PP/ Grd-based Overlay Archtecture to Support VoIP Servces n Large Scale IP Networks We Yu *, Srram Chellappan # and Dong Xuan # * Dept. of Computer Scence, Texas A&M Unversty, U.S.A. {weyu}@cs.tamu.edu # Dept. of Computer and Informaton Scence, The Oho-State Unversty, U.S.A. {chellapp, xuan}@cs.oho-state.edu Abstract: Communcaton servces such as, Voce over IP (VoIP), over the Internet have ganed much attenton n recent years, as next generaton Internet applcatons are requrng the ntegraton of voce and data n the sngle IP nfrastructure. In ths paper, we propose a Peer-to-Peer (PP)/ Grd-based archtecture to effcently provde VoIP servces n large scale IP networks. Partcularly, our technologes nclude: 1) A Mult-overlay archtecture that leverages the resources/ capabltes of ndvdual VoIP components. ) We desgn and develop several models and protocols for realzng VoIP servces n our archtecture. We conduct extensve performance evaluatons on dfferent schemes proposed. The evaluaton results show that the load-aware weght-based call routng scheme can acheve much better performance than statc selecton schemes n terms of average call routng delay. The expermental results also demonstrate that the PP+herarchy model for conferencng applcatons can acheve better performance than all other models n terms of mnmzng the network bandwdth overhead. eywords: Communcaton Servces, Grd Computng, PP-based Computng, Overlay, VoIP I. Introducton In ths paper, we propose a Peer to Peer (PP)/ Grd-based archtecture to provde scalable and effcent Voce over IP (VoIP) communcaton servces n large scale IP networks. We leverage the resources/ capabltes of all components present n the VoIP networks nto two grd overlays: a call routng overlay and a user/servce endpont overlay. In ths PP/ Grd-based archtecture, we study varous ssues related to desgn, analyss and evaluaton of VoIP servces on these overlays. Provdng scalable and effcent communcaton servces such as VoIP over the Internet has become an actve research area n recent years, as the next generaton Internet applcatons are requrng the ntegraton of voce and data n the IP nfrastructure. However exstng system for VoIP have followng drawbacks: 1) Centralzaton of call control. A majorty of the exstng approaches to deploy communcaton servces n the IP network adopt the clent/ server model. For example, H.33-based communcaton system has two types of components: the centralzed call control agent as the server and endpont IP phone user agent as the clent. As the server provdes the call control and feature servces for all clents, t easly becomes the bottleneck due to the centralzaton of data access and control. ) Centralzaton of feature delvery. Smlar to the approach for call control, the exstng approach to delver features s also centralzed. The drawbacks can be llustrated by the followng example: Centralzed approaches make the system unable to effectvely support some basc features such as, conferencng. Tradtonally the conference feature s conducted by the call control server whch drects the call to the centralzed conference brdge wth a large amount of conferencng resources. When conferences have large number of users, the network wll easly be congested due to a large amount of traffc drected to certan area of the network. 3) Manual call routng confguraton. To support the communcaton servce wth clents at dfferent locatons, the tradtonal approach requres that each call agent manually confgures a route-pattern to each other call agent. In ths sense, n a system wth N locatons, each call agent needs to mantan N-1 route-patterns. Ths approach cannot s not scalable and wll have much mantenance overhead and wll be unable to realze global deployment of the servce n the near future. In the mgraton perod, a scalable call routng archtecture s hghly requred to support a varable network topology consderng call agents jonng/ leavng the system dynamcally. Ths wll be necessary for transparent and automatc deployment of new call agents wthout 1
much mpact on exstng communcaton systems, call agent falng due to DoS (Denal of servce) attacks n the IP network [1]. Followng from the above observatons, we propose a novel and effectve approach to address the problem of VoIP communcaton servces. We propose a scalable and effcent PP/ Grd-based archtecture ntegratng all the components present n VoIP systems. Our archtecture leverages the resources/ capabltes of all components such as, IP phones, servce agents etc, that can be shared, managed, coordnated and controlled n an effectve way. Our archtecture s thus completely dstrbuted, as each entty mantans comparatvely small states and ncreases the capacty of the global system to better conduct the call control and provde features. In recent years, PP and Grd computng have become effectve methodologes to support the large scale servces. Examples of such PP systems are applcaton level overlay systems such as CAN [], Chord [3] and Tapestry [4] are scalable, decentralzed and self organzng systems. Globus toolkts are evolvng towards an Open Grd Servces Archtecture (OGSA) n whch a grd provdes an extensble set of servces for the vrtual organzatons [5]. Leveragng PP/ Grd technologes for global communcaton servces brng n two valuable benefts mproved effcency and better QoS (qualty of servce). Snce the IP phone endponts n ths nfrastructure can exchange nformaton drectly, rather than through ntervenng dedcated servers, work and results can be dstrbuted quckly and effcently. In ths paper, we employ the followng technologes for VoIP servces. 1) Mult-overlay archtecture: We propose an ntellgent PP/ Grd based archtecture wth two functon grd overlays that leverages the resources of all components n the VoIP systems. The two overlays are: call routng grd overlay and user/ servce endpont overlay. The call routng grd overlay s composed of call agents, whch are dedcated to perform the call routng. The endpont grd overlay s composed of all the user endpont agents and servce agents, whch are dedcated for the call control and feature delvery. Our archtecture s hghly dstrbuted and s also compatble wth IETF standard SIP (Sesson Intal Protocol) [6]. ) PP-based call routng grd overlay: We propose a PP based call routng grd overlay. We also study protocols for overlay constructon and mantenance under dynamcs, and for forwardng the receved call routng request messages by desgnng effcent mult-path routng algorthms. We propose two schemes, namely a Statc routng scheme and Weght random selecton scheme for path selecton. Our protocols use easly avalable system nformaton to acheve better QoS performance. 3) PP-based user/servce endpont overlay: Ths overlay takes care of call control and feature delvery. We develop four overlay models, namely centralzed, herarchcal, PP and PP+herarchy models to provde such servces. We conduct extensve performance evaluatons on dfferent models proposed. The rest of paper s organzed as follows: In Secton II, we ntroduce the network model and some termnologes. In Secton III, we present our archtecture wth two grd overlays: call routng overlay and endpont overlay. In Secton IV, we dscuss the call routng overlay along wth network dynamcs. In Secton V, we dscuss the endpont overlay for large conferences. In Secton VI, smulaton and evaluaton results are presented. In Secton VII, we gve the survey of related work. The summary of ths paper and some future work are gven n Secton VIII. II. Network Models In the communcaton system, we consder a world-wde voce sgnalng network wth a large number of regon call agents. The call agent s a sgnalng mddleware-box and s capable of handlng call related sgnalng messages for the IP phone endponts. The regon call agent can support dfferent applcaton layer sgnalng protocols n the IP network, such as SIP, H.33 and MGCP (meda gateway control protocol). Each regon call agent s assgned wth a route-pattern for global call routng purpose. Assumng that each regon call agent can support 1, phones, 97943 s the correspondng route-pattern for the partcular regon call agent wth dalng number 97943XXXX. Assume that each regon control agent manages a gven number of regstered IP phones. When the IP phone makes the outgong call by dalng the destnaton dgts, the regon agent fnds the destnaton regon call agent through the call routng nfrastructure. The destnaton regon call agent wll negotate wth the source regon agent and wll setup the call. We restrct our work to a sngle doman or multple domans wthn our jursdcton. In ths sense, we can deploy call agents anywhere n domans wth known network topology. We assume that endpont IP phone can contrbute to the communcaton servce, ts avalable CPU and memory resource. For example, t may have lmted DSP (dgtal sgnal processng) resources to support
ad-hoc conferencng features. In other words, besdes the normal DSP resources to play the tone and rng, we also assume the each IP phone has extra DSP resources to conduct conferencng wth lmted number of partcpants. Central Controller Central Controller Central Controller Endpont IP Phone SW SW SW SW SW SW UNI NNI NNI UNI Fgure 1: Tradtonal Centralzed Call Model Endpont IP Phone III. Intellgent PP/Grd-based IP Communcaton Archtecture In the tradtonal VoIP system, servces are delvered by the clent/server model shown n Fgure 1, where the centralzed call control server conducts the basc call sgnalng and feature nteracton for the IP phone endpont user agents. In ths model, the user has lttle control over the call processng and feature delvery as the user devce s consdered as a dummy devce wthout any ntellgence or resources to delver features. All the system ntellgence s located at the centralzed call control agent, whch becomes a very complex system. The servces and features are mplemented on the top of the central call control server by usng feature prmtves. Wth more features requred n the communcaton servces, software desgn becomes the bottleneck due to complexty n handlng the nteractons of dfferent features and call control coupled wth fast delvery demanded by next generaton communcaton servces. In order to develop a scalable, lghtweght and easly deployable archtecture to effcently provde communcaton servces, we propose an ntellgent PP/ Grd-based archtecture that leverages the resources/ capabltes of system components as shown n Fgure. The hghlghts of our archtecture are: 1) The workloads of both call control and feature servce are dstrbuted to all components ncludng the user endpont IP phone. Thus call control and feature delvery becomes scalable, as there s no centralzaton n mantanng large number of states for the call control and feature delvery. ) Call routng s conducted by the call routng grd, whch s composed by a number of call routng agents whch are dedcated to perform the call routng. In the call routng grd, the PP-based approach s adopted to effcently support large scale systems. It can also handle network dynamcs under call agents jonng/ leavng the system. 3) Feature delvery s handled by the endpont overlay grd, whch s composed by all endpont user agents and servce agents n the system. In ths archtecture, components contrbute ther avalable resources/ capabltes and make the system effectvely support features such as, large conferencng etc. A 1 A 1 A A Endpont Grd Overlay R R R R Call Routng Grd Overlay Fgure : PP-based Communcaton Servce Archtecture (A 1 User Endpont Agent, A Servce Endpont Agent, R Call Routng Agent) Generally, our communcaton servce archtecture provdes an ntellgent envronment, whch can enable all components n the system to effectvely manage ther resource, complement each other, hence 3
makng the system scalable to support features for communcaton servces. In the followng sectons, we wll dscuss these two overlays n detals. IV. Call Routng Overlay A. Overvew Our call routng problem s defned as follows: Gven a sgnalng network wth call agents and destnaton route-pattern, we need to fnd the correspondng call agent for the gven destnaton routepattern. In other words, we need to fnd the mappng between the dalng number and the correspondng call agent. Accordngly, we have followng dfferent approaches: 1) Centralzed approach: Just lke Napster [7] and the DNS system for the Internet fle sharng and namng servce respectvely, ths approach has a centralzed drectory server or a herarchcal tree wth dfferent drectory servers. All call agents regster wth the drectory server. When the call agent receves call setup requests, t checks the daled number. If the daled number s out of ts regon, t sends the request to the drectory server for the destnaton lookup and the drectory sever responds wth the correspondng call agent for the daled number. The man advantage of ths scheme s ts smplcty and easy mplementaton. But ths server could be a bottleneck and a sngle pont of falure. ) Purely dstrbuted approach: In ths approach, each call agent mantans the nformaton of all other call agents by means a fully connected topology. Smlar to the centralzed approach, the call routng path for each source/destnaton par wll be just one hop. Wth ncrease n system sze, ths approach s not scalable as each call agent needs to mantan a large number of states for other call agents. 3) PP-based overlay approach: To allevate the above problems, we propose a PP-based overlay approach. Ths approach requres an overlay sgnalng network consstng of call agents and each call agent n the system has only a lmted number of neghbors. Thus, each call agent only needs to mantan local nformaton and call agents cooperate to acheve the global call routng objectve. The beneft of ths approach s that t s hghly scalable and fault tolerant. The PP overlay approach s naturally scalable under large system szes. In our PP-based overlay, the followng two tasks need to be conducted: 1) Overlay constructon and mantenance: Ths task deals wth how the overlay topology s constructed and ts mantenance under network dynamcs such as node jons/ leaves. ) Call routng forwardng: Ths task deals wth how to forward the call routng request n the overlay to acheve the call routng objectve. In the followng sectons, we wll dscuss the above two components n detal. B. Overlay Constructon/ Mantenance Protocol The objectve of overlay constructon and mantenance protocol s to establsh call routng table for the routng message. Ths s acheved by, 1) generatng canddate entres for constructng and updatng call routng tables, ) determnng elgblty of canddate entres, 3) constructng or updatng call routng tables wth the elgble entres, whch wll be used by the call routng forwardng protocol, dscussed later. We can use exstng approaches to collect the network nformaton, e.g., Landmark approach proposed n [8] can be used for montorng the node dstance. Once the nformaton on network status s collected, the routng table for each node can be easly constructed, by say, selectng the set of nodes wth shortest dstance. For the overlay constructon and mantenance, the man task s to handle dynamc behavor of nodes n the network (node jons/ leaves). Protocol 1 lsts two sub-protocols to handle the cases of node jons and node leaves. Durng the system ntal tme, the overlay can be automatcally constructed by each node runnng the node jon sub-protocol. Durng the run-tme wth new node jonng/leavng, these two sub-protocols wll be executed correspondently, the new updated call routng table wll be automatcally generated. Protocol 1: Overlay Constructon and Mantenance Protocols (OCMP) H: the set of neghbors of node N (node leavng case), d: the number of neghbors, N: the route pattern of node jonng/ leavng the system, where N[] represents -th dgt, S: the set of neghborhoods for node N (node jonng case), R: root node // Node Jon Sub-protocol (for node N) We assume that new node N can always fnd the network root node R n ts local regon Sends message(r, N) to the root node R and sets S empty R broadcasts the message to ts neghbors When node L receves the message Calculate the dfference of route-pattern (L, N) wth mask LN 4
The mask LN [l]( l [ 1, T ] ) s calculated as mask LN [l] = 1, f L[l] < N[l] mask LN [l] =, f L[l] > N[l] mask LN [l] =, f L[l] = N[l] If only one of mask LN [l] for l [ 1, T ] ) s not equal to and there s no neghbor, whch s nearer to N than L Send message response back to R Both L and N update ther call routng tables Else Forwardng the message based on protocol 1 Wth recevng d responses, the node N fnalzes the call routng table // Node Leavng Sub-protocol (for node N) Node N sends message(x) to all local neghbors H[1,...,d] and each neghbor H[] just deletes the entry N from ts call routng table For entry H[], fnd the symmetrc neghborhood H[x] satsfyng followng condton Assume only -th dgt s dfferent between H[] and N If (N > H[]) Fnd the neghbor X wth -th dgt larger than N[] H[x] = X If(N < H[]) Fnd the neghbor X wth -th dgt less than N[] H[x] = X H[] and H[x] set up the neghbor relatonshp. C. Overlay Call Routng Forwardng Protocol The objectve of the call routng request forwardng protocol s to forward the receved call routng request messages. Ths s acheved by performng the followng three tasks n sequence: 1) locate the call routng entry (entres) for the ncomng call request, ) select one entry among the located entres and 3) forward the message accordng to the felds n the selected entry. To perform these tasks, we propose the call routng request forwardng protocol below. Protocol : Call Routng Request Forwardng Protocol (CRRFP) T: the length of route pattern for the dalng dgts, S: the route pattern of the callng party number, D: the route pattern of the called party number, P : the route pattern at call agent node, M(S, D): routng request message wth source S and destnaton D For call agent, t receves the call routng request message M(S, D) Calculate the mask SD to dentfy the dfference between route pattern P and D The mask P D [l] ( l [ 1, T ] ) s calculated as mask P D [l] = 1, f P [l] < D[l] mask P D [l] =, f P [l] > D[l] mask P D [l] =, f P [l] = D[l] If mask P D [l] are for all l [ 1, T ] The call agent s the destnaton The call agent fnds the correspondng IP phone as the destnaton and forwards the request to the IP phone user agent Else Assume there are d non-zero entres n mask P D [1,.., T] Select 1 from d non-zero entres (recorded at the mask SD ) set wth selecton schemes by formula () For example, j-th entry s selected Fnd the next hop,.e., N Forward message M(S, D) to the next hop N In connecton wth Protocol, we have followng theorem. Theorem 1 (Correctness of the protocol ): If the network has no faults, then wth our call routng protocol, the call requestng message from a source wll eventually be delvered to the destnaton. 5
Proof: Let us say an agent R receves a message from node R (R can be source node). From the call routng algorthms mentoned above, we can easly acheve the followng concluson: gven any path taken by our protocol, we can prove that R s always one hop near to the destnaton than R (recall that we choose one of d-non-zero enttes n the mask calculated n Protocol ). In ths sense, the call routng s loop free, whch guarantees that the call routng request wll eventually arrve at the destnaton. Q.E.D. Wth QoS consderatons, our motvaton for the call routng s to mnmze call routng message delay. As there are multple paths from the source to the destnaton, we adopt mult-path routng selecton approach. As we know, the throughput of routng messages depends on the network topology and load of network node. Hence, a mult-path selecton decson should take all avalable nformaton nto account. The path selecton nformaton represents the system nformaton whch can be used for path selecton. We consder two mportant factors: the dstance of the path P and load of network node L, where P can be obtaned by call routng algorthms and L can be obtaned by exchangng the node load nformaton among network neghbors perodcally. From here, we wll regard the avalable nformaton for the path selecton as metrcs. A sngle type of nformaton corresponds to one metrc. In the case where there s only one type of the nformaton (n other words, only one metrc), say, the dstance of the path, t s straghtforward to fgure out that the path wth shortest dstance s better than one wth longer dstance. In the case of multple metrcs, say, the dstance of the path, and load on a node, t s dffcult to judge whch one s better among the followng two paths: one s wth shorter dstance and heavy load, or one wth a longer dstance and lghter load. Hence, we need to combne dfferent ndvdual metrcs together to get a composte one, whch can be done n many dfferent ways. The basc dea s to desgn a functon, takng these metrcs as nput parameters. The output of ths functons n the composte metrc C,.e., C = f(p, L). Note that ths functon should be monotoncally ncreasng or decreasng for each metrcs. A smple functon can be C = L / P. (1) Once the composte metrc s ready, the followng ssue s how to use t to select a path. We propose followng weghted random selecton approach. The basc dea of the weghted random selecton approach s that each path s assgned to a weght, and a path s chosen randomly based on weght assgnment. The path wth hgher weght value has a hgher probablty of beng selected than one wth lower weght values. Ths approach may effectvely dstrbute the traffc over dfferent routes and potentally mprove the throughput performance. Weght assgnments for mult-path routng have been studed n the lterature. For example, an optmal weght assgnment was proposed n [9] n whch the average delay of message can be mnmzed. We wll take a heurstc approach for weght random selecton. Wthout loss of generalty, we assume that elgble entres are n the routng table for a gven destnaton route pattern. We let these entres be ndexed by 1,,, ; and ther assocated weghts are denoted as W 1, W,.., W k, respectvely. We can assgn the weght W ~ C (C s calculated by (1)) for each neghbor node. To normalze W to be a probablty such that = W = 1 1, we should assgn W as follows: for =1,,, L / P W =. () L / P j = 1 j j V. Endpont Overlay In the endpont overlay, the most mportant component s the endpont user agent,.e., IP phone. In ths overlay, the PP-based approach s adopted to make all components contrbute ther avalable capablty/ resource and make the system effectvely support features such as, large conferencng and so on. For large conferences (a large number of endpont users are jonng n a conference), dependng on sequence of each node jonng n the conference, all the conference members construct an overlay dedcated to the conference. In ths overlay, each endpont node can contrbute ts local DSP resources to conduct the conference meda mxng. Thus, more effcent conferencng can be supported by ths approach. We dscuss ths n detal below. Note that our analyss s generc and can be used for provdng other features too. Assumng that N endpont user agents are n a large conference, we consder followng models for the conference feature: 6
1) Centralzed Model: Each endpont user agent keeps two meda sessons connectng to the centralzed meda servce agent. From the endpont user agent perspectve, the outgong meda sesson sends the RTP packets from the endpont user and the ncomng meda sesson receves the RTP packets from the meda servce agent. ) Herarchcal Model: The system has multple meda conference servce agents, whch are herarchcally organzed. For example, the two layer herarchcal network can be the system wth R regons. Each regon has the N/R enttes and 1 centralzed meda servce agent s deployed. All endpont user agents just connect to the local regon meda servce agent just lke the centralzed model. All regon meda servce agents can be reconnected to support the large conference. 3) PP-based Model: In ths system, we assume that end pont user agents have lmted capablty to mx the meda connected to t. For example, consder user A and B havng an actve call. Then, A decdes to conduct a conference by nvtng user C by makng a consultaton call to C. There s no call set up drectly between B and C. A receves meda stream from both B and C, and mxes them. A sends a stream combnng A s and B s streams to C. In ths sense, the endpont user agent wthn the small area can be grouped as a cluster wth a leader (A s the cluster leader n ths case) whch takes the responsblty to conduct the meda mxng for the local cluster. Dependng on sequence of all nodes jonng n the conference, all the conference members construct an overlay dedcated to the conference. In ths overlay, each endpont node can conduct the conference meda mxng locally to acheve effcent conferencng. 4) PP+Herarchy Model: In the system, the Herarchcal model and PP-based model are combned. In ths sense, the network has R regons and each regon adopts the PP-based model. Based on above models, we defne followng performance metrcs: Bandwdth Overhead (BO) defnes the bandwdth consumpton for dfferent models. Based on the meda topologes for each model, we can easly calculate the bandwdth overhead as followng: 1) Centralzed Model: log N 1 BO c = + +... + ( log N 1)* (3) ) Herarchcal Model: log ( N / R) 1 BO h = R + R * ( + +... + ( log ( N / R) 1) * (4) Smlarly, we can fnd R to mnmze the overhead for the network wth sze N. 3) PP-based Model: Each ntellgent endpont user agent can handle a part of meda sesson locally. For example, several nearby IP phone user agents can construct a PP topology to mx the meda. Suppose each endpont user has the capablty to support M meda sessons, the bandwdth overhead s, log N / M 1 BO p = + +... + ( log N / M 1) * (5) 4) PP+Herarchy Model: The Bandwdth overhead s gven by, log N /( RM ) 1 BO = R + R * ( + +... + ( log ( N /( RM )) 1) * (6) ph VI. Smulaton Results and Analyss In ths secton, we evaluate the performance of the system that uses our proposed algorthms and protocols. We wll frst descrbe the expermental model and then report performance results. A. Performance of Call Routng Overlay 1) Expermental Model Network model: For the performance evaluaton of the call routng algorthms, we evaluate two types of network topologes: random and real networks; 1) Random network: we randomly generate a random topology wth network sze, 4. ) Real network: we consder the MCI ISP backbone network wth 19 nodes, whch are nterconnected by lnks wth 1 kbps for voce sgnalng traffc. We assume each node s deployed wth a call agent. Traffc Model: We assume that call requests form a Posson process wth rate [, 1] per second. Baselne System: For the call routng algorthms, we consder followng baselne systems A { SR, WR }, where SR denotes the statc routng (the next hop s always chosen by shortest path selecton) and WR denotes the weght random selecton scheme (the next hop s calculated by the weght formula ()) 7
Performance Metrcs: For the routng algorthms, we consder the followng performance metrc - Average Routng Delay (RD): t s defned as the average call routng message delay n the system. The hgher the RD value s, the worse s the performance. ) Performance Results In ths secton, we report results of our experments along wth our observatons. Due to space lmtatons, we only present a lmted number of cases here. However, we found that the conclusons we draw are generally held for many other cases we have evaluated. Average Delay 5 4 3 1 Average Routng Delay Performance (4 nodes) 1 3 4 5 6 7 8 Request Rate WR SR Fgure 3: RD Performance of Routng Algorthms (network wth 4 nodes) Average Delay 9 8 7 6 5 4 3 1 Average Routng Delay Performance (MCI Network) 1 3 4 5 6 7 8 9 Request Rate SR WR Fgure 4: RD Performance of Routng Algorthms for Real Network Topology Fgures 3 and 4 show the senstvty of RD n a random network of sze 4 and real MCI backbone network, respectvely. We see that RD s ndeed senstve to the dfferent network szes and call routng schemes. We have followng observatons. RD of weghted random routng selecton scheme performs better than the shortest path routng scheme consstently for dfferent network szes. The reason s that weghted random selecton scheme reles on the fact that the hgh congeston paths are not beng selected. RD s senstve to the servce request rate R. The value of RD ncreases as R ncreases. The reason s obvous: a large R mples that a large number of requests are sent to call agents ncreasng the processng tme. B. Performance of Endpont Overlay Fgure 5 shows the performance of conference bandwdth overhead for dfferent models. In ths fgure, Peer(#) represents the system where each cluster has # nodes, Herarchcal($) represents the system wth $ regons, and Herarchcal+Peer($+#) represents the system wth $ regons wth each cluster havng # nodes. We have followng observatons: The bandwdth overhead becomes larger wth ncrease n network sze. The reason s smple. The large network sze ncreases the overall network path length, whch causes the large bandwdth overhead. Gven the network sze, the bandwdth overhead for the centralzed model gves the worst performance compared to other models. The PP+herarchy model acheves the best performance. 8
The reason can be explaned. The PP+herarchy model makes network meda traffc wth shorter network paths. As local endpont user agents contrbute the possble meda streamng capablty, the PP+herarchy approach s a good soluton to extend the system scalablty for large conference feature. Bandwdth Overhead (Hearchcal Model) Bandwdth Overhead (Peer-to-Peer Model) 5 5 Bandwdth Overhead 15 1 5 Bandwdth Overhead 15 1 5 1 3 4 5 6 7 8 9 Number of Nodes (3^x) Centralzed Hearchcal() Hearchcal(5) 1 3 4 5 6 7 8 9 (a) Herarchcal Model Number of Nodes (3^X) Centralzed P eer() P eer(5) Peer(1) (b) PP Model Bandwdth Overhead (Peer+Hearchcal) Bandwdth Overhead (Dfferent Models) 5 5 Bandwdth Overhead 15 1 5 Bandwdth Overhead 15 1 5 1 3 4 5 6 7 8 9 Number of nodes (3^X) Centralzed Harchcal and P eer (4 + ) Harchcal and P eer (4 + 4) 1 3 4 5 6 7 8 9 Number of Nodes (3^X) Centralzed Hearchcal(5) P eer(5) Harchcal + P eer(4 + 4) (c) PP + Herarchy Model (d) All Models Fgure 5: Bandwdth Overhead for Conference Feature VII. Related Work In ths secton, we revew research work n the area of PP/ Grd computng and communcaton systems n the IP network related to our study. Recent applcaton level overlay networks, e.g., CAN [], Chord [3] and Tapestry [4] are scalable, decentralzed and self organzng systems. Wth dstrbuted hash table (DHT) [1], nodes n these systems collectvely contrbute towards admnstraton-free storage space and acheve effcent content lookup. Our approach s dfferent from these systems due to the salent features of call routng. Moreover, as the communcaton sgnalng overlay s comparatvely stable and our call routng protocols beng much smpler, more effcent mechansms can be realzed for call routng. From the applcaton perspectve, Grdbased computng has drawn much attenton n recent years [11]. Future network systems need to have certan machne-understandable semantcs to make the ntellgent peer actvely and effcently provde ondemand computng [1]. Buldng an ntellgent socal grd, semantc resource grd and knowledge grd are also becomng mportant research areas [13] [14]. In ths sense, our work can be nterestng as an ntellgent grd applcaton. There s a lot of research work amng to mprove the avalablty of voce over IP servces. H.33 sute of protocols s developed by ITU, whch has been supported by several products [15]. Because of ts nablty to readly provde new servces, SIP s becomng popular and currently standardzed by IETF n RFC 543 [6] [16]. As SIP takes the dstrbuted control approach and supports MIME and URI, t can easly support combned servces n the Internet and make the servce delvery fast. In both SIP-based and H.33-based call control models, the call routng feature s performed by the controller (H.33 world) or SIP proxy (SIP world), whch are just examples of the generc call agent n ths paper, whch manly conducts the call routng task. The call routng problems have also been dscussed by adoptng nterdoman routng protocol (BGP) [17] [18]. Our approach s dfferent from these studes by adoptng the PPbased overlay approach to construct ntellgent grds at the applcaton layer to effcently support the communcaton servce. In ths sense, our approach s an extenson of prevous work from the framework of supportng large scale communcaton systems wth enough ntellgence to easly support exstng features and extend t to new features. 9
Much work related to mprovng the voce qualty from the network data plane s done. For example, aram et al. [19] studed the delay and jtter analyss of voce traffc n the Internet by analyzng dfferent schedulng algorthms. Chuah et al. [] studed the statstcal approach to make the bandwdth resource management more effcent. Our work dffers from the above by manly focusng on the network control plane. VIII. Conclusons and Future Work We have studed dfferent ssues regardng the provsonng of VoIP servces n large-scale IP systems by adoptng ntellgent PP/ Grd computng technologes. To the best of our knowledge, ths s the frst study that addresses ssues n ths feld. In summary, our technologes nclude the followng: 1) An ntellgent PP/ Grd based archtecture, ) PP-based call routng overlay, 3) PP-based endpont overlay to effcently support features. We conduct extensve performance evaluatons on dfferent archtectures and algorthms. The evaluaton results show that the load-aware weght-based call routng scheme acheves much better performance than the statc selecton scheme n terms of average call routng message delay. The expermental results also demonstrate that PP+herarchy model can effcently support large conferences n terms of mnmzng bandwdth overhead. Our work has broad mpacts. Wth a tremendous spurt n communcaton servces demanded by Internet applcatons, tradtonal approaches to delver communcaton servces fnd t ncreasngly dffcult to do so. Such servces can be easly deployed wth our approach presented n ths paper. There are several drectons to extend our study: 1) as the endpont (IP phone) s a generc devce, t can mplement a broader spectrum of other servces such as, personal computer vrus detecton and frewalls; ) our Grd-based overlay desgn approach s generc to provde dfferent servces; our desgn approach can easly be adopted n other areas lke, hgh performance nstant messagng system etc, whch wll be part of our future work too. References [1] J. Xu, W.Y. Lee, Sustanng Avalablty of Web Servces under Dstrbuted Denal of Servce Attacks, accepted to IEEE Transacton on Computers, specal ssue on Relable Dstrbuted Systems, 5() (3) 195-8. [] S. Ratnasamy, P. Francs, M. Handley, R. arp, S. Shenker A Scalable Content Addressable Network, Proceedng of ACM SgComm, 31(4) (1) 161-17. [3] I. Stoca, A Scalable PP Lookup Servce for Internet Applcatons, Proceedng of ACM SgComm, 31(4) (1) 149-16. [4] B. Y. Zhao, J. D. ubatowcz, Tapestry: An Infrastructure for Fault-reslent Wde-are Locaton and Routng, Techncal Report UCB/CSD-1-1141, Unversty of Berkeley, 1. [5] I. Foster, C. esselman, Grd Servces for Dstrbuted System Integraton, IEEE Computer, 35(6) () 37-46. [6] M. Handley, H. Schulzrnne, SIP: Sesson Intaton Protocol, RFC 543, 1999. [7] M. Macedona, Dstrbuted fle sharng: barbarans at the gates, IEEE Computer, 33(8) () 99-11. [8] S. Ratnasamy, M. Handley, R. arp, S. Shenker, Topologcally Aware Overlay Constructon and Server Selecton, Proceedng of IEEE INFOCOM, 3(3-7) () 119-1199. [9] R. G. Gallager, Mnmum Delay Routng Algorthms Usng Dstrbuted Computaton, IEEE Transacton on Communcatons, 5(1) (1997) 73-85. [1] M. Naor, U. Weder, A Smple Fault Tolerant Dstrbuted Hash Table, IEEE -th Internatonal Workshop on Peer-to-Peer Systems, 3, 88-97. [11] A. Chervenak, I. Foster, C. esselman, C. Salsbury, S. Tuecke, The Data Grd: Towards an Archtecture for the Dstrbuted Management and Analyss of Large Scentfc Datasets, Journal of Network and Computer Applcatons, 3(1) (1) 187-. [1] J. ephart and D. Chess, The Vson of Autonomc Computng, IEEE Computer Magazne, 36(1) (3) 41-5. [13] H. Zhuge, Semantcs, Resource and Grd, edtoral of the specal ssue of Future Generaton Computer Systems, (1) (4) 1-5. [14] H. Zhuge, Chna s E-Scence nowledge Grd Envronment, IEEE Intellgent Systems, 19(1) (4) 13-17. [15] ITU, Regstraton, Admsson and Status Sgnalng (Recommendaton H.5/RAS), Internatonal Telecommuncaton Unon (ITU), 1998. [16] J. Rosenberg, Dstrbuted Algorthms and Protocols for Scalable Internet Telephony, Ph.D Dssertaton, Dept. of Electrcal Engneerng, Columba Unversty, 1. [17] J. Glasmann, W. ellerer, H. Muller, Servce development and deployment n H.33 and SIP, 6-th IEEE Symposum on Computer and Communcaton, 3(5) (1) 378 385. [18] D. Hampton, D. Oran, The IP Telephony Border Gateway Protocol (TBGP), Internet Draft, draft-etf-ptel-glp-tbgp-1.txt, work n progress. [19] M. J. aram, F. A. Tobag, Analyss of the Delay and Jtter of Voce Traffc Over the Internet, Proceedng of IEEE INFOCOM, () (1) 84-833. [] C. N. Chuah, R. H. atz, Network Provsonng and Resource Management for IP Telephony, Computer Scence Dept., Report No. UCB/CSD-99-161, UC at Berkeley, August 99. 1
We Yu s bo: We Yu receved hs B.S (199) from Nan Jng Technology Unversty, M.S. (1995) from Tong J Unversty, and Ph.D degree (1998) from Shangha Jao Tong Unversty. All are n Electrcal Engneerng. Snce 1999, he s a Ph.D canddate n the Dept. of Computer Scence at Texas A&M Unversty. Currently, he s workng for Csco Systems, Inc. Hs research nterests nclude network securty and dstrbuted systems. Srram Chellappan s bo: Srram Chellappan s a graduate student n the Dept. of Computer and Informaton Scence at The Oho-State Unversty. Hs current research nterests are n network securty, dstrbuted systems and wreless networks. He holds a Masters Degree n Electrcal Engneerng also from The Oho-State Unversty and a Bachelors degree n Instrumentaton and Control Engneerng from the Unversty of Madras. Dong Xuan's bo: Dong Xuan receved hs B.S. and M.S. degrees n Electronc Engneerng from Shangha Jao Tong Unversty (SJTU), Chna, n 199 and 1993, and Ph.D degree n Computer Engneerng from Texas A&M Unversty n 1. Currently, he s an assstant professor n the Department of Computer and Informaton Scence, The Oho State Unversty. He was on the faculty of Electronc Engneerng at SJTU from 1993 to 1997. In 1997, he worked as a vstng research scholar n the Department of Computer Scence, Cty Unversty of Hong ong. From 1998 to 1, he was a research assstant/assocate n Real- Tme Systems Group of the Department of Computer Scence, Texas A&M Unversty. Hs research nterests nclude real-tme computng and communcatons, network securty and dstrbuted systems. 11