A Hierarchical P2PSIP Architecture to support Skype-like services

Transcription

1 A Hierarchical P2PSIP Architecture to support Skype-like services Isaias Martinez-Yelmo, Carmen Guerrero, Ruben Cuevas Departamento e Ingeniería Telemática Universia Carlos III e Mari Av. Universia Leganés. Mari Spain) {imyelmo, guerrero, rcuevas}@it.uc3m.es Anreas Mauthe Computing Department, InfoLab 21 Lancaster University Lancaster LA1 4WA U) anreas@comp.lancs.ac.uk Abstract A Hierarchical DHT Overlay Architecture base on P2PSIP is propose to support a Skype-like service. The IETF P2PSIP Working Group is stanarising a protocol to support any DHT in orer to eploy services insie a omain. We exten its functionality to allow the interaction between peers of ifferent omains. Furthermore, we perform an analysis of the Routing Performance an Resource Consumption uner a Skype-like scenario where VoIP calls are more likely to happen among users of the same omain. 1. Introuction Nowaays, one of the most wiely researche topics is the provision of Multimeia services on the Internet. However, ue to their requirements, currently there is a lack of evelopment effort. Although, some services are successful, Skype is a goo example), they are not easy to esign an evelop. Thus, aitional research an tools are neee in orer to eploy easily Multimeia ae value services. In this paper, we examine the benefits of aopting a Hierarchical Overlay Network to evelop a Skype-like service. Skype is a wiely use application, but it is a proprietary solution an a stanarise solution is preferre. Some etails about Skype can be foun in [6] or [16]. In orer to obtain a low cost Skype-like solution, the IETF P2PSIP Working Group has been create to work on a similar but stanarise alternative. The iea is to provie a ecentralise architecture that replaces the functionalities of proxy an registrar SIP servers with a peer-to-peer overlay network [2]. The result of this work is the P2PSIP protocol [7] that is being efine to support any kin of DHT overlay network such as Content Aressable Network CAN) [14] or aemlia [11] where the esire information can Figure 1. Hierarchical Overlay Architecture be store in a ecentralise fashion. Furthermore, the connections woul be NAT Traversal capable using the ICE protocol [15] base on STUN an TURN relays publishe as services in the P2P network. However, the connectivity is only possible between peers of the same P2PSIP omain.however, our proposal enables the connectivity between ifferent omains. This proposal is shown in Fig.1 where ifferent P2PSIP omains eploy their own overlay an global connectivity between the ifferent omains is obtaine with an Interconnection Overlay in which each omain is represente by at least one super-peer. This approach has a number of avantages: first of all, VoIP calls usually are more frequent among peers that belong to the same omain. Several examples inclue intraomain VoIP calls insie the heaquarters of a private company or VoIP calls an Instant Messaging between people subscribe to some Community/Social Network. Secon, a hierarchical architecture allows fault isolation since if a network failure happens in the gateway router of a company supporting P2PSIP for its VoIP calls, it oes not affect other P2PSIP omains. The Interconnection Overlay is only slightly affecte since only this super-peer is not reachable. In aition,

2 the avantage of a better scalability is obtaine, which is a known fact in hierarchical architectures [18]. Other well known avantages are explaine in literature [1]. This paper aresses how a Hierarchical DHT Overlay Network can be use to interconnect ifferent P2PSIP omains an it stuies analytically the Routing Performance in terms of hops to reach the estination. The rest of the paper is structure as follows. Section 2 presents the propose hierarchical DHT architecture an how to apply the propose architecture to provie a global P2P VoIP service. The Routing Performance of this architecture consiering the intra-omain hit probability is etaile in section 3. The efficiency of this architecture is explore in the case of a Hierarchical CAN Overlay Network in section 4. Finally, in section 5 the relate work associate with Hierarchical Overlay Networks is presente an section 6 presents the conclusions. 2. Hierarchical DHT Overlay Architecture The propose hierarchical architecture is base on creating two ifferent levels: i) the P2PSIP omains an ii) the Interconnection Overlay. If a peer wants to retrieve information from a ifferent omain, it must route its query to the super-peers, which maintain the reachability between the ifferent omains through the Interconnection Overlay. Thus, any omain is reache through the Interconnection Overlay. Prefix ID n-bits) Suffix IDm-bits) Figure 2. Hierarchical ID The IDs that supports this architecture are compose by two fiels a Prefix an a Suffix as it is illustrate in Fig.2. Peers insie a omain use the Suffix ID for the omain overlay network. On the other han, the super-peers route the queries along the Interconnection Overlay using the Prefix ID. Thus, if a peer nees to fin a peer or item with a ifferent Prefix ID, the query is sent to its super-peer. This super-peer routes the query to the super-peer that manages the omain in which the estination peer is place an finally the query is propagate in the target omain. Peers obtain several benefits with this approach. The maintenance state is reuce in comparison with the flat counterpart, a peer only has to take care about the peers of its own omain. If a flat overlay network is use to support all the peers of the ifferent omains, the routing tables of all the peers woul be larger. With the hierarchical approach, peers have to maintain in average an overlay of N/ peers where is the number of omains), whereas in the flat overlay the number of peers is N. In fact, the reuction of routing entries with this architecture in the peers is approximately log B log B N 100 % if the routing size has a logarithmic epenency with the number of peers which is quite common). This reuction not only implies less memory an CPU consumption, but the banwith neee for the maintenance of the overlay routing entries is also reuce. This is an avantage in comparison with other proposals like [4], [19]. The rawback is the overloa in the super-peers [1], especially in the case of banwith an CPU consumption. In this paper, we assume that a mechanism exists to select suitable super-peers [12], [13] Interconnection of P2PSIP omains One of the main problems in eveloping a ecentralise architecture the mapping between the available information an the peers in the system. If we consier a VoIP environment, the most usual way of referencing a buy is using URIs e.g. user@example.com). Associate with each URI, a Resource contains the user location information. An user URI can be ivie in two fiels: user ID user) an omain ID example.com), which are separate Thus, this separation can be use to map user URIs to a Hierarchical DHT Overlay Network. The Prefix ID or omain ID is create with the hash of the omain ID: Prefix ID = hashexample.com). The Suffix ID or peer ID is generate with the hash of the URI: Suffix ID = hashuser@example.com). Once the mapping is performe, a tuple containing the esire contact information an the generate key is store on the overlay. This operation is equivalent to the registration in the registrar SIP server in a SIP environment. At this point, the user information is available for the rest of the users. The worst case scenario for retrieving this information occurs when the target buy is locate in an other omain. In this case, the Prefix ID is generate from the omain ID an the Suffix ID from the complete URI. Once this actions are performe, a request is sent to the super-peer of the own omain. Then, the super-peer is able to route the query to the super-peer of the target omain using the Prefix ID. Once the super-peer of target omain receives the query, it forwars the query insie the omain using the Suffix ID. If the contact information is available, it is immeiately sent back an the establishment of a Multimeia Session starts.

3 Figure 3. Hierarchical P2PSIP Signalling 2.2. Signalling Exchange An example of the signalling on the propose hierarchical scenario is illustrate in Fig.3, consiering the actual status of the P2PSIP protocol [7]. We must take into account several things in orer to unerstan the signalling flow. First of all, when the peer in omain.b requests the information of the query in the Fetch message is in plain text. This is because a peer in a omain oes not have to know what hash function is being use on the interconnection overlay an what hash function woul be use in the other P2PSIP omain, which can have any length an also gives the Prefix ID length. Thus, the super-peer in omain.b performs hashomain.a) in orer to obtain the information of the super-peers in omain.a through the interconnection overlay. Insie this information, the hash use on the other omain is inclue hash a ), an therefore a request for the esire item is correctly forme with the correct length as hash a user1@omain.a). Thus, the esire interoperability is obtaine. Finally, the peers taking care of the esire Resource-ID answers to the super-peer on omain.a, which forwars this information to superpeer in omain.b. Super-peer in omain.b sen the esire Resource-ID to the peer in omain.b. Once this flow finishes, a legacy SIP negotiation is initiate for a VoIP call. We must highlight that Fig.3 represents only a subset of the real flow where the intermeiate hops in each overlay are not shown. 3. Routing Performance in a Hierarchical Overlay Network for a Skype-like service In this section we stuy the Routing Performance of the architecture propose. We exten the methoology efine in [10] with a more etaile explanation an moel as well as a complete explanation of the consiere assumptions. The parameters in the analytical moel are efine as follows: : number of P2PSIP omains. M i : number of peers in the P2PSIP omain i. N: all peers in the ifferent P2PSIP omains. It is consiere that a peer cannot be attache to multiple omains, hence N = i=1 M i. ρ ij : probability of launching a query from the P2PSIP omain i to the P2PSIP omain j. Cx): number of hops neee to fin a super-peer on the interconnection overlay epening on the number of super-peers x. It epens on the type of overlay use in the interconnection overlay. D i x): number of hops neee to fin a peer in P2PSIP omain i as function of the number of peers x belonging to the P2PSIP omain. We assume that all the peers in a P2PSIP omain know their super-peer attache in the interconnection overlay. This assumption implies that only one hop is neee to reach the super-peer. Taking into account the above efinitions, we obtain the Routing Performance R.P.) of this DHT-base hierarchical overlay network. Let us efine the cost of fining a peer on each overlay: D i M i ): cost of fining a peer on its own omain. C ): cost of fining a super-peer on the interconnection overlay. If the probability of obtaining an item in a P2PSIP omain from its super-peer is consiere negligible this can be assume because the number of superpeers is N/ an N ), the average Routing Performance for a peer in omain i can be written as follows: RP i = ρ ii D i M i )+ j=1,j i ρ ij [1 + D j M j ) + C )] 1) The first term in the previous expression is the cost of searching within the P2PSIP omain of a peer, whereas the secon term is the cost for searches in other omains. Thus, the average number of hops is expresse as: RP = 1 N M i RP i 2) i=1 However, some simplifications can be one in Eq.2 if the number of peers is the same in all omains: RP = 1 RP i 3) i=1

4 an D x) = l x 1 l, so if we use them on Eq.6, we have that the Routing Performance R.P.) is: ) R.P. = l N/) 1 l + 1 ρ ii ) 1 + i 1 i 7) Figure 4. Hierarchical CAN network Since we consiere an equal number of peers in all P2PSIP omains an each look-up in the overlay is ranomly inepenent, the probability of looking for a peer attache to other omain can be equally istribute between all the foreign omains.however, we may still have situations where the probability of looking for a peer in the own P2PSIP omain is ifferent from the probability of looking for a peer in other omains, in which case Eq.1 is expresse as follows: RP i = ρ ii D i M)+ 1 ρ ii [1 + D j M) + C )] 1 j=1,j i 4) This relation is useful for some type of scenarios. For instance, it can be use in a VoIP scenario where the probability of performing a call within the same omain is higher than a call to an external omain. Finally, if the same overlay is use on all the P2PSIP omains the sum can be eliminate from Eq.4: RP i = ρ ii D M) + 1 ρ ii ) [1 + D M) + C )] 5) From Eq.3 an Eq.4 we obtain the next equality: RP = RP i = D M) + 1 ρ ii ) [1 + C )] 6) We efine ρ ii as the intra-omain hit probability an it efines the probability of establishing a connection insie the own omain. 4. Hierarchical CAN Network In this section we stuy the performance of a hierarchical CAN overlay with two levels, this is shown in Fig.4. Accoring to [14], we have that C x) = i x 1 i Eq.7 represents the Routing Performance for a hierarchical CAN overlay where l is the number of imensions for the overlays in each omain an i is the number of imensions for the interconnection overlay. Each value can be optimise inepenently consiering the number of peers at each level. Thus, it can be state that l = lnn/) an i = ln these values can be obtaine minimising the Routing Performance of the flat CAN overlay network [14]). However, if the evelopment of a hierarchical CAN base application is consiere, it woul be reasonable to use the same value in both levels of the hierarchy, since just one version of CAN nees to be implemente an the coe can be easily reuse. Eq.8 presents the Routing Performance when only one value of is use in both layers: ) R.P. = N/) ρ ii ) ) The stuy that can be applie to this scenario is to optimise the R.P. accoring to the number of imensions when the number of P2PSIP omains an the number of peers per omain is given. This stuy makes sense since P2PSIP omains are inepenent entities that cannot be merge or splitte, an therefore an optimisation in relation with is not possible an only can be use to optimise the R.P. Theorem 1. The value of that minimises the number of hops on a hierarchical CAN overlay network with N peers an omains is unique. Proof: First we prove the existence of a minimum by eriving Eq.8 with respect to : f ) = ) 1 N 1 ln + 1 ρ ii ) 1 N 1 ln 1 ) 1 ) + Taking into account that 1, ), then: f = 1) = N an ) 9) 1 ln N ) + 1 ρ ii ) 1 ln) 10)

5 lim f ) = 2 ρ ii 11) N Thus, if, > e, then f = 1) < 0 an f = ) > 0 an consiering that f is a continuous function, we can conclue that one or more solutions exist for f ) = 0 in our range of work. Now we prove the unicity of the solution. Therefore, the secon erivate with respect to is calculate: f ) = 1 ) 1 N ln N ) 1 ) ) + 1 ρ ii ) 1 ) 2 ln 1 Taking into account again that f ) > 0, 1, ), then f ) is a monotonic increasing function an there is only one solution of f ) = 0 an this value is a minimum. The result of the theorem is interesting because it assures an optimal value for the number of imensions in a hierarchical CAN overlay network. Thus, let us explore if some result can be obtaine. We apply the following variable substitution: A = N/) 1, B = 1 13) If the above substitution is applie in Eq.9, we have: f ) = A 1 lna) + B 1 lnb) 14) Furthermore, if A an B are transforme to fin a relation between them, we can obtain: lnb = ln lna 15) lnn/) ln Suppose x =, we obtain that that B = lnn/) A x. With this result, we can rewrite Eq.14 as: f ) = A 1 lna) + 1 ρ ii )A x 1 lna x ) 16) The bisection metho can be use to solve f ) = 0. The solution gives the optimum value for accoring to the number of P2PSIP omains an the number N of peers. One scenario where these parameters are given coul be VoIP, is the number of groups that must be attache in the interconnection overlay in orer to be reachable between them an N is the overall number of peers. In orer to unerstan the importance of the number of imensions on a Hierarchical CAN Overlay : number of imensions Optimum parameter when [2, 1e5]. N=1e+006 Bisection result ρ=0 Roune bisection result ρ=0 Bisection result ρ=0.1 Roune bisection result ρ=0.1 Bisection result ρ=0.3 Roune bisection result ρ=0.3 Bisection result ρ=0.6 Roune bisection result ρ=0.6 Bisection result ρ=0.9 Roune bisection result ρ= Number of clusters x 10 4 Figure 5. Optimum value of epening on parameter =100 =333 =1000 =3333 =10000 parameter vs ρ parameter ρ Hit Ratio) Figure 6. Depenency of parameter with ρ ii Network, Eq.16 has been solve using the numerical metho of bisection. The value of N = 10 6 has been use for this analysis an a large range of values has been explore. The results are shown in Fig.5, the soli line is the result of the bisection metho an the ashe line is the nearest integer to the bisection results for each value of. We can observe how the value of changes with. It is interesting to note that the variation of is not a monotonic function with an it has a minimum value epening on the intra-omain hit ratio ρ ii. Fig.6 presents the epenency of with the intraomain hit probability. In soli lines is represente the result of the bisection metho applie to ifferent values of ρ ii for N = 10 6 an fixe values of. Dash-ot lines are use for the nearest integer values to the bisection results. The legen of the figure has the corresponence between the markers an the ifferent values of. If < N, the epenency of with respect to ρ ii is a monotonous increasing function. Whereas, if > N, the epenency of

6 45 40 Mean number of hops epening on an 2,1e5). N=1e+006 ρ=0, l, i ρ=0, opt ρ=0.1,, l i ρ=0.1, opt the maintenance state. 5. Relate Work Mean number of hops ρ=0.3, l, i ρ=0.3, opt ρ=0.6, l, i ρ=0.6, opt ρ=0.9, l, i ρ=0.9, opt Number of clusters x 10 4 Figure 7. Routing Performance Hops) with respect to ρ ii is a monotonous ecreasing function. Furthermore, for = N the value of with respect to ρ ii is a constant. This is the point where the monotonous increasing epenency changes to monotonous ecreasing. When ρ ii 0, 0.65), the value of [7, 8] inepenently of. This implies a wie range of work for if ρ ii 0, 0.65). However, for higher values of ρ ii, [5, 9]. Thus, more attention must be pai to ρ ii since the parameter has a smaller range of work. Furthermore, in orer to see the performance of a hierarchical CAN overlay, the Routing Performance complexity for an optimum flat CAN overlay is plotte in Fig.7 with a ashe line with stars flat = 14). The Routing Performance for ifferent values of the intraomain hit ratio is in ashot lines when l an i are configure inepenently. On the other han, the Routing Performance when an unique parameter is use is epicte with ashe lines. The values of were obtaine from Fig.5. The key shows the ifferent values of ρ ii that were use. Obviously, the Routing Performance is improve if the intra-omain hit ratio increases. A better Routing Performance than the flat counterpart can be obtaine with small values of ρ ii. Aitionally, we obtain a smaller maintenance cost if we use the optimal values of from Fig.5. We can also observe how the ifference between using an unique value of or a value of for each level ecreases if the intra-omain hit ratio increases. In aition we obtain a reuction of the state maintenance. This state is reuce from flat overlay routing entries [14] to the values shown in Fig.5. The optimal value of for the Hierarchical CAN Overlay Network epens on the number of omains, the number of peers an the intra-omain hit ratio, but usually, for most of the cases, we get a reuction of the 50% in flat Stuies relate with hierarchical overlay networks have been one to improve the performance of flat overlay networks. When a hierarchical architecture is consiere, the ifferent trae-offs that arise with these types of architecture have to be taken into account. Furthermore, these architectures have benefits [8], [9] in comparison with the flat ones. One approach is to elegate all the work to superpeers [5], [20], which maintain the overlay network an perform all the necessary actions, peers only have to register their information to their super-peers. In our proposal, super-peers only have to store the information neee of the routing table for its P2PSIP omain, which is lighter logm i ) an esirable for a more scalable solution. Other stuies focus on ecrementing the elay in peer-to-peer overlay transactions. In [19], a low elay hierarchical overlay network base on Chor is propose. The rawback is the high state maintenance neee memory, CPU an banwith) because all the peers in the overlay are attache to all the levels in a n-level hierarchy. A less aggressive esign with the same objective is presente in [4] where a hierarchical structure is built with the constraint of limiting the maintenance cost to the flat counterpart. The proposal in this paper only overloas super-peers, which is esirable because peers can be more lightweight. This feature is attractive for hanhel evices that are caniates to use Skype-like applications in the future. In relation with Hierarchical P2PSIP architectures, there is some work presente in [17]. However it oes not provie a very etaile analysis on the Routing Performance an how to eal with non-fixe lengths in the Prefix an Suffix IDs neee for interoperability. 6. Conclusions In this paper a Hierarchical DHT-base Overlay Architecture has been efine using the ongoing work in the IETF P2PSIP WG. In our proposal, there are ifferent P2PSIP omains where an overlay is maintaine using the P2PSIP protocol. Each omain has a super-peer that allows the connectivity with the rest of the omains through the Interconnection Overlay. The routing in each omain is base on a Suffix ID, whereas the routing in the Interconnection Overlay is base on the Prefix ID. If the Prefix ID an the Suffix ID are efine respectively as Prefix ID=hashexample.com)

7 an Suffix a global Skype-like service can be provie between the ifferent P2PSIP omains. We provie a iscussion on the Routing Performance epening on the intra-omain hit probability ρ ii ), which is the probability of oing calls insie the own omain. A reuction in the Routing Performance with respect to the flat counterpart is achieve since the hierarchical architecture takes avantage of the intraomain hit probability, which is not possible in a flat overlay network. In the case of a Hierarchical CAN Overlay, the optimal value of for a given number of omains an values of ρ ii is obtaine. Furthermore, we observe how the intra-omain hit ratio reuces the Routing Performance on a hierarchical CAN overlay network an how the optimal values of can be use in a wie range of work. Aitionally, the complex relation between the optimal value for the number of imensions an the value of the intra-omain hit probability an the number of groups is explaine. We conclue that most of the cases, the maintenance costs is also reuce up to a 50% in comparison with the flat one. In orer to continue this work, simulations using the PeerFactSim.om [3] simulator are being evelope. Acknowlegement This work has been supporte by the European Union uner the IST Content NoE FP IST ), by the Regional Government of Mari uner the BioGriNet project CAM, S-0505/TIC-0101) an by the Ministry of Science an Innovation uner the CONPARTE project MEC, TEC C03-03/TCM). References [1] B. Beverly Yang an H. Garcia-Molina. Designing a super-peer network. In Data Engineering, Proceeings. 19th International Conference on, pages 49 60, [2] D. Bryan, P. Matthews, E. Shim, an D. Willis. Concepts an terminology for peer to peer sip, July Internet Draft raft-ietf-p2psip-concepts-02.txt. [3] V. Darlagiannis, A. Mauthe, N. Liebau, an R. Steinmetz. An Aaptable, Role-base Simulator for P2P Networks. Proceeings of the International Conference on Moeling, Simulation an Visualization Methos, pages 52 59, [4] P. Ganesan,. Gummai, an H. Garcia-Molina. Canon in g major: esigning hts with hierarchical structure. In Distribute Computing Systems, Proceeings. 24th International Conference on, pages , [5] L. Garces-Erice, E. W. Biersack,. W. Ross, P. A. Felber, an G. Urvoy-eller. Hierarchical p2p systems. In Proceeings of ACM/IFIP International Conference on Parallel an Distribute Computing Euro-Par), [6] S. Guha, N. Daswani, an R. Jain. An experimental stuy of the skype peer-to-peer voip system. In In IPTPS [7] C. Jennings, B. Lowekamp, E. Rescorla, S. Baset, an H. Schulzrinne. Resource location an iscovery reloa), July Internet Draft raft-ietf-p2psipreloa-00.txt. [8] M. won an S. Fahmy. Towar Cooperative Interoverlay Networking. Proceeings of IEEE ICNP, [9] M. won an S. Fahmy. Synergy: an overlay internetworking architecture. In Computer Communications an Networks, ICCCN Proceeings. 14th International Conference on, pages , [10] I. Martinez-Yelmo, R. Cuevas, C. Guerrero, an A. Mauthe. Routing performance in hierarchical htbase overlay networks. In In Proceeings on 16th Euromicro International Conference on Parallel, Distribute an network-base Processing, Feb [11] P. Maymounkov an D. Mazieres. IPTPS 2002 Cambrige, MA, USA, March 7-8, Revise Papers, volume 2429/2002 of Lecture Notes in Computer Science, chapter aemlia: A peer-to-peer information system base on the XOR metric, pages Springer, [12] S.-H. Min, J. Holliay, an D.-S. Cho. Optimal superpeer selection for large-scale p2p system. In Hybri Information Technology. ICHIT 06. Vol 2. International Conference on, volume 2, pages , [13] A. T. Mizrak, Y. Cheng, V. umar, an S. Savage. Structure superpeers: leveraging heterogeneity to provie constant-time lookup. In Internet Applications. WIAPP. Proceeings., pages , [14] S. Ratnasamy, P. Francis, M. Hanley, R. arp, an S. Schenker. A scalable content-aressable network. In SIGCOMM 01, pages , New York, NY, USA, ACM Press. [15] J. Rosenberg. Interactive connectivity establishment ice): A protocol for network aress translator nat) traversal for offer/answer protocols, October Internet Draft raft-ietf-mmusic-ice-19.txt. [16] D. Rossi, M. Melia, an M. Meo. A etaile measurment of skype network traffic. In In IPTPS 2008, [17] J. Shi, Y. Wang, L. Gu, L. Li, W. Lin, Y. Li, Y. Ji, an P. Zhang. A hierarchical peer-to-peer sip system for heterogeneous overlays interworking. Global Telecommunications Conference, GLOBECOM 07. IEEE, pages 93 97, Nov [18] H. Simon. The architecture of complexity. In M. Press, eitor, The Sciences of the Artificial, pages [19] Z. Xu, R. Min, an Y. Hu. Hieras: a ht base hierarchical p2p routing algorithm. In Parallel Processing, Proceeings. International Conference on, [20] S. Zoels, Z. Despotovic, an W. ellerer. Cost-base analysis of hierarchical ht esign. In Peer-to-Peer Computing. P2P Sixth IEEE International Conference on, pages , 2006.