Vol. 02, No. 07, 200, 2227-2233 Traffc Engneerng wthout Congeston Hot Spots n MPLS A.Padmaprya(Author) Assstant Professor Department of Computer Scence & Engneerng Alagappa Unversty, Karakud, Taml Nadu, INDIA L. Rajeswar (Author) Programmer Computer Centre, Alagappa Unversty, Karakud, Taml Nadu, INDIA Abstract Traffc Engneerng [] broadly relates to optmzaton of the performance of the operatonal IP network. In networkng, network congeston occurs when a lnk or node s carryng so much data that ts qualty of servce deterorates. Typcal effects nclude queueng delay, packet loss or the blockng of new connectons. A consequence of these latter two s that ncremental ncreases n offered load lead ether only to small ncreases n network throughput, or to an actual reducton n network throughput. Ths paper dscusses methods lke PNP approach [2] and HITS method for mprovng QoS [3], whch are used for traffc engneerng n MPLS. Ths paper wll examne the two approaches; dscuss solutons n both PNP approach and HITS method for mprovng QoS and pont to topcs for research and advanced development. Keywords-component; Traffc Engneerng, Congeston, PNP approach, HITS, QoS I. INTRODUCTION The unprecedented growth of the Internet has lead to a growng challenge among the ISPs to provde a good qualty of servce, acheve operatonal effcences and dfferentate ther servce offerngs. ISPs are rapdly deployng more network nfrastructure and resources to handle the emergng applcatons and growng number of users. Traffc Engneerng [4] has drawn much attenton n recent years. Two mportant components of traffc engneerng are traffc estmaton and routng. A good understandng of the nterplay between these two nter-related components wll make sgnfcant contrbuton to network management and performance. Awduche et al [] note that a dstnctve functon performed by Internet Traffc Engneerng s the control and optmzaton of the routng functon, to steer traffc through the network n the most effectve way. Load balancng s an mportant approach to address network congeston problems resultng from neffcent resource allocaton [6]. A routng specfes how to route the traffc between each Orgn-Destnaton par across a network. The objectve n desgnng a routng s to provde good qualty of servce and to optmze the utlzaton of network resources. Measurng or estmatng traffc demands accurately s non-trval. Desgnng a routng robust to changng and uncertan traffc s desrable. Common objectves of Traffc Engneerng nclude balance traffc dstrbuton across the network and avodng congeston hot spots. To meet the objectves, demands have to be placed over the lnks n order to acheve balanced traffc dstrbuton and to avod congeston hot spots n the network. A technque for montorng network utlzaton and manpulatng transmsson or forwardng rates for data frames to keep traffc levels from overwhelmng the network medum. The assumpton that statstcal multplexng can be used to mprove the lnk utlzaton s that the users do not reach ther peak rate values smultaneously, but snce the traffc demands are stochastc and cannot be predcted, congeston s unavodable. Whenever the total nput rate s greater than the output lnk capacty, congeston occurs. When the network becomes congested, the queue lengths may become very large n a short tme, resultng n buffer overflows and cell loss. Congeston control s therefore necessary to ensure that users get the negotated Qualty of Servce (QoS) [7]. We provde the background n Secton II. In Secton III we study the performance of both PNP approach and HITS method for mprovng QoS. We compare these two methods n Secton IV. Then we draw to conclusons n Secton V. II. BACKGROUND A. Graphcal Models Bayesan Networks Graphcal models are nothng but fuson of probablty theory and graph theory. They provde a natural tool for dealng wth two problems that occur throughout appled mathematcs and engneerng uncertanty and complexty and n partcular they are playng an ncreasngly mportant role n the desgn and analyss of machne learnng algorthms. Probablstc graphcal models [8, ] are graphs n whch nodes represent random varables, and the arcs represent condtonal ndependence assumptons. Hence they provde a compact representaton of jont probablty dstrbutons. Undrected graphcal models, also called as Markov Random Felds (MRFs) or Markov networks, have a smple defnton ISSN : 097-3397 2227
Vol. 02, No. 07, 200, 2227-2233 of ndependence: two (set of) nodes A and B are condtonally ndependent gven a thrd set, C, f all paths between the nodes n A and B are separated by a node n C. By contrast, drected graphcal models also called Bayesan Networks or Belef Networks (BNs), have a more complcated noton of ndependence, whch takes nto account the drectonalty of the arcs. For a drected model, we have to specfy the Condtonal Probablty Dstrbuton (CPD) at each node. If the varables are dscrete, ths can be represented as a table (CPT), whch lsts the probablty that the chld node takes on each of ts dfferent values for each combnaton of values of ts parents. B. Identfcaton of Congeston Hot Spots usng Bayesan approach Currently the so-called Bayesan network approach [9] s used to dentfy the congeston hot spots n MPLS. Wth ths approach, n addton to the network topology, t s necessary to specfy the parameters of the nodes n the network. For each node n the network the Condtonal Probablty Dstrbuton (CPD) [0] s specfed. Let us llustrate ths wth a smple example. A traffc demand and capacty of the network are selected. The Condtonal Probablty Dstrbuton for the fve nodes s dsplayed n the followng tables. The new Bayesan approach s proposed for achevng traffc TABLE I CPD FOR THE NODE A P(A=0) P(A=) 0. 0. TABLE II CPD FOR THE NODE D A P(D=0) P(D=) 0 0 0.0 0.99 TABLE III CPD FOR THE NODE G D P(G=0) P(G=) 0 0 0.9 0. TABLE IV CPD FOR THE NODE C D P(C=0) P(C=) D 0 0 0. 0.9 G C TABLE V CPD FOR THE NODE F E G C P(C=0) P(C=) Fgure. Bayesan Network for fndng Congeston Lkelhood The network represented n Fg. conssts of fve nodes each havng lnk capacty of unts. If the nodes n the network are dscrete the CPD can be represented as a table (CPT), whch lsts the probablty that the chld node takes on each of ts dfferent values for each combnaton of values of ts parents. Here the chld nodes are destnaton nodes and parent nodes are source nodes. Optmal routes havng mnmum congeston lkelhood can then be calculated usng the followng formula. Pe ( R rpr ) ( r) P( R r e) Pe ( ) () where P(R = r e) denotes the probablty that random varable R has value r gven evdence e. Instead of routng the demands over the congested routes, routes that sut the current engneerng n the backbones. Instead of relyng on the mappng of logcal connectons of physcal lnks to manage traffc flows n the network, we run IP routng natvely over the physcal topology, and control the dstrbuton of traffc flows through settng approprate lnk weghts for shortest path routng. III. 0 0.0 0.0 0 0.0 0.99 0 0.2 0.8 0.0 0.99 TRAFFIC ENGINEERING WITHOUT CONGESTION HOTSPOTS Multprotocol Label Swtchng (MPLS) s a mechansm n hgh-performance telecommuncatons networks whch drects and carres data from one network node to the next. MPLS makes t easy to create "vrtual lnks" between dstant nodes. It can encapsulate packets of varous network protocols. ISSN : 097-3397 2228
Vol. 02, No. 07, 200, 2227-2233 MPLS s a hghly scalable, protocol agnostc, data-carryng mechansm. In an MPLS network, data packets are assgned labels. Packet-forwardng decsons are made solely on the contents of ths label, wthout the need to examne the packet tself. Ths allows one to create end-to-end crcuts across any type of transport medum, usng any protocol. The prmary beneft s to elmnate dependence on a partcular Data Lnk Layer technology, such as ATM, frame relay, SONET or Ethernet, and elmnate the need for multple Layer 2 networks to satsfy dfferent types of traffc. MPLS belongs to the famly of packet-swtched networks. A. PNP Approach The Pragat Node Popularty (PNP) approach dentfes congeston hot spots n MPLS wth mnmum efforts when compared to Bayesan Network approach. Pragat Node Popularty (PNP) s a numerc value that represents how popular a node s on the gven network. When one node lnks to another node, t s effectvely castng a vote for the other node. The more votes that are cast for a node, mples that the node s more popular among all other nodes n the network. PNP value s the way of fndng a node s popularty. Wth the help of ths PNP value, the congeston hot spots n the network can be dentfed. Let dgraph G = (V, E) represent the IP network, where V s the set of nodes and E s the set of lnks. Please note that the lnks and ther capactes are drectonal,.e. lnk j s consdered dfferent from lnk j, each wth ts own capacty. IN() and OUT() denote the number of edges nto and out of node respectvely. To calculate the Pragat Node Popularty (PNP) value of a node, all of ts nbound lnks are taken nto account. A PNP value of node, s defned as PNP() ( ) (2) where δ s a constant value whch can be set between 0 and, and γ s the share of the PNP value of every node that lnks to node. PNP( j) node" j" edge : j, (3) OUT ( j) Mn PNP() (4) where Tk, k, S D Subject to V, PNP( ) 0 () V, PNP( ) N (6) The objectve functon (4) s to mnmze congeston by selectng a route wth nodes havng least PNP values as ntermedate nodes. Constrants () and (6) are node popularty conservaton constrants. Equaton () says that, nodes havng zero PNP value must not be consdered for routng process. Because they ndcate that the path has been broken down. Equaton (6) says that the net popularty value of a network s equal to the total number of nodes n the network. B. HITS method to mprove QoS Gven a network topology and exstng traffc flows, the Hts algorthm can be used to dentfy the congested nodes n the network. The nodes n the IP network can be consdered as the pages of HITS algorthm. Jon Klenberg's algorthm called HITS (Hyperlnk Induced Topc Search) [] dentfes good authortes and hubs for a topc by assgnng two numbers to a page : an authorty weght a, and a hub weght h. These weghts are defned recursvely. Pages wth a hgher a number are consdered as beng better authortes, and pages wth a hgher h number as beng better hubs. a h v hu (7) u v E u av (8) u v E Let A be the adjacency matrx of the graph G, A t be the transpose of the matrx and v be the authorty weght vector and u be the hub weght vector. Then, k In equaton (3) the share means the lnkng node s PNP value dvded by the number of out bounds lnks on the node. So the PNP value s determned for each node ndvdually. Further, the PNP value of node s recursvely defned by the PNP of those nodes whch lnk to node. Let K be the set of pont to pont demands. For each k K, let S k, D k, T k be the source node, destnaton node and ntermedate node respectvely. Let N be the total number of nodes n the network. Then the reactve routng usng PNP approach can be formulated as where, v A t. u (9) u Av. (0) ISSN : 097-3397 2229
Vol. 02, No. 07, 200, 2227-2233 h h 2 u... hn and a a 2 v... an The ntal hub and authorty weghts of the nodes are, u v 0 0.. t A... () (2) After k steps, the authorty and hub weghts are calculated usng The objectve functon s to mnmze congeston by selectng a route wth nodes havng least hub and authorty weghts. Here n equaton () only ntermedate nodes are taken nto consderaton. Equatons (6) and (7) say that the authorty and hub weghts may be ether zero or some postve value. It never takes negatve values. Equaton (8) says that the number of ntermedate nodes n a path may be zero or less than the number of ntermedate nodes n the path. The above LP formulaton s not restrcted for the gven example network but they can be generalzed to any network topology. IV. COMPARISON BETWEEN PNP AND HITS Let us consder the network shown n Fg. 2. It shows a smple network topology, lnk weghts, and traffc demands. Each lnk has a capacty of unts and each demand needs bandwdth of 4 unts. The lnk capactes, lnk weght and traffc demands are drectonal n IP networks. Snce the network here s rather small, the process of traffc engneerng can be done manually. The optmal routes for achevng balanced traffc dstrbuton are as follows. By usng equal-cost load balancng n the OSPF routng protocol [2], 3 A 3 v t k ( A. A). vk (3) F 2 B u ( AA. t ). u (4) k k D Let dgraph G = (V, E) represent the IP network, where V s the set of nodes and E s the set of lnks. Please note that the lnks and ther capactes are drectonal,.e. lnk j s consdered dfferent from lnk j, each wth ts own capacty. Let A represents the Adjacency matrx of the IP network. The hub and authorty weghts (h and a respectvely) are represented by sngle column matrxes. For any gven traffc flow S be the source of the flow, D be the destnaton of the flow and ntermedate nodes are denoted by Ij. Then the congeston hot spots dentfcaton usng HITS algorthm s formulated as follows. Subject to Mn I : Ij Ij j I a h () j aij 0 (6) hij 0 (7) 0 j n (8) G E A->B: 4 A-B A->F : 4 A-F B->F : 2 B-C-D-G-F B->F : 2 B-C-E-G-F A->E : 2 A-D-G-E A->E : 2 A-D-C-E F->B : 2 F-G-D-C-B F->B : 2 F-G-E-C-B Fgure 2. Topology, Lnk Weghts and Traffc Flows Demand A to B uses path AB, and demand A to F uses AF. Demand B to F has two paths. Half of the demand goes over BCEGF and the other half over BCEGF. BCDGH and BCFGH appear as equal-cost paths, so routng protocols such as OSPF wll perform load sharng over them. Smlarly for A to E. Half of the demand traverses path ADCE and the other half traverses through ADGE. Optmal routes and traffc dstrbuton are gven n Fg. 2. A. Calculaton of PNP values Because of the sze of the nodes n the Internet backbones, the Pragat Node Popularty (PNP) approach uses an approxmate, teratve computaton of PNP values. Ths means C ISSN : 097-3397 2230
Vol. 02, No. 07, 200, 2227-2233 that each node s assgned an ntal startng value and the PNP values of all nodes are then calculated n several computaton crcles based on the equatons (2) and (3) determned by the PNP approach. The teratve calculaton s llustrated usng the example network n Fg.2, whereby each page s assgned a startng PNP value of and constant value as 0.8 to strke better results. The PNP values for Fg. 2 are lsted n Table. I. TABLE I PNP VALUES FOR THE NETWORK IN FIGURE 2 Iteraton Pragat Node Popularty values A B C D E F G 0.000.000.000.000.000.000.000 0.0 0.77.80.000 0.77 0.77.80 2 0.0 0.77.489.24.98 0.77.489 3 0.0 0.64.796.036 0.994 0.64.796 4 0.0 0.70.3.20.68 0.70.3 0.0 0.627.77.062.020 0.627.77 6 0.0 0.690.68.88.4 0.690.68 7 0.0 0.637.728.08.039 0.637.728 8 0.0 0.682.92.72.29 0.682.92 9 0.0 0.644.708.09.02 0.644.708 0 0.0 0.676.60.60.8 0.676.60 0.0 0.649.693.0.062 0.649.693 2 0.0 0.672.622.2.09 0.672.622 3 0.0 0.62.683.2.069 0.62.683 4 0.0 0.669.63.46.03 0.669.63 0.0 0.6.67.7.074 0.6.67 6 0.0 0.667.638.42.099 0.667.638 7 0.0 0.67.669.2.078 0.67.669 8 0.0 0.66.642.38.096 0.66.642 9 0.0 0.68.66.23.08 0.68.66 20 0.0 0.664.646.36.094 0.664.646 It s seen that good approxmaton of the real PNP values can be calculated usng few teratons alone. It s clear that the nodes C and G are the most congested node n the network. Because three dfferent traffc flows, AE, BF and FB are usng these two nodes. Next to them node D s the most congested one. Lkewse t can easly predct the congested nodes n any gven network. B. Calculaton of Hub and Authorty values usng HITS The nodes n the IP network can be consdered as the pages of HITS algorthm. The Source nodes S are consdered as Hubs and the Destnaton nodes D are consdered as Authortes. The Intermedate nodes I between any source and destnaton may act ether as a hub or as an authorty. Let us frst llustrate wth a smple example how the HITS algorthm can be mplemented n a network. Let us consder the network shown n Fg. 2. The vertces and the edges are represented below. Here edges here G V ( V, E) ( A, B, C, D, E, G, F ) E ( AB, AF, AD, BC, CD, CE, DC, DG, EG, GE, GF ) Here S={A, B,F}, D={B,E,F} and I={B,C,D,E,G} 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 A 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 Usng (), (2), (3) and (4), we get, 8 3 6 0 0 9 a0 0 h0 47 9 0 0 28 9 0 Then I s calculated usng equaton (). j 70 0 6 I 6 0 70 Accordng to HITS the congeston ndex s hgh for the node D. From the calculatons t s clear that when we take D as ntermedate node the congeston wll be very hgh. Next to that there comes A and G. Lkewse we can predct whether the path s congested or not usng HITS. C. Comparson between PNP and HITS congeston ndex For example, let us consder n Fg. 2, there comes a new flow from EA. Multple paths are avalable from EA. They are ECDA, ECBA, EGDA and EGFA etc. ISSN : 097-3397 223
Vol. 02, No. 07, 200, 2227-2233 Whle applyng the PNP approach, the objectve functon (4) wll select the route ECBA. Because, t s the one havng least net PNP value when compared wth the remanng routes. All other routes are havng hgher net PNP values. So the path ECBA, can be chosen for data transmsson between node E and A. It s llustrated n Table II. TABLE II PATH SELECTION USING PNP APPROACH Feasble Paths Summaton of hub and authorty values of Intermedate nodes E-C-B-A.646+0.664=2.30 E-C-D-A.646+.36=2.782 E-G-D-A.646+.36=2.782 E-G-F-A.646+0.664=2.30 If we calculate the congeston ndex usng HITS method of mprovng Qualty of Servce, the objectve functon () wll select the route ECBA. Because, t s the one havng least hub and authorty weght when compared wth the remanng routes. So the path ECBA, wll be selected for data transmsson between node E and A. TABLE III HUB AND AUTHORITY CALCULATION Feasble Paths Summaton of hub and authorty values of Intermedate nodes E-C-B-A 6+0=06 E-C-D-A 6+=67 E-G-D-A 70+=8 E-G-F-A 70+0=20 From the above calculaton t s clearly shown that both the methods wll yeld same result. For makng comparson wth PNP values, the HITS congeston ndex s converted to numbers wth two decmal places. Values 2.000.00.000 0.00 0.000 Comparson between PNP and HITS A B C D E F G PNP Nodes Fgure 3. Comparson chart V. CONCLUSION NEW HITS We observe that both the PNP approach and HITS method of mprovng QoS, wll effectvely dentfy the congeston hot spots n operatonal IP networks. Both the methods can dentfy the congeston hot spots wth no knowledge of lnk weghts and shortest paths but wth the help of the network topology alone. Furthermore, t overcomes the draw back on N-square problem of Bayesan approach. Smlar approaches have been tred by some servce provders n the past. When a lnk s experencng congeston, servce provders typcally ncrease the weght for that lnk n the hope that traffc wll be moved away from t. These experments, however, were done based on smple heurstcs. The lack of systematc strategy and comprehensve studes of lnk weght change mpact has prevented t from beng wdely adopted n operatonal backbone networks. Whenever there s new traffc demand and path selecton procedure by ether PNP approach or HITS method of mprovng qualty of servce, the IN and OUT values of the nodes and the adjacency matrx must be updated accordngly. The updaton s made n order to reflect the changes n traffc flows n the network. The complexty of the HITS method s greater when compared to that of PNP approach. Irrespectve of the network topology these methods dentfy the congeston hot spots effcently and acheve the common goal of Traffc Engneerng []. Wth the help of PNP approach or HITS method of mprovng QoS, we not only fnd the congested nodes n the network, but also fnd a better route than the congested ones. Even though the path suggested by ether PNP approach or HITS method may be longer than the shortest path, but t reduces the queung due to congested routes. Thus, wth these methods we mprove the qualty of routng to almost perfecton there by avodng congeston hot spots n the network. REFERENCES [] D. Awduche et al, Overvew and prncples of Internet Traffc Engneerng IETF, Internet draft, draft-etf-tewg-prncples-02.txt, Jan 2002. [2] Dr.E.Ramaraj and A. Padmaprya, Pragat Node Popularty Approach to Identfy Congeston Hot Spots n MPLS, Proc. ICCESSE 200, Internatonal Conference on Computer, Electrcal and System Scences and Engneerng, France, Issue 68, pp. 628-632, July-200. [3] Dr.E.Ramaraj and A. Padmaprya, Congeston Hot Spots Identfcaton usng HITS and Improvng Qualty of Servce, n Internatonal Journal of Computer Scence and Network Securty (IJCSNS), August 200, n press. [4] Yufe Wang, Zheng Wang, Leah Zhang, Internet traffc engneerng wthout full mesh Overlayng, Infocom 200. [] ATM Forums, Prvate-Network-Network-Interface Specfcaton Verson.0 (PNNI.0), Internatonal Standart-of-pnn-oo.000, March 996. [6] D. Awduche, J.Maledm, J. Agogbua, M.O Dell and J. McManus, Requrements for traffc Engneerng over MPLS, IETF RFC 2702, Sep 999. [7] Fang Lu. ATM Congeston Control. http://www.cse.ohostate.edu/~jan/cs788-9/ftp/atm_cong/ndex.htm [8] S.Russell and P.Norvg. Artfcal Intellgence: A Modern Approach. Prentce Hall, Englewood Clffs, NJ,99. ISSN : 097-3397 2232
[9] Yufe Wang and Zheng Wang, Explct routng algorthms for nternet traffc engneerng, Proceedngs of ICCCN 99, Sept 999. [0] J Moy, OSPF verson 2, Internet rfc, IETF, May 998, Techncal report RFC 2328. [] Kevn Murphy, A bref ntroducton to Baye s rule, Sept 2000. Web: http://www.cs.ubc.ca/~murphyk/bayes/bayes.html. http://www.cs.ubc.ca/~murphyk/bayes/bayesrule.html. [2] C. Vllamzer, OSPF optmzed multpont, Internet draft draft-etfospf-omp-00.txt, IETF, March 998. [3] J. Klenberg. Authortatve sources n a hyperlnked envronment. Journal of the ACM, 46():604 632, 999. A. Padmaprya et. al. / (IJCSE) Internatonal Journal on Computer Scence and Engneerng Vol. 02, No. 07, 200, 2227-2233 AUTHORS PROFILE Mrs. A.Padmaprya s wth Department of Computer Scence & Engneerng, Alagappa Unversty, Karakud 630 003, Taml Nadu, Inda as Assstant Professor. She s currently pursung Ph.D n Computer Scence at Alagappa Unversty. She has been a member of IACSIT. She publshed 6 papers n Natonal and Internatonal conferences. Mrs. L. Rajeswar s wth Computer Centre, Alagappa Unversty, Karakud 630 003, Taml Nadu, Inda as Programmer. Her areas of nterest nclude Traffc Engneerng and Qualty of Servce and Data mnng. She s havng more than 20 years of servce wth Alagappa Unversty. ISSN : 097-3397 2233