You are on page 1of 13
Multicast Routing Algorithms and Protocols: A Tutorial Laxman H, Sahasrabuddhe and Biswanath Mukherjee, University of California Abstract Mlicsing sho billy of @ commnication network Yo accept o single message ftom on application ont fe dslverCopos of he message ro multe rciplen at different locations. Recently, there has been an explosion of research literature on ‘ruticos commnicotion. This work presents a torisksum survey of fhe vert suo roan lar ot lon th lea ug toe for packetawitched wide rea networks, Our conirbuion should be of paricular benefit othe generic networking audience fond, toa lesser exlen, to the expert on his sobeeh ultcasting iste ability of a communication (LANs) and wide area notworks (WANs). A LAN spans & network to accept a single message from an to deliver copies of the mes YF Ssgo to mutple recipionis at ditteront loca tions, One of the challenger is to miaimnize the amount of rtuark resources employed by multicasting, To illustrate this point, fetus assume that a video server wants (0 transmit movie to 1000 recipients (Fig. 1a). I the server were to ‘employ 1000 separate point-to-point connections (e.g, TCP connections), 1090 copies of the movie may have to be sent over a single link, thus making poor use af the available bundwidth, An eff ‘ent implementation of multicasting permits ‘ch better use of the available bankideh by transmitting at mast one copy of the movie on each fink in the network, as Shown in Pig 1b ‘Rocenty, thero has boon a lot of research the area of multicast communication, Alto ‘many excellent surveys and books exist which txamine various aspocts of miltcesting [1-6], in the courso of aur studies we have found need for a tutoril-cum-survey ot the various ‘nuldcast routing algorithm and their relati ship with multicast routing protocols. In this Work we present a tutorial-cum-survey of the following two important topics in multicasting: * Malticast routing algorithms * Multis routing protocals ‘Communication networks can be classified Into to categories: Tocal ares networks ‘Ti ook has eo suprted part ythe Natal Seize Foundation (NSF) under Grants os. NCR. DSUSD3S and ANE IR0528, Matias desnations 300 sos4nnis10.00 © 2000 TEE small geographical arca, typically a single building ora cluster of buildings, while a WAN spans a large geographical area (ea, anaiton). Often, nodes comnected toa LAN communi- sale over a broadcast network, while nodes connected t0 4 WAN communicate via a s¥itched network, In broadcast LAN, a transmission fram any one nade is received by al the nodes on the network; thus, sulticsting is easily implement- Sours Source ] || 1000 pot topo Single multiast connection “aommection Of luneastng movi to 1000 iffren ser b) maulicasting the raovie. (R= star- ‘trl rouder, MR= rutcast router) TEBE Nework «Janay ary 2000 (W Figure 2a) an example ofa WAN. Data is ransmitedjrom source Hos I to destination Hest 2 trough te route consising of nades CAL, UT, Ml, and NY; b)a directed raph that models the WAN shown ina). ed on a broadesst LAN, On the other hand, imple- ‘menting multicasting on a setched network is quite challenging; hence, throughout this work, we will focus on the multicasting problem in a WAN which is ased on a switched network, ‘Today, many multicast applications exist, such as news feed, ile distribution, interactive games, and videoconfereacing, but the implementation of these 'pplications is not necessurly efficient heeasse today’s WANS were designed to mainly support point-to-point (unieast) communication, tm the Tuure, as mallicast applications become more pop- ular and bandwidth-intensive, there will emerge @ pressing nced to provide efficent multicasting sup- port on WANS, "A WAN consists of nodes (ie., switebes or outers) interconnectad by communication links. A {ransmission from a source to.a destination is routed through these interconnected nodes. Figure 2a shows an example of a out of s transmission from a source to a destination on & WAN? ‘A WAN can be modeled by a dicted graph. Figure 2) shows a directed graph that models the communication network Shove in Fig 2, 4 directed graph consists of a Set of nodes anda set of inks. link connecting node w to node v i rep~ resented by an ordered tuple (u,v). Nodes in the directed raph represent nodes in the WAN, while links ia the deected [graph represent communication links in Uae WAN. (Note that the graph in Fig. 2p is special one inthe sense that if there is ‘link (0), there also exists lnk (9, ); this characteristic, however, snot «necessity in a general directed graph.) ‘Communication links in a network may bave different properties. For example, «fiber aptie communication link may have very large bandwidth compared to a copper wire communication link. A property of a communication link is tepresented by a weight Of the corresponding link in a graph. For example, ifthe propagation delay ofthe communication link (CA2,FX) i Tm tis information ean be represented by assigning a weight equal to 1 tothe link (C42,7) in Fig. 25, with the weights of the other links being thelr corresponding propagation delays in milliseconds “The communication links in Pig. 2can be of two ypes: sme "thera of work (a) we wae the ters “than "eat" mean; an) he combination of and he oe ‘sponding ith gh wich eras tothe never lb ated as “antacid and willbe efoto ana: Frenne, (source) ‘oa Hot | a sich CAT il be refered 0 a crc) nade CAL IEEE Network + Janaarebroary 2000 IFigue 3. An exp of Sener ve Mian group = (CAI, TK Tl NON Cost als 1 Castf Siar te =. ‘metric and asymmetric. Symmetric links have the same weight in both dircetions, while asymmetric links have different ‘weights depending on the dicection, Thus, in Fig. 2, which shows weights on only four links, the link between nodes CA2 fand 7X is symmetric, while the link between nodes 7X and [MD is asymmetrie, If all the links in a WAN ave symettic, ‘we can model the WAN by an undirected graph, a8 showin i Fig. 3. In an undirected graph, the direction ofa ink is unin portant; Renee, a link between node w and node » canbe Fop- Fesented by an unordered tuple (u, »). Traditionally, ccoramunication networks have been modeled by undirected araphs. Henceforth inthis work, unless otherwise stated, the term graph wil reter to an undirected graph, ‘Ix unicast (point-to-point) communication, souting is often treated es the shoreeaepath problem in apts, When tro nodes Wish to communicate, a mintmum-welght path (shortest path) connecting the corresponding pair of nodes i selected. In mule ‘casting, a group of more than two nodes (also called the mul- ticast group) ish to communicate with one another, Now, instead ofthe shortest path, we are fntersted in tae mma weight ee which spans all the nodes inthe multicast group. Tn general, different multicast applications have different soquitoments, For example, a reliable data transfer multicast application, such as software distribution, has very different requirements from 2 real-time multimedia multicast applica tion, such as nationwide videoconferencing, Ths, itis helpful to classify multicast communication into tw typos * Source specific: In souree-specific multicast communication, ‘only one node inthe multicast group sends data, wale al the other nodes receive dat, + Group-shared: In group-stiated multicast communication, ‘each node in the multicast group ean send data to the mul- a ticast group as well as receive data from other nodes in the multicast gromp, The next Section discusses multicast routing algorithms. We then study the imp ion of multicast routing protocols oon the Internet. Note that the current Internet uses IPv4, hile the next-generation Internet (NGT) will employ [PV6. Since some topics discussed are specifi to IPv4, they are not applicable to the NGI, although the general principles dis- cussed will stil be applicable. On the other hand, the subsec tions on multicast routing algorithms are relevant to both [Pus ‘and 1Pv6 because they do not presuppose any partiewlar net ‘work-layer protocol. Finally, we provide concluding remarks Muticast Routing Algorithms Figure 3 shows an undirected graph G = (VE), where Vis the set of aodes and E the set of links, Note tha, since graph G is undirected, it models a communication neiwork wh has symmetric links. Let M = (Cal, TX, IL, NY) be a multi ‘east group. (Shaded nodes in Fig. 3 Belong to the multicast {group.) Now, in order to perform multicast communication, the nodes inthe multicast group must be interconnected by & tree. Thus, she problem of multicast routing in communication networks is equivalent to finding a tree in graph G such that T spans al vertices inthe multicast group M. Such a tree is called 2 multicast tee and is shown in Fig. 3 by thick lines. (The term Steiner ee used in Fig. 3 will be claitied next.) Just as multicast communscation ean be of two types, multicast tuees ean also be classified into two corresponding categories souree-specifie (ar source-rooted) and group-shared., For the icast example as in Fig. 3, Fig. 4a shows a source- specific multicast tree which employs unidirectional links? (vith source = CAL), while Fig. 4 shows a group-shared ml ticast ee. The key diference between a sourcespecifc multi- cast tree and a group-shared multicast tree is that @ souree-specific multicast tree is optimized for source-specific ‘multicast communication, while a group-shared malticast tree {s optimized for group-shered multicast communication. For example, ifwe want 10 minimize the average delay for source- specific communication, we need to minimize the average sourcespecific delay which is calculated by taking the average Of the end-to-end delays overall (source, maliieast-member) pairs. Now, assuming tht each link in Fig, 4a has delay equal to I, the source-pecifc delay ofthe source-specific tee rooted 2 Throughout sis work he defale eight fl inks, ues specie ot erie requ 1 3 Ne haa ours muds ee connect a sur node ae es the ua roup yep ether nde or biden ks, wie group-hana macau eels nl bition Enki 2 at CAL i equal to 233 (the average ofthe delay fom source CA to nodes TH IL, and NY) In comparison, he surcespe- sii day (vith C1 a the sores) of he groupsated ml Gan tre shown n Fig equal t0323,00n fhe other hand, ite were to clenlate the average eroupshared delay of the souircespestic tes by taking the mverage ofthe endetovend ‘delays overall (rulticast-memiber, maticast-member) pairs, the avertgeproupatared deny ea 1038 in Fi, we the average sroup stated delay of he group shared ree in Fig. 3b is equal to 267. Thus, the application requirements dette stich pe of mutica trees are “beter “The following ia list of the properties of 2 ood matcast Since for most mukieast applications some prepertics are ‘more important han ars, we have divided the properties ‘nto tees priory levels! high meson, end lw High Puiorty + Tw cost: The cost (or weight) of «multicast tee is the ‘of the costs (or weights) ofall the links inthe multicast {eco A good mlicast ee tres to sinimize Ui ost Low delay: The end-to-end delay trom the sourge node to the destination node ithe sum ofthe todividal link delays along the route. A good multicast tee tees to minimize the endsto-end dele for every souree-destination pat in the ‘mutica group. Sealab A good multicast te i scalable in two respec First, consteting a multicast tzce for a large multicast group should requite reasonable amounts of time and Fesources. Second the switehes in the communication ete ‘work should be ale to simultaneously support large mam bor of multicast trees Mediu Priority + Support for dynamic multicast groups: Multicast groups can be clasiied ts staic and dynamic. The members oft static multicast group do not change overtime; in a dynamic mal cast group, now members may join oF existing members leave, A good multicast tce should allow multicast members to join of leave the multicast tree in a seamless fashion, ‘Moreover, the properties of a good! multicast tree should not degrade due to the dynamic nature ofthe multicast group, + Survivabliy: A good multicast tree should be able to vive multiple node and fink failures. Note tha tepals my be deren for cain applications, For xu, whie the fumes propeny of a muicas ree nt very impor tanta geerl maybe ema omponant propery of the malas ee ifthe multicast re s being employed by a mip game Moreover, ‘lug some properties are onsderefoepriony fir today's api hans, heyy Become more portant in theatre de to omer ‘application which maybe bend ou comprcension today TEE Network Januuryebruary 2000, low Prioviy ‘Fairness: A good multicast tree is fair in eo espees. it it ies to provide a minimum quality ‘service (eg, bounded delay) to each member in ‘he multicast group. (Its not Fai to unnceessarily ppunish one member in order to improve the quality Df service to other members.) Second, i tries (0 evenly divide the multicasting effort (c.g, packet ‘uplication effort!) among the participating nodes “Most algorithms that have been proposed in the literature mainly focus on cost and delay optimiza tion, although the other properties have also been dressed ta lesser extent. Before we examine each Gf the above properties in detail, [et us examine Some important theoretical concepts and definitions which will help us better undorstand the natuee of the multicast outing problem. “he cls optimicaton problem in mcs routing i called the Sie ee prolem nnetwors (SPN)y andi Sete sans Ged “Anandietedpsph = (%F) + oon fetid WE eg pov eles yt ink « Aid ot goes Me ¥ which belong the matics! group fd ave P= (ry Ey) which pe sah at eet Cr i minnie. Such smite las fee Jerre Nose ht se paphG unde imadel«communation network wc hs tocol inks ihm Stcer ce Pi ousted mules tee Tigre shows» Stier ee ch conc ult soup coating of sodes C41, Tf, and NY Ie are fhe tana each ote equal t ihe cost ofthe Steiner tre lb egal to Note tut nodes Ca and Pd do not belong to theta ao, but ae pao te Stee as Sec tases ae eal Se ney, Although SPN is NP-complete [7], there are some trivia cases SPN that an Be soho pebmoma ine shown below [1]: Tl 2 2 (anias ease: There are only wo node inthe Ilan group. SPN vedues rhe weltioown shores Prt probe, Plyomaltne signs forts proses fee known (3) + T= 17 reese: In hi case he mts group Contin Ye sox nthe newors Thun SPN feds the welomnmisanan sprang to elem, Pele Ta for robles ken [111 ties here noone es ch spans the mun ro hs sae te slo fo SPN Moreover, for ta ens of SPN, ne en ede the sz of the pro by employing the flowing rae (9 Nee ht tach rule cn be performed in puyol tne Let seg) tao he degree te ode 727, THE conte node th eye) = 2 then v andthe ok {ace enue removed fom GIy «At nad uM then sed he malin opin the educed fap. Note thal’ © Monk Gy) blogs othe Stee es 218 contains a node v «Hf with deg) = 2, tenth two Ts Gt) and cone tepnced nk Go a ost yo ey ee ret ro nk erome pal the she wnt lanper cst ean be emoved om ‘pH conan sky) such ty > dee dhe onto the hata bemeen nodes sa then ik |W Figure 5. An example ofa graph for which the Sener ce can be found by ping hc edt Co of al inks =; tothe Sete = 4 4) can be removed from G. Furthermore, ey = dy and ‘there is a path of cost dy from i to, sot contataing {ij then link (can be removed from 6. 4NIEG contains throe distinet nodes u,v, w © M, sock that w andy ae adjacent, ye > dyn A ay > dy the ink (u,v) ‘can be removed from G. Tn other words fu, », and w are any three nodes in the multicast group such that the cost of the link (u, ») is mote than the cos of 2 path from node w to node 1 3s wel as node v, then link (u, 1) does not belong, to the Steiner tee S)Let w= M, Let v and be the closest and second closest adjacent nodes to, respectively. Now, ify + min (dy |P 4¢ M and pu} < om then the link (u,v) belongs to the Steiner tree and G can be contracted along (ut, ¥) In other words, ifthe closest adjacent node (») brings you nearer *0 other members of the multicast group, then link (u, v) ‘should belong tothe Steiner treo. For example, the graph shown in Fig. 5 (nodes in the multi- cast set are shaded) can be reduced to a single node by employing reduction 5 repetitively as follows. First, we con fact along fink (C42, Cl); second, we contract along link (MT, NY); thd, we contract along link (CA1, UT); and finally, wwe contact along link (U7, Mf}. ‘Unfortunately, asthe following lemma demonstrates, these reductions eannot be applied toa large number of instances ‘of SPN which oceur in typical communication networks. Usui aly, hese redvetions anno be apie o casos whieh 1, G isnot sparse and G satisfies the triangle inequality {Go be explained shorty. The folowing Lemma descibes & ‘sufficient condition for an instance of SPN to be “reducible.” {Lemma 1 — If an instance of SPN (say P)satstes all ofthe following three conditions, then P eannot be reduced t0 a smaller instance of SPN by sing the aforementioned eedue- |The graph satisies the triangle inequality, that is, the cost ur Of Link (1) Is strictly less than the cost of any path fom node w to nade v which does not include link (4). 2.The minimum degree of the graph is 3, thatis, Vee Vy ego) 23 3.None of the nodes in the multicast group are adjacent to ‘one another, that, Wu, ¥ @ M, (u,v) Proof — Reductions 1 and 2 cannot be applied because the ‘minimum degree of the graph is 3. Reduction 3 cannot be applied because the graph satisfies the triangle inequality Reduction 4 eannot be applied because none of the nodes in the multicast group are adjacent to ane another. Final, reduc tion 5 cannot be applied because the graph satisfios the tiangle S Thee ree other special ates of SPN for which pobwomia ie ago ex (1) 8 fine raph tobe spare ial te spanning es of rap can fe cmumensted in polnenal ne TREE Network + Sannayabrasry 2000 0 ‘wa Ma ed ed oo Figure 6. An example of «Steiner tee consircted by employing the KMB algorithm. noquality, and none of the nodes in the multicast group are Adjacent fo one another, The graph shown in Fig, 3 satisfies all the above conditions thus, it san “irteduciblo" graph. “Thus, for typical communication networks, it may be impossi- ble to find a Steiner tre in & reasonable amount of times hence, itisimportant to develop appracimation alarms for SPN. ‘Approximation algorithms for SPN run in polynomial ime and produce good- prriymulcest node) > _prony(Steiner node), No ordering i performed among aay two ‘multeast nodes or any two Steiner nodes. Then the fst x nodes inthe priority queue are ‘aken and placed ina sot B. he multe ‘ist nae v that i selected isthe one tha i closest to Thais, famoag all directed paths from a node in B to a multicast node ‘ot in the tree, the path of least cost and the multicast node on hich this pth terminates are chosen, The parameter 82 rob that can be uscd to set the algorithm to a desired trade-off botween expected tree cost and ruaning time that is, as we increase , the tre-cost decreases ut tho running time increases Delay Optimization This section studies th problem of optimizing the delay of a toulicast ree, Note thatthe definition of delay fora source: Specific multicast tree Is diferent from thal for a group- Shared mullicast treé, If ls the soures, the average ource-specifle day (wth respect 0 sores 3}, DS,, ofthe anulticast tee define’ a8 1 Paes ® where P(s, v) is the delay of the path from source s 10 malti- Tae 2 he maa ece Sry the serge ou. a @ where DS, is the average souree-specific delay (with respect 10 souree ¥) of the mlticat tree. ‘The problem of finding a source-specific multicast tree which minimizes DS, has 4 simple solution, described below, On the other hand, the problem of finding a group-shared multicast tree which minimizes the value of DG is NP. plete [18], and will be discussed later inthis subsection, ‘An optimum source-specific delay multicast tree (for bath unidirectional and bidirectional link cascs) is also called the shores path re aa is defined as follows, Lets be the source of ‘a soureespecifie mulicast tee and SPU, ») be the shortest path fromsto modo © A (s}. Consiruet & graph Gsp by taking the union of ll the shortest paths SPs, v), where $18 the source node and vis « multicast node. Now, the shores-puth multicast tree is oblained by removing all the loops in Gy la Fig. 7, tho shortest paths from Source CAL ro destinations 7X, IL, aud NY are CAI-CA2-PN, CAL-WAAL, and CA1-UT-MLNY, respec: tively. Thus, the cost ofthe shortest-path tree i 7, 2nd it consists of links (CA1, CA2), (CA2, TX, (CA1, WA), OFA, IL), (CA, UD), (UT, MD, and (MI, NY), as shown in Fe, 7a ‘As mentioned before, finding « multicast tree which opti izes the average pronp-shared delay (DG) is NP-complete, although polynomial-time approximation algorithms exist ‘hich have a constant performance guaranteo. One such algo "thm finds an optimum center-based tree which is defined 2s follows, Let T, be the shortest-path multicast tree rooted at ade v aud DG! be tho average group-shared delay of Ty. ‘Then an optimum ester based tee is defined asthe shortest path tree with the minimum value of DGF. This tee can easly be found in polynomial time by computing the DG vatuos for all nodes v € G, and taking the minimum. An optimum cen: ter-based tier has a performance guarantee of 2, that is, the average group-shared delay of an optimum center-based tree iguaraniced to be within two cimes an optimum grou shared delay (19), Note thatthe root of the shostest-path wee may not be a multicast member, it may be any node in the fgraph If we choose the center only from the moltigast mem bers the performance guarantee of such a center-based tree is 3 [19]. Figure 7b shows an optimum center-based tree. ‘enter is 7X, and the average group-shared delay is 2.07. The CostDelay Trade-off In the previous two subsections, wo studied algorithms for ‘optimizing the cost and delay of a multicast routing tee. In {is subsection we examine the problem of finding a souree~ specific multicast tree which attempts to optimize both cost fan delay. In general, a single multicast ree cannot have ‘minimum cost and minimum delay. For example, if is the source of a multicast connection and Tis the Steiner tro ound by the KMB algorithm, the average source-specific delay (with respect to sources) of is hounded from above by ([91 + 1/3 times the minimum average source-specitie delay 20], Similarly, the shortest-path tree optimizes the sourte-apecitie delay (with respect 0 soured), fut it can be [af times costice than the Steiner tec, although empirical lata suggests that, on am average, the shortes-path tree may ‘only be slightly (30 percent) costlier than the Steiner tee found by approximation algorithms such as KMB (20). On the other hand, the average souree-specific delay of an Approximation algorithm, such as KMB, i typically larger? 7 Not tha hse nets ast at every ln has the sane cst and lay vals (ey the con of ik hen he delay of he ink ms aso ben, IEEE Network+ anusryFebrvary 2000, 95 (G0 percent than the average source-specifc delay of the shortest-path mutiast tee [21] “Thus, itis natural to ask if an algorithm exists to improve the sources diy carats ol mua ep dhe by on appresiation alpthm ut os MB. One “lgortitn, whichis described low, i base onthe following, intuition, Given an optimum-cost (or near-optimum-cost) 1 treo, if we find the multicast destination for which the olay inthe Steiner tree diflers the mos fom the delay of the ‘corresponding shortest path, and conneot this destination (0 the source by the shortest pith, we can decrease the average sourcespecific delay. Thus if Py(s,») i the delay of the path the optimum

You might also like