A QoS-Aware Intelligent Replica Management Architecture for Content Distribution in Peer-to-Peer Overlay Networks

S.Ayyasamy et al /Internatio nal Journal on Computer Sci ence and E ngineering Vol.1(2), 2009, 71-77 71 A QoS-Aware Intelligent Replica Management Architecture for Content Distribution in Peer-to-Peer Overlay Networks S.Ayyasamy 1 1 Asst. Professor, Depa rtment of Information Technology, Tamilnadu Col lege of En gineering Coimbatore-6 41 659, Tam il Nadu , INDIA, Email: ayyasamyphd @gmail.com S.N. Sivanandam 2 2 Professor a nd Head, Depart ment of Com puter Science and Engineeri ng, PSG Col lege of Tec hnology, Peelamedu, Coimbatore-6 41 004, Tamil Nadu, INDIA. Abstract — The large scale cont ent distribution systems were improved broadly using the r ep lication techniques. The demanded contents ca n be brought closer to the clients by multiplying the source of information geogra phically, which in turn reduce both th e access late ncy and the network traffic. The system scalability can be improved by distributing the load across multiple s ervers which is proposed by replication. I f a copy of the requested object (e.g., a web pa ge or an image) is located in its closer proximit y th en the clients would fee l low access latency . Depending on the position of the replicas, the effectiveness of replication tend s to a large extent. A QoS b ased overlay network architecture involving an intelligent repl ica placement algorithm is proposed in this paper. Its main goal is to improve the network utilization and fault tolerance of the P2P system. In addition to the replica pl acement, it also has a caching technique, to reduce the search latency. We are able to show that our proposed architecture attai ns less latency and bet ter throughput with reduc ed bandwidth usage, through the simulation results. Keywords-Clusters, Conten t, Overlay, QoS, Rep lica, Routing I. I NTRODUCTION 1.1 Overlay Networks To share the computer resources like content, storage, CPU cycles directly without using an i ntermediate sy stem or a centralized server, distributed computer architecture, called “peer-to-peer” are designed. They are distinguished by their failure adaptation capabilities a nd maintenance of acceptable connectivity and per formance [1]. Significant research attention has been appl ied to Content di stribution, whi ch is an important peer- to-peer applicat ion on the Internet . By allowing personal computers to work as a distribut ed storage medi um, they normally contribute, search and obtain digital co ntent. Overlays are flexible and depl oyable approaches that allow users to perform distributed operat ions without modifying t he underlying physica l network. Peer -to-peer (P2P) ov erlay systems have been proposed t o address a vari ety of probl ems and enable new applications. Th e attraction of these systems, when compared to client/ser ver frameworks, is in t heir robustness, reliability and cost efficiency. Unlike tradition al distributed computing , P2P networks aggregate large number of comp uters and possibly m obile or handheld devices, whi ch join an d leave the network frequently. Nodes in a P2 P network, called peer s, play a variety of roles in their interaction with other peers. When accessing information, they are clients. When serv ing information to other peers, they are servers. When forwarding in formation for other peers, they are routers. This new breed of systems creates application- level virtual netwo rks with their own overlay to pology and routing prot ocols. To search for data or resources, m essages are sent over multiple hops fr om one peer t o another wit h each peer responding to qu eries for information it has stor ed locally. Structured P2P overlays implem ent a distributed hash table data structure in which every data item can be lo cated within a small number of hops at the expense of keeping some state information locally at the nodes. 1.2 Replica Placement for QoS-Aware Content Distribution Replication tech niques are widely employed to improve the availability of data, enha ncing perform ance of query latency and load balancing in co ntent distributi on systems. We can geographically multiply the sou rce of information b y distributing mu ltiple copies o f data in the netw ork. By forwarding each quer y to its nearest copy , the query search latency can be effectively reduced. The ability to i mprove system scalabilit y through distributing the lo ad across multiple servers [2] is also offered by replication. If a replica of the requested object (e.g., a web page or an imag e) is kept in its nearer proximity then the clients would feel low access latency. Depending on the position of the replicas, the effectiven ess of replication tends to a large extent. The centralized servers be come a bottleneck as the requirement of the information increases. The performance problem is managed by the content providers, system administrat ors or end users by them selves through del ivering replicas of web content to m achines, spread throughout the network. Th e load on the central server [3 ] is reduced by replicas through respon ding to the local clien t requests. The load which is delivered to the cooperate nodes includes: ISSN : 0975-3397 S.Ayyasamy et al /Internatio nal Journal on Computer Sci ence and E ngineering Vol.1(2), 2009, 71-77 72 • Communicati on bandwidth for sending t he data to the requesting cont ent • Storage used for hosting the replica and • CPU resources for query pr ocessing. The problem of decidin g how many replicas is to be delivered to each file and its lo cation is given by the Replica manageme nt to this circum stances. To handle m ore requirements for each file, enough replicas should be present. Servers become overloade d and clients observe l ower performance by having o nly few replicas. On the other h and the waste bandwidth of ext ra re plicas and the storage which could be reassigned t o the other fil es, and also the m oney spent to rent, power and also f or host machine cool ing. In this paper, we propose QO S-aware Intelligent Replica Management (QIRM ) architecture for peer-t o-peer overlay network. It con tains a replica placem ent algorit hm and a robust query searching technique for data retri eval. This paper is organized as f ollows. Section 2 gives the detailed rel ated work done. Sect ion 3 present s the system model and algori thm overview f or the proposed a rchitecture. Section 4 presen ts the intelli g ent replica placement algorithm, followed by the searchi ng technique. Sec tion 5 gives t he experimental results and sect ion 6 concludes the paper. II. R ELATED W ORKS Most of the research efforts to im prove the perform ance of Gnutella-like P2P syste ms can be classified into two categories: P2P search and routing algori thms and P2P overlay topolo gies. Most of the pr oposed routi ng or search algorit hms in the first category, disregard the n atural peer heterogenei ty present in most P2P systems, and more importantly the potentia l performance hurdl e caused by the rand omly constructed overlay topolo gy. B. Mortazavi_ and G. Kesidis [4] have provided a survey of reputation s ystems. Based on a reputati on framework, th ey have designed a game in whic h users play to maxim ize the received files from the system. Fo r this, the users adjust their cooperation le vel, there by obt aining a g ood reputat ion as a result. Raphael Chand and Pascal Felber have desi gned for publish or subs cribe system based on peer -to - peer para digm. A containment-based pr oximity met ric was proposed which allows us to buil d a bandwidth-efficient network topolo gy that produces no false neg atives and very few false po sitives. They have also developed a proximity metric based on subscripti on similarities w hich yields a more solid graph stru cture with negligible false neg atives ratios and very few false positives [5]. Anwitaman Dat ta have discussed som e of the import ant issues concerning structure d P2P systems and interplay between the tw o P2P and MANET self-organizin g networks from a data ma nagement persp ective which aim s to achieve efficient and robust informat ion search and access schemes [6]. Paraskevi Raftopoulou and E uripides G.M. Petrakis have presented iCluster, a self-org anizing peer-to-peer overlay network for s upporting full-fl edged inform ation retrieval i n a dynamic environm ent. They de fined the criteria for peer similarity and peer sel ection, and also p resented the protoc ols for organizing the peers into clusters and f or searching within the clustered organization of peers [7]. Carvalho, N. Araujo, F. Rodrigues. L, have presented the IndiQoS architecture, a scalable QoS-aware publish- subscribe system with Qo S-aware publications and subscription s that preserves the decou pling which makes the publish-subscrib e model so appealing. To support such model, the proposed architecture Indi QoS includes a decentralized message-broker based on a DHT that leverages on underly ing network-level Qo S reservation mechanisms [10]. Guillaume Pierre and Maarten van Steen have presen ted Globule, a collaborat ive content delivery network. The Proposed network was composed of Web ser vers that cooperate across a wide-area ne twork to provide pe rformance and availability g uarantees to the sites they ho st [12]. Yan Chen et al. [14] have proposed the dissem ination tree, a dynamic conten t distribution syste m built on to p of a peer-to- peer location service. They have presented a replica placement protocol that has b uilt the tree while meetin g QoS and server capacity constraints. The number of replicas as well as the delay and band width cons umption for update propa gation wa s significantly redu ced. Jian Zhou et al. [15] have s hown that the replica placement problem in P2P networks has represented as a Clustered KCenter problem (which essen tially differed from the classic kcenter proble m) and is proven t o be NP-compl ete. To solve this problem , they bring for ward an approxim ation algorithm in the form of a distance graph for the network topolog y; when their defined feasib ility condition has hold at a cer tain point; the replica placement solut ion has built out o f (m-1) power of current distance graph. Kan Hung Wa n and Chris Loeser [16] have propose d techniques and alg orithms for point-to-point strea ming in autonomous syst ems as it might occur in large compani es, a campus or even in large hotels . Their major aim was to create a replica situation that inter-sub network RSVP streams are reduced to a minimum. Therefor e, they have introduced the architecture of an overlay ne twork for intercon necting sub networks. Each sub networ k cont ains a so-called local active rendezvous server (LARS) which do es not just act as directory server but also co ntrols availability o f movie conten t in its subnet work. Due to this, they have considered data placement strategies depending on restri ctions of network bandwidt h, peer capabilities, as well as the movie’s access frequency . Spiridon Bakiras and Tha nasis Loukop oulos [17] have discussed that the caching and re plication have emerged as the two primary t echniques for reducing t he delay experienced by end-users when downl oading web pa ges. They have investigated the po tential performance gain by using a CDN server both as a replicator an d as a proxy server. They have developed an analytical model to quantify the b enefit of each ISSN : 0975-3397 S.Ayyasamy et al /Internatio nal Journal on Computer Sci ence and E ngineering Vol.1(2), 2009, 71-77 73 technique, under vari ous system parameters, and t hey have proposed a greedy algorithm to solve the co mbined caching and replica placem ent problem. Jie Su and Douglas Reeves [18] have proposed tha t bounding the l atency of cl ient requests are a n important factor in solving the replica p lacement problem for content distribution net works. They ha ve proposed two alg orithms for placing replicas with latency constraints efficiently, one centralized, and one dist ributed. They have shown that t he impact on t he num ber of repli cas required as t he latency constraint has become more stri ngent. In the case where client request patterns were unknown, they have shown that the additional number o f replicas needed is reasonable. Unfortunately, most existing work on repl ica placem ent has focused on optimizing an av erage performance measure of the entire client community su ch as the mean access latency [8], [9]. While an average performance measure m ay be important from the syst em’s point of vi ew, it does not differentiate the l ikely diverse perform ance requireme nts of the individuals. So far, to the best of our knowledge, t here has been no study o n QoS-aw are replica placement. III. S YSTEM M ODEL AND A LGORITHM O VERVIEW 3.1 Algorithm Overview In our QOS aware t opology, nodes are grouped into stron g and weak clusters based on t heir weight vector which comprises the following parameters: Available capacity CPU speed Memory size Access Latency In the replica placement algorith m, we classify the content as Class I and Class II, based on their access patterns. (i .e.) The most frequently accessed cont ents are ranked as Class I and the less frequently accessed contents as Class II. Then more copies of Class I content ar e replicated in strong clusters (having high weight values). Routing is perform ed hierarchical ly by broadcast ing the query only to the strong clusters. Thus our proposed architecture achi eves Low bandwidth Consumption, Reduced Late ncy, Reduced Mai ntenance Cost, Strong Connectiv ity and Qu ery Coverage. 3.2 System Model Let us consider a collectio n of N server nodes which form a peer to peer (P2P) overlay network. In additi on to being part of the overlay, each node functi ons as a server responding to requests (queries) which com e from clients outside of the overlay network. An example coul d be that each node is a web server with the overlay l inking the servers and cl ients being web browsers on rem ote machin es requesti ng content from the servers. We assume each node always stores one copy of its o wn content item which it serves to cl ients and that it has additi onal storage space to store k repli cated content items from other nodes which it can also serve [3]. The object is a ssociated with an authoritativ e origin server (OS) in the network wh ere the content provider makes the upda tes to the object. The object copy located at the orig in server is called the origin copy and an object copy at any re maining server is called a rep lica . IV. I NTELLIGENT R EPLICA P LACEMENT A LGORITHM 4.1 Clustering the Nodes let n, 1,2....... i , N node each i  For bandwidth Available - i BW speed CPU - i SP Latency Access - i AL MZ i - Memory Size 1. The weight of the node Ni can be calculated as i i i i i AL MZ SP BW W ) (    2. Form the vector } , { i i W S W  , which denotes the node ids and their correspon ding weight values, sorted on the descending order. 3. Let {Sk} denote t he set of st rong cluster nodes ) k (0 n    , which satisfies the following condition   k W , where  is the minimum threshold valu e for the weight. 4. Then the set {Wj} = {Ni} – {Sk}, denote t he set of weak cluster nodes ) (0 n j    , which satisfies the conditi on   k W 4.2 Replica Placement Let QS be the query server which regi sters the query of each client. The query server stor es the clust er information of each node along with the node id as “S” or “W” for strong and weak clusters, respectively. At time Tk, let m clients generates query requests {Qm} of the form q{nid, ckwd}, where ni d is the node id of the client and ckwd is the keywo rd of the content to be retrieved. The queries {Qm} are register ed in the query server QS. The requested content of the queries are classified and categorized as class1 or cl ass2, depending on the access frequencies. (i.e.) A query Qj, j < m , is considered to be class1 If n (Qj) >= Amin and class2, ISSN : 0975-3397 S.Ayyasamy et al /Internatio nal Journal on Computer Sci ence and E ngineering Vol.1(2), 2009, 71-77 74 If n (Qj) < Amin Where n (Qj) is the num ber of access of the content pattern for the given query and Amin is the minimum access threshold value. Then the query server QS assi gns the class1 contents to the strong cluster nodes and cl ass2 contents to the weak cl uster nodes. After the assignment, QS trans mit these replication pattern information to the origin server OS. OS performs the replication pl acement, accordin g to the pattern inform ation obtai ned from QS. The weight val ue Wi of each node is stored along with the content. OS then broadcasts th e replication in formation to th e respective clients in the fo llowing format {Nid, Clid (“S” or “W”), c1, c2 …} Where Nid is the n ode id, Clid is the clus ter id and c1, c2… are content database ids. 4.3 Query Search and Data Retri eval A route discovery algorithm is neede d to determ ine if and where the requested ite m is rep licated when the requester does not have knowle dge of the destinati on. By reducing the comm unication costs, the speed and efficiency of the informati on retrieval mechanism can be improved. So, t he number of messages exchanged between the nodes and the number of cl uster nodes that are queried for each query request, are to be m inimized. For this, a robust searching algorithm is proposed. In this algorith m, each node main tains a profile wh ich contains the detail s of queries processed by i ts neighbors, within the last t seconds . Node Id (Ni) Query Id (Qid) Query Hits (Qhit) and No. of Results (NoR) This profile info rmation is then used to forward the queries to the neighbors who are having m ore chances of replying to those queries. In order to f orward a query Q, t o its neigh bors, a node N 1 assigns a score to each of its neighbors based on their profile. To calculate the score of each node Nj, (j= 2, 3…) N1 compares .Q with all queries sto red in Nj’s profile. If there is a query hit for Q, then the scor e of Nj can be calculated as m Score (Nj,Q) =  NoR(Nj,Qk) α k=0 Where NoR (Nj, Qk) is the number of results returned by Nj for query Q k, which are sim ilar to Q. So the nodes w hich return more results get the higher score. If α allows us to add more weig ht to the most similar queries. For example, when α is large, then the query with the largest similarity NoR (Nj, Qk) dom inates the formula. If we set α = 1, all queries are equall y counted, whereas setting α =0 allows us t o count only the number of results returned by each peer. (i) When a data requ est is initiated at a clie nt, it first looks for the data item in its own cache (local hit). If there is a local cache miss, the client sends the requ est to the set of stro ng cluster nodes. (ii) On receiving the request, each strong cluster node which has th e requested con tent, will send an ack p acket to the query client to ackno wledge that it has the data ite m. The ack packet will contain the following fields: time stamp Ts and weight W. Th e time stamp field h elps to choose th e latest copy of the searched item and the wei ght value field helps t o choose the best client node. (iii) When the quer y client receive ack packets from the strong cluster, it s elects the best n ode Sbest with max ) , ( W T s and sends a confirm packet to the c lient Sbest. The ack packets for the sam e item received from other nodes are discarded (iv) When the node Sbest receives a confirm packet, it responds back with the actu al data value to th e requested query node. (v) Suppose if the req uested data is not a vailable in any strong cluster no des, the request is directed to the server from the query client. Then the necessary data is sent to the clien t from the server. If the client ha s the available memory size (MZ) and bandwidth (BW), then it caches the data in its buffer. Then it is also considered as a strong cluster node and it is propagated to other nodes as {Nid, Clid (“S”), d1} Where Nid is the no de id, Clid is th e cluster id and d1… is the content database id . (vi) Subsequ ently if the same data is required for any oth er client, then it sen ds its request to this s trong cluster nod e which caches the data and receive s the required data. Caching frequen t data which is not foun d in the replica, into the local cache of a node increases the query efficiency and decreases the late ncy significantly. We now summarize the above steps into the fo llowing algorithm. Algorithm 1. The Node N1 get s a query Q from a client C. 2. N1 compares Q with Qi d kj , where Qid kj is the k th query of nod e Nj. k=1, 2, 3…. and j= 1, 2….. 3. If Qhit k > 0 , then 4. Score(Nj,Q) =  NoR(Nj,Qk) α i. k=0 ISSN : 0975-3397 S.Ayyasamy et al /Internatio nal Journal on Computer Sci ence and E ngineering Vol.1(2), 2009, 71-77 75 5. Select the Nodes Nj where Score (Nj, Q) is maximum. 6. Forward the query Q to Nj. 7. Nj sends ACK to cl ient C. 8. C selects the node Sbest fro m Nj, with max ) , ( W T s . 9. C sends a confirm packet to Sbest. 10. Sbest send requested data fo r query Q, to C 11. If Qhit k > 0 for all j, then 12. C send query request Q to Se rver. 13. Server sends requested data f or query Q, to C 14. If MZ C > min(MZ) and BW C > min(BW) (Where MZ C – Memory size of C and B W C – i. Bandwidth of C) then 15. C caches the data item. 16. C becomes a strong cluster node. 17. C propagates {Nid , Clid (“ S”) , d1 } to other nodes (Where Nid is the n ode id, Clid is the cluster id and d1 is the cont ent database id). V. E XPERIMENTAL R ESULTS 5.1 Simulation Setup This section deals with th e experimental performance evaluation of our al gorithm s through simulat ions. In order t o test our protoc ol, the NS2 sim ulator is used. NS2 is a general- purpose sim ulation tool t hat provides di screte event sim ulation of user defined networks. We have used the Bittorrent p acket-level simu lator for P 2P networks [13]. A net work topology i s only used for the pac ket- level simulato r. Based on the assumption that the bottlen eck of the network is at the access links of the users and not at the routers, we use a sim plified topol ogy in our s imulations. We model the network with the help of access and overlay links. Each peer is connected w ith an asymmetric link to its access router. All access routers are connected directly to each other modeli ng only an overlay link. This enable s us to simulate different upl oad and downl oad capacities as well as different end-to-end (e 2e) delays bet ween different peers. Figure 1. Topology of P2P Overlay Network 5.2 Simulation Resul ts We have compared our QIRM ar chitecture to Virat, a node capability aware P2 P middleware [11] architecture for managing replicas in l arge scale distributed sy stems. Based On Load In our initial exp eriment, the load of the requested content is varied from 2.0mb to 5 .0 Mb. The response delay i n seconds and received throughput in packets are measured. In Figure 2, we can see that, when the lo ad increases, the delay also increases. It is eviden t that the delay of QI RM is significantly less than the delay of VI RAT. Figure 3 shows the aggregated th roughput of all the client nodes which obtain ed their resp ective share of data. From the figure we can see that the QI RM has more throughput than VIRAT and the thro ughput valu es are decreasing, when the load increases. Lo ad v s D e la y 0 0.2 0.4 0.6 0.8 1 1.2 2.0 2.5 3.0 3.5 4.0 4.5 5.0 Loa d(M b) Del ay VI R AT QI R M Figure 2. Load Vs Delay (s) Load Vs Thr oug hput 0 200 400 600 800 2.0 2.5 3.0 3.5 4.0 4.5 5.0 Load(M b) Thr oughput(pk ts ) VI R AT QI R M Figure 3. Load Vs Throughput Based On Rate In our second experiment, the query sendi ng rate is varied from 250Kb t o 1Mb. The response delay i n seconds and q uery efficiency are measured. Query efficiency is a measure of the percentage of data queries that get served during an enti re simulation. ISSN : 0975-3397 S.Ayyasamy et al /Internatio nal Journal on Computer Sci ence and E ngineering Vol.1(2), 2009, 71-77 76 Figure 4 shows that the average query efficiency of the client nodes increases when the rate is increased. From the figure, we can see that the QI RM has m ore efficiency than VIRAT. Que ryE f fici e ncy 0 20 40 60 80 100 250 500 750 1000 R ate (k b) ef f ici en cy( % ) VI R AT QI R M Figure 4. Rate Vs Packet Delivery F raction E nd-to- E nd D e la y 0 0.1 0.2 0.3 0.4 0.5 0.6 200 400 600 800 R ate(k b) Del a y(s ) VI R AT QI R M Figure 5. Rate Vs Delay I n F i g u r e 5 , w e c a n o b s e r v e that, when the rate increases, the delay remains al most constant. From the figure, it can be seen that the delay o f QIRM is sign ificantly less than the delay of VIRAT. In Figure 6, the throu ghput against rate i s shown. From the figure, we can see that the thro ughput of QIRM is more wh en compared to VIRAT, and incr eases when rate increases. In Figure 7, the band width uti lization of client s against the rate is shown. From the figure, we ca n see that, bandwi dth utilization o f QIRM is nearly 80-90 %, when compared to VIRAT, which is 60-70%. R ate V s Thr oughput 0 0.2 0.4 0.6 0.8 1 1.2 250 500 750 1000 r a te (k b) Thr oughput(Mb/ s) VI R AT QI R M Figure 6. Rate Vs Throughput Ra t e V s Ba n d w i d t h Ut i l z a t i o n 0 20 40 60 80 100 120 250 500 750 1000 ra t e ut iliz at ion ( % ) VI R AT QI R M Figure 7. Rate Vs Utilization VI. C ONCLUSION A QoS based overlay net work architecture including an intelligent rep lica placement algo rithm is used to i mprove the network utiliza tion and th e fault tolerance of th e P2P and also to reduce the search latency. Ba sed on th e weight vecto r which includes available capacity, CPU speed, and memory size and access latency the nodes are cla ssified into strong and weak clusters. Based on the access pattern the content is classified into class I or class II by the replica m anagement algorithm. Then class I contents are replicat ed into strong grou ps for more copies. Routing is p erformed only to the stro ng clusters through broad casting the query hi erarchically. In addition to the replica placement, it also has a caching technique, to reduce the search lat ency. Low bandwid th Consum ption, Reduced Latency, Reduced Maintenance Cost , Strong Connectivity and Query Coverage are achiev ed in the proposed architect ure. Thus we have shown that our proposed architecture attains less la tency and better throughput with reduced network band width usage, throu gh the simulation results. References [1] [1] Stephanos, Routsellis-Theotokis and Dio midis Spinellis, “A Survey of Peer-to-Peer Content Distribution Technologies”, ACM Computing Surveys, Vol. 36, No. 4, December 2004, pp. 335–371. [2] [2] Xueyan Tang Jianliang Xu “On Replica Placement for QoS- Aware Content Distribution” INFOCOM 2004. Tw enty-third AnnualJoint Conference of the IE EE Computer and Communications Societies, 7-11 March 2004. ISSN : 0975-3397 S.Ayyasamy et al /Internatio nal Journal on Computer Sci ence and E ngineering Vol.1(2), 2009, 71-77 77 [3] [3] David Hales, Andrea Marcozzi and Giovanni Cortese, "Towards Cooperative, Self-Organised Replic a Manageme nt", in proc. of First IEEE International Conference on Se lf-Adaptive and Self-Organizing Systems, pp: 367-370, 9-11 July 2007, Doi:10.1109/SASO. 2007.62. [4] [4] B. Mortazavi and G. Kesidis, "Cumulative Reputation Systems for Peer-to-Peer Content Distribution", in proc. of IEEE Annual Conference on Information Sciences and Sy stems, pp:1546-1552, 22-24 March 2006, Doi: 10.1109/CISS. 2006.286385. [5] [5] Raphael Chand and Pascal Felber , "Semantic Peer-to-Peer Overlays for Publish/Subscribe Networks”, in proc. of International Euro-Par conference, vol.3648, pp: 119 4-1204, 2005. [6] [6] Anwitaman D atta,"MobiGrid: Peer-to-Peer Overlay and Mobile Ad- Hoc Network Rendezvous - a Data Ma nagement Perspective”, in proc. of 15th Conference on Advanced Inform ation Systems Engineering, 16- 20 June 2003. [7] [7] Paraskevi Raftopoulou and Euripide s G.M. Petr akis, "iCluster: a Self-Organizing Overlay Network for P2P Infor mation Retrieval", in proc. of ECIR, 30 Marc h - 3 April 2008. [8] [8] P. Krishnan, D . Raz, and Y. Shavitt, “The cache location proble m,” IEEE/ACM Transactions on Networ king, vol. 8, no. 5, pp. 568–582, Oct. 2000. [9] [9] L. Qiu, V. N. Padmanabhan, and G. M. Voelker, “On the placement of web server replicas,” in Proceedi ngs of IEEE INFOCOM’01, Apr. 2001, pp. 1587–15 96. [10] [10] Carvalho, N. Araujo, F. Rodrigues, L. “Scalable QoS-Based Event Routing in Publish-Subscribe Systems” Network Computing and Applications, Fourth IEEE International Symposium on Publication Date: 27-29 July 2005. [11] [11] A Vijay Sri nivas and D Janakiram, "Node Capability Aware Replica Management for Peer-to-Peer Grids”, Repot Accepted by IE EE Transactions SMC Part A, October 29, 2008. [12] [12] Guillaume Pierre and Maarten van Steen," Globule: A Collaborative Content Delivery Netw ork", in proceedings of IEEE Communication Magazine, vol.44, no.8, August 2006, Doi:10.1109/MCOM.2006.1678120. [13] [13] Kolja Eger, Tobias Hopfeld, Andreas Binzenhofer," Efficient Simulation of Large-Scale P2P Networks: Packet-level vs. Flow-level Simulations", in proceedings of 2nd Workshop on the Use of P2P, GRID and Agents for the Development of Content Networks, pp: 9-16, June 2007. [14] [14]. Yan Chen, Randy H. Katz a nd John D. Kubiatowicz"Dynam ic Replica Placement for Scalable C ontent Delivery", Lecture Notes In Computer Science; Vol. 2429, pp.306 -318, 2002. [15] [15]. Jian Zhou, Xin Zhang, Laxmi Bhuyan and Bin L iu,"Clustered K- Center: Effective Replica Placement in Peer-to -Peer Systems", IEEE conference on Global Telecommunica tions, pp.200 8-2013 Nov. 2007. [16] [16]. Kan Hung Wan, Chris Loeser, "An Overlay Network for Replica Placement within a P2P VoD Networ k”, Journal Of High Performance Computing And Networking, Vol.3, No.5/6,pp.320-335,Decem ber 2005. [17] [17]. Spiridon Bakiras and Thanas is Loukopoulos, “Combining re plica placement and caching techniques in content distribution networ ks", Computer Communications Vol.28, pp.1062 –1073, 2005. [18] [18]. Jie Su and Douglas Reeves, "Replica Placement Algorithms with Latency Constraints in Content Distribution Networks", T echnical Report, 2004. About the Authors Mr.S.Ayyasamy completed his B.E. (Electronics and Comm unication Engineering) i n 1999 from M aharaja Engineering Col lege and M.E. (Com puter Science and Engineering) i n 2002 from PSG College of Technolo gy, both under Bharathiar University, Coimbatore. Currently he is purs uing PhD degree from Anna University, Coimbat ore. He is working as a Assist ant Professor, Depart ment of Inf ormation Technology at Tam ilnadu College of Engineering, Coimbatore. He is a member of CSI. His research areas include P2P ne tworks, Overla y Networks, Cloud comput ing and Quality of Services and having 8 y ears of teaching experience in Engineering Colleges. Dr. S. N. Siv anandam completed hi s B.E. (Electrical Engin eering) in 1964 from Governme nt College of Technology, Coimbatore, and MSc (Engin eering) in Power Systems in the year 1966 from PSG College of Techno logy, Coimbatore. He acquired PhD in cont rol system s in 1982 from Madras University. He recei ved best teacher award in the year 2001 and Dhakshina Mu rthy Award for teaching excellence from PSG College of technology. He received the citation for best teaching and tech nical contribution in the year 2002, Govern ment Coll ege of Technology , Coimbatore. His research areas include Mode ling and Simulation, Neural Networks, Fuzzy Syste ms and Genetic Algo rithm, Pattern Recognition, Multidimensional syste m analysis, Linear and Non linear control system, Signal and Image processing, Control System, Power System, Num erical methods, Par allel Computing, Data Mining and Database Security. He is a member of various p rofessional bodies li ke IE (India), ISTE, CSI, ACS and SSI. He is a t echnical advisor for various reputed indus tries and engineerin g institutions . ISSN : 0975-3397