the loator, of at least one of the already onneted routers, to be used as overlay attahment point. It ahieves this by requesting that the hannel identifier be mapped to the loators of potential overlay parents, in essene, by sending a Map-Request for (S-EID,G). The request propagates through the mapping system up to the oordinating MS, whih ensuing the request s reeipt, starts a searh for overlay 9

10 members with spare apaity. The searh may be done randomly or, if additional topologial information exists, in aordane to a predefined heuristi that ould ensure that an optimal attahment point is hosen. One obtained, the result, onsisting of a list of one or multiple RLOCs pertaining to the overlay members able to aommodate new hildren, is onveyed to the joining ETR in a Map-Reply. Typially, the MS will offer an ETR the possibility of hoosing its upstream either when not optimizing the overlay or when all the hoies have an equal ost in the distribution tree. Using loal poliy (e.g., shortest AS path) and the priority and weight values assoiated to the list entries, the ETR hoses the best RLOC and sends it a Join-Request message to request the setting up of an overlay branh between the two. The parent router appends the RLOC of the joining ETR to the list towards whih it performs uniast enapsulated repliation, therefore onluding the ETR s attahment. Alternatively, if the two routers an be onneted by inter-domain multiast, the joining ETR first performs a protool dependent multiast join to the parent in the underlying inter-domain network. Afterwards, it indiates the multiast hannel identifier, to be used as destination for the multiast enapsulated pakets (as opposed to uniast enapsulated), in the Join-Request. A speial ase arises when an ETR is first to join the overlay. In this situation, the ETR requests the multiast ontent from the ITR, but it may happen that the ITR is not yet subsribed to (S-EID,G). Therefore, on reeipt of the Join-Request, the ITR must first subsribe to the multiast soure, using IGMP or a PIM Join [41] message, to obtain the streamed ontent to be repliated towards the joining ETR. The seond step in a member s subsription proedure is to signal that it an perform repliation within the overlay. To this end, one the ETR is attahed, it starts registering (S-EID,G) with the MS. The Map-Register message onveys the ETR s RLOC, that of the hosen parent and the number of intra-domain lients it serves at a ertain time instant. To be noted that the estimating of the membership in a multiast session, although traditionally a diffiult task, an be ahieved within a domain using the expliit traking apabilities of both IGPM and PIM. Then, having for a hannel identifier the registration messages of all the overlay members, allows the MS to build an aggregated (S-EID,G) mapping that provides a omplete view of the overlay s topology, i.e., of how the members organize in a distribution tree. This mapping, together with underlay topologial information, if any exists, is used by the MS to answer Map-Requests of joining ETRs and to optimize the distribution tree. Two additional benefits of the proedure, sine registrations are refreshed periodially, are that the MS impliitly detets the failure of an ETR and also beomes aware of the hanges in lient population within a domain. The way the distribution tree is built has two advantages. First, it ensures that multiast pakets in the soure domain do not reah the ITR if no ETR is joined and the ITR does not partiipate in a loal multiast group. Seond, pakets are forwarded from ITR to all ETRs without mapping database lookups thus, with minimum overhead. 10

11 4.3. Member Failure and Unsubsription If a member loses network onnetivity, its data path hildren will sense the failure either as a lak of multiast pakets or by means of a LISP speifi mehanism, alled RLOC-Probing. This proedure, used by xtrs, onsists in the use of Map-Request messages to determine reahability of peer xtrs and to estimate round-trip times (RTT). One the hildren detet the failure they look for new overlay parents by either sending Join-Requests to other RLOCs in the mapping assoiated to (S-EID,G) or, if no other exists, by redoing the whole subsription proedure. Still, suh irumstanes will result in paket loss for all members of the subtree headed by the affeted router and out of band mehanisms would be required for remedying the failure. Suh mehanisms are out of the sope of the urrent paper. However, sudden loss of network onnetivity for a domain s border router should be a seldom ourrene. An ETR initiates its unsubsription from the Last overlay when the last of its lients leaves the intra-domain multiast group. First, if the ETR repliates ontent to other overlay members, it inrements the priority of the RLOC registered with the MS to the least preferable value and replies to RLOC-Probing messages indiating that its RLOC is unreahable. The update fores the MS to avoid assigning the ETR new hildren and fores the existing ones to find new overlay parents. The lookup proedure is idential to the one followed in the event of a parent failure however, in this ase there are no paket losses. Alternatively, when the MS senses the departure of an ETR, it ould proeed to optimizing the whole affeted subtree to avoid performane degradation. As a seond step, the ETR sends a Leave-Request to its overlay parent and stops registering (S-EID,G) with the MS, onluding the unsubsription Distribution Tree (Re)Configuration The position of an overlay member on the data-path is established at subsription time, however the MS ould be onfigured to further optimize the distribution tree, if provided with information about the overlay s topology. In this ase, distribution effiieny is ontrolled by the MS through optimal plaement of joining ETRs and/or through periodi or enfored tree reshaping. When reorganizing the distribution tree, the MS informs members of their new positions through updated mappings. To avoid paket loss and to assure a seamless transition, members use a make before break proedure when swithing parents. Hene, prior to sending Leave-Request to their overlay parents, members first attah to those indiated in the updated mappings. If any dupliate pakets arise, they should be disarded by end-hosts at appliation layer. This type of entralized management enables the easy ustomization of the distribution tree as routers are oblivious to optimization algorithm hanges. In fat, a key feature of Last, is that an operator in ontrol of the MS an perform on-line swithing between multiple optimization algorithms or topology disovery protools, if more are supported, to better adapt the overlay to operational onstraints. In the next setion, we provide as example a possible tree optimizing algorithm and ways of obtaining topologial information that ould be implemented in Last. 11

12 5. Overlay Optimization The onfiguration of the Last overlay ontroller is appliation and operator speifi. To illustrate this point, in this setion we onsider the lassial problem of delivering lateny onstrained ontent (e.g., live streaming of news and sports events) and show how Last ould be used to solve it. We first propose an algorithm to ompute the distribution tree and afterwards point out how ITR-loal BGP routing tables and organized inter-domain lateny measurements ould be used to approximate overlay topology. For brevity, we further refer to the ombination of an optimization algorithm and a topology disovery mehanism as an optimization strategy Distribution Tree Optimization Algorithm In what follows, we use the term distane when referring to a relative length or amplitude of a metri, observed on a path onneting two points, but when the exat nature of the metri is of no interest. Considering our goal of delivering ontent for delay sensitive appliations, the funtion we minimize in our experiments is the maximum distane (e.g. lateny or number of AS hops) from a lient to the multiast soure. Notie that the referene is the end-host and not the domain border router (ETR). Thus, what matters in deiding an ETR s position in the overlay tree is not solely its distane to the ITR but also the number of lients it serves. Then, a router lose to the soure but serving few lients might find itself lower in the hierarhy than another with a slightly higher distane but with a larger lient set. In other words, the algorithm tries to improve average end-host quality of experiene by optimizing the router overlay onsidering two dimensions, inter-router distane and the size of the lient set served by a router. This also ensures the algorithm is fair to members. Domains with fewer lients are more likely to beome leaves while those with larger user sets, the ones that benefit most from Last, are required to ontribute by repliating. The problem desribed above, heneforth named minimum average distane, degree-bounded spanning tree (MADDBST), may be formally stated the following way: Definition 1. Given an undireted omplete graph G=(V,E), a designated vertex r V, a degree bound d(v) d max, v V, d max N, a vertex weight funtion (v) N and an edge weight funtion w(e) R +, edge e E. Let Pr,v T be the set of edges e on the path from vertex r to v in the graph s spanning tree T. Also, let Wr,v T = e P w(e) represent the ost of the path linking r and r,v T v in the spanning tree T. Find the spanning tree T of G, routed at r, satisfying d T (v) d max, suh that v V,v r (v)w r,v T is minimized. We note that [42] and [18] have previously defined and solved similar optimization problems. Shi et al. [42] also proved that a partiular instane of the problem, where all verties have weight 1, is NP-omplete for degree onstraints 2 d max V 1. Similarly to our approah, they were interested in 12

13 a entralized solution whereas Banerjee et al. [18] have suessfully managed to distribute the algorithm. The heuristi we used to solve the MADDBST problem is similar to the one used by Banerjee and it is a variant of the one proposed by Shi. In short, the algorithm works by inrementally growing a tree started at the root node r until it beomes a spanning tree. For eah node v, not yet a tree member, it selets a potential parent node u in the tree T, suh that the metri δ(v) = (W T r,u + w(u, v))/(v), i.e., the distane to the soure per lient, is minimized. At eah step, the node with the smallest metri value is added to the tree and the parent seletion is redone BGP-based Topology Map One of the best soures of topologial information that is not or an not be ommonly used by appliation layer overlays is the BGP routing table. The BGP information an AS router holds attempts to present an Internet wide interonnetion map. But, due to the algorithm s distributed nature and its use of poliy, both inauraies and inomplete information may exist. The ITR has two options for obtaining BGP topologial information. First, it may aggregate partial BGP feeds from multiple overlay members (global view) or seond, it may itself onnet to BGP (loal view). The former ould ensure a more detailed desription of the topology, and thus grounds for better deisions, while the latter a more restrited, partial view of the interonnetion map and seemingly worse performane. Another aspet to be onsidered regarding the global view is that an off-line obtained BGP map may be rendered inadequate due to hurn whereas one obtained through on-line aggregation of multiple BGP tables may be a tehnially hallenging task. Even more so as some domains may be relutant to provide suh information whih they often deem as sensitive. By ontrast, the loal, on-line topology gathering mehanism requires nothing more than BGP feeds from the ITR. Additionally, there is no need for a ommuniation protool between the MS and the overlay members for the onveying of BGP reahability information. To ompare the two alternatives, we take as global view the Internet-like topology we use in our evaluation and as loal view the routing table of the ITR. More details on how we obtained the dataset an be found in Setion 6.2. Using these two topologies we omputed the relative AS path length inrease of the loal view and the distribution of the path lengths for both. Results are depited in Figure 3. If we fous on Figure 3a, we see that 99% of the loal view paths are at most 2 hops longer than in the global view and about 20% have an idential length. On average, path length in the loal view inreases only about 1.1 hops. This is also illustrated in Figure 3b where we an also note that, save for the average 1 hop inrease, the distributions of hop lengths are similar. Given the relatively small differene, we are lead to onlude that the loal view presents a reasonably aurate desription of the topology. Due to its relatively good auray and, more importantly, due to the implementation simpliity, we opted in our experiments for the BGP topology disovery mehanism based on loal information. The metri it provides, inter 13

14 Cumulative Probability Cumulative Probability global view loal view relative path length inrease (AS hops) path length (AS hops) (a) Cumulative probability of the relative path(b) Cumulative probability of the AS path length inrease due to loal view lengths Figure 3: Comparison of BGP loal and global views AS hops, in ombination with the optimization algorithm results in a degreeonstrained shortest AS path tree optimization strategy. For brevity, we shall refer to it as bgp. Should there be interest in obtaining a minimum AS hop ost tree, at the expense of larger number of hops to the soure, a degree-onstrained minimum spanning tree heuristi should be used Lateny-based Topology Map Inter-member lateny is a metri ommonly employed by appliation layer overlays in topology optimizations. Yet, obtaining an inter-member lateny map may sale poorly with the population size and therefore its implementation may be both expensive and tehnially hallenging. For instane, in a topology onsisting of N members, a naive approah, whereby eah member measures all possible peers, would require N-1 measurements per member. To prevent saling the number of measurements with the size of potentially large overlays, a more intelligent approah for the seletion of link latenies worth estimating is needed. We avoid performing a large number of measurements and assure they are arried out in an optimized order by exploiting a mehanism similar to the one used by Banerjee et al. for the group management of NICE [16]. The solution onsists in lustering nodes that are lose to one another in terms of lateny and limiting the inter-member measurements to just the pairs finding themselves in lose proximity. The amortized ost analysis shows that the number of ontrol plane peers (i.e., the number of peers measured) at an average member is onstant O(k) while in the worst ase it an reah O(k log(n)). Where, k is a onstant limiting the node degree and the size of the luster and N is the number of overlay members. Even in the worst ase, given that N may be in the range of thousands to tens of thousands, this is a onsiderable derease from O(N). 14

15 Another advantage of the entralized group management is that the lateny disovery mehanism, when implemented in Last, has a lower per member ommuniation overhead than in NICE, as members do not partiipate in a separate ontrol plane protool. However, LISP s extension to provide for a simple mehanism to onvey lateny measurements between ETRs and the MS is required. Sine ETRs hek the liveness of the loators assoiated to ahed mappings with RLOC-probing, the extension requires just the implementation of a message, similar to a Map-Reply for reporting the RTT estimates. The ombination of the lateny topology disovery protool and the optimization algorithm results in an optimization strategy we further refer to as lat. 6. Evaluation Methodology To ompare the performane of the overlay optimization strategies proposed in the previous setion, we implemented an event-based simulator. In what follows we desribe the simulator s omponents and our evaluation methodology. We start by desribing the datasets and the proedure followed to build an Internet-like inter-domain topology. Subsequently, we present the methodology used to generate traes that emulate realisti lient behavior and explain our simulation setup. We onlude the setion with a brief presentation of the metris used to evaluate overlay performane Simulation Methodology Our experimental evaluation simulates a set of 140k end-hosts spread in 3k autonomous systems, wathing a live stream over an Internet-like topology with the help of Last. For this purpose, we developed a event-base simulator apable of handling large sale Last overlays and several optimization strategies. This resulted in the partial implementation of Map-Server and ETR funtionality. In all the experiments we employ as ontent soure an arbitrary autonomous system as we observed that the hoie does not influene the results. Client ASes partiipate in the overlay with one ETR and their deision to subsribe or unsubsribe is triggered by the ativity of intra-domain users they serve. To simulate various types of user behavior, the latter is provided as input to the simulator in the form of trae files that log end-host join and leave events. ETR subsriptions are not optimized, but done at the first randomly found free position in the distribution tree not to bias the effet of the optimization strategies and are always based on uniast onnetions. The distribution tree is optimized by the MS periodially (10 min) or if more than a third of the members sustain an inrease of the served lient set above 10 or drop to 1, join or leave the overlay. These values were hosen to balane the omputation osts and the overlay s ontent delivery effiieny. Additionally, to evaluate the influene of tree optimizations on ommuniation overhead, we require that all member departures trigger the optimization of the affeted sub-trees instead of only having the affeted hildren reperform the subsription proedure. We 15

16 detail the optimization algorithm, the Internet-like topology and the three traes that desribe user behavior in the next setions. The performane of eah optimization strategy is evaluated by running simulations with respet to the lient trae and fan-out values, whih we vary between 2 and 10 to understand how repliation fators influene performane. For eah suh simulation run, we sample and store for analysis overlay state one per minute and ontrol traffi overhead one per seond. Finally, to better gauge the performane of bgp and lat, we also define and evaluate a very simple overlay management strategy that does not perform topology disovery or tree optimizations. In this senario, we further refer to as rnd, members join at random positions in the distribution tree and member departures always require the affeted hildren to repeat the subsription proedure Internet Inter-Domain Topology To obtain a realisti global inter-domain topology we aggregated datasets that estimate how autonomous systems interonnet from multiple soures: iplane [43], RouteViews [44], CAIDA [45] and RIPE [46]. All the data used is from April The dataset laks link speifi BGP poliy information that ould transform part of the AS graph s edges in ars (direted links). Most affeted by this assumption are the links between ustomers and their upstream providers and peering links between stub ASes. The first type may not be used by upstream providers for transiting traffi to destinations other than those found in their lients network. The seond type may not be used to transit traffi to destinations outside a peer s network. Sine, Last only repliates traffi between stub domain border routers, the two types of links may be misused only when a non-member stub domain transits traffi to or from an Last member. However, stub domains generally have a muh less diverse onnetivity than transit domains thereby, suh situations should be a seldom ourrene. For the resulting inter-as topology, we observed that the log-log plot for the omplementary umulative distribution funtion (CCDF) of the AS-node degree follows a straight line, a property found in power law distributions. Aordingly, as previously shown in [47] and [48], the Internet AS topology is a sale-free network with power law node degree distribution. Further, the average path length in our topology is 3.5 or 5.4% lower than the 3.7 observed [49] in the Internet. These two results orroborate our laim that the aggregate topology has properties similar to those of Internet s AS graph. For estimating inter-as lateny, we made use of iplane s [43] proven lateny predition abilities for IP pairs [50]. Beause we needed to estimate the lateny between domain border routers we had to elet for all partiipating ASes a representant. We did so by using iplane s estimations that assoiate points of presene (PoP) to ASes and their inter-onnetion map. For any domain, the PoP with the largest degree was eleted as the representant. In about 30% of the ases, when iplane failed to provide an answer, we used a lateny estimator based on geographial distane desribed in [51]. 16

17 6.3. The Client Traes To ensure a thorough evaluation of the optimization strategies, we make use of lient traes that emulate omplementary types of user behavior. The domains that partiipate in the overlay and their respetive number of lients were obtained from a passive distributed apture of several P2P TV hannels whereas the lient hurn was modeled in aordane to reent results in the field. We detail both efforts in what follows. SopCast [13] is one of the P2P TV appliations frequently used for streaming of live sports events. Wanting to model lient distribution for large events of global, or at least wide-spread interest, we aptured the traffi pertaining to several SopCast overlays during an 2011 UEFA Champions League semifinal. To this end, we used 2 vantage points in USA, 5 in Europe and 2 in Asia, spanning a total of 6 ountries. We were interested in understanding how lients luster in autonomous systems, not in the speifi performane of a hannel s overlay. Thus, depending on the upload apabilities of eah vantage point, we joined a number of P2P hannels, streaming the same event, at eah node. As a result, the traes finally ontained more than 145k unique IPs spread in over 3.8k ASes. Out of them, in our simulations we used 3k ASes for whih we ould ompute pairwise lateny estimates. More information about the traes we aptured and their properties an be found in [52]. In spite of the large size of our aptured dataset, lak of logs from the overlay s bootstrapping server made it impossible to approximate lient lifetime in the overlay. We thus resorted to syntheti modeling of lient hurn. As shown by several studies [53, 25, 24, 27, 26], it is generally aeptedthat lient arrival proess, at least for periods spanning dozens of minutes, an be modeled by a Poisson proess. Furthermore, Sripanidkulhai et al. observed in [53], after analyzing 3 months worth of Akamai logs, that short duration events, whih last a ouple of hours, present flash rowds whereas non-stop streams have a time of day behavior. These findings were onfirmed by Veloso in [27] who also noted that for long streams lient inter-arrivals an be modeled through a Pareto or a pieewise stationary Poisson proess. For lient session lengths however, onsensus ould not be found. Thus, depending on stream length or the type of system being analyzed by either paper, they may follow different distributions. Still, with the exeption of [25], there seems to be an agreement that sessions should have lengths distributed aording to a power law but opinions diverge when assessing the weight of the tail. Considering the works disussed above, in order to perform an evaluation of our proposed arhiteture that aknowledges the wide range of lient behavior, we generated 3 traes with omplementary properties. The goal was to model a short event, spanning 2h 30min, with a piee-wise Poisson arrival proess but with different shapes for the session length distributions. In order to apture the flash rowd effet we required that 80% of the lients join during the first 30min, and the rest spread over the time left. For the session lengths we used a Pareto distribution with a shape parameter of 1.5 and a sale parameter, 17