Evaluaton of Delay Performane n Valant Load-balanng Network Yngd Yu a, Yaohu Jn a, Hong Cheng a, Yu Gao a, Weqang Sun a, We Guo a, Wesheng Hu a a State Key Laboratory on Fber-Ot Loal Area Networks and Advaned Otal Communaton System, Shangha Jao ong Unversty, Shangha 23, P. R. Chna ABSRAC Network traff grows n an unredtable way, whh fores network oerators to over-rovson ther bakbone network n order to meet the nreasng demands. In the onsderaton of new users, alatons and unexeted falures, the utlzaton s tyally below 3% []. here are two methods amed to solve ths roblem. he frst one s to adjust lnk aaty wth the varaton of traff. However n otal network, rad sgnalng sheme and large buffer s requred. he seond method s to use the statstal multlexng funton of IP routers onneted ont-to-ont by otal lnks to ounterat the effet brought by traff s varaton [2]. But the routng mehansm would be muh more omlex, and ntrodue more overheads nto bakbone network. o exert the otental of network and redue ts overhead, the use of Valant Load-balanng for bakbone network has been roosed n order to enhane the utlzaton of the network and to smlfy the routng roess. Rasng the network utlzaton and mrovng throughut would nevtably nfluene the end-to-end delay. However, the study on delays of Load-balanng s lak. In the work resented n ths aer, we study the delay erformane n Valant Load-balanng network, and solate the queung delay for modelng and detal analyss. We desgn the arhteture of a swth wth the ablty of load-balanng for our smulaton and exerment, and analyze the relatonsh between swth arhteture and delay erformane. Keywords: Load-Balanng network, delay erformane, swth arhteture. INRODUCION Network traff grows n an unredtable way, whh fores network oerators to over-rovson (utlzaton tyally below 3%) ther bakbone network n order to meet the nreasng demands. wo methods are amed to solve ths roblem. he frst one s to adjust lnk aaty wth the varaton of traff. But n otal network, rad sgnalng sheme and large buffer s requred. he seond method s to use the statstal multlexng funton of IP routers onneted ont-to-ont by otal lnks. However, the routng mehansm would be muh more omlex. o exert the otental of network and redue ts overhead, the use of Valant Load-balanng for bakbone network has been roosed n order to enhane the utlzaton of the network and to smlfy the routng roess. 2 3 2 3 4 4 6 5 6 5 a) b) Fg..a) Examle of the Dret Mode. b) Examle of the Valant Load-Balanng Mode. he mehansm of ordnary bakbone network s shown n Fg. a): When data streams enter network from an ngress node, for examle, Node, the node would selet a ath avalable to ther destnaton address to transort data streams aordng from the ngress node to the egress node. We refer ths mehansm as Dret Mode. In Dret Mode, traff may exerene one or more logal hos, and eah logal ho onssts of several hysal hos. he Valant Load- Balanng mehansm s shown n Fg.2 b) wth the dashed lne. When akets enters network from an edge node, gnorng the destnaton address, they would be dstrbuted to all other nodes, where they wll be routed to the destnaton Network Arhtetures, Management, and Alatons V, edted by Janl Wang, Gee-Kung Chang, Yosho Itaya, Herwg Zeh, Pro. of SPIE Vol. 6784, 678428, (27) 277-786X/7/$8 do:.7/2.745336 Pro. of SPIE Vol. 6784 678428-
node. hs mehansm s referred as Load-Balanng Mode. In Load-Balanng Mode, traff traverses two logal hos at most. Due to ts essentalty of wo Ho, rasng the network utlzaton and mrovng throughut would nevtably nfluene the end-to-end delay. However, the study on delays of Load-balanng s lak. In the work resented n ths aer, we study the delay erformane n Valant Load-balanng network, and solate the queung delay for modelng and detal analyss. And we dsuss the nfluenes of arhteture of swth suortng Load- Balanng by our smulaton and exerment. Seton 2 desrbes and models the Valant Load-Balanng mehansm; Seton 4 analyzes the delay erformane n Load-Balanng network; Seton 4 shows the affet brought by swth arhteture on delay erformane; Seton 5 desrbes the smulaton and analyzes the smulaton results. 2. VALIAN LOAD-BALANCING MECHANISM In Load-Balanng model, n nodes are full-mesh onneted. We defne the traff that enters network from node as Loal Ingress raff I, the one leavng network from node as Loal Egress raff E, and other traff assng the node as the Byass raff B. Note that, n I n We use traff ntensty to reresent Loal Ingress raff I = I. I denotes the length of the akets, and denotes the akets ntensty. Noton j denotes the orton of traff I destned to node j. hus, the traff ntensty matrx would be: P 2 I n n 2 B 2 n2 n E n 2n 2 I n We denote the lnk onnetng node to j wth l j, and ts aaty wth j. And the traff t on l would be: hus, the total traff on l j ould be reresented as: B E B E k I n ( n ) n k I k I n ( n ) If we use the rato of t and to reresent utlzaton u of the lnk l k l: U t l n k I l n I k l (5) ( n ) k From the equaton above, we ould note that, for node k, f k s small, t ould afford qute a lot of traff from other nodes; f k s large, t only take /(n-) effet on ts lnk utlzaton, thus always mantan the utlzaton near the average utlzaton. Sne the network traff nreases n an unredtable way, there are always some lnks wth hgh load, whle others wth low load (Current lnk utlzaton s usually less than 3%). Load-Balanng mehansm wll ull down the hgh utlzaton lnks and ush u low utlzaton lnks to avod hot lnks n bakbone networks. From another roset, () (2) (3) (4) Pro. of SPIE Vol. 6784 678428-2
a hgh lnk utlzaton of Load-Balanng network ndates that the total network load s muh heaver, and the only method to solve the throughut roblem ould be enhanng the lnk aaty. Besde the advantage above, the Load-Balanng mehansm resent an outstandng erformane n oeraton omlexty [3]. As every node s full-meshed, all traff ould be forwarded to the destnaton n only one routng, whh s somewhat smlar to the MPLS. But wthout alulatng the status of the network to enable raff Engneerng, Load- Balanng mehansm smly dstrbutes the traff evenly to every other node n order to mrove network utlzaton, whh would greatly redue the omutng burden of whole bakbone network. However, t s worthy to note that all traffs have to traverse the network twe. And as the result of ths wo-ho haraterst, more delay elements wll be ntrodued nto the end-to-end delay, and the erformane of Load-Balanng network ould be questonable. Moreover, the traffs are dstrbuted to dfferent edge nodes, mlyng that the traff traversng dfferent aths may exerene dfferent delay, brngng out the dsorder roblem, and rasng rguous requrements on synhronzaton erformane. 3. ANALYSIS ON LOAD-BALANCING DELAY PERFORMANCE As we mentoned above, under the Load-Balanng mehansm, the mrovement on throughut erformane s aheved on the ost of deteroratng ts delay erformane. But how the delay erformane s affeted, to whh extents they are affeted stll requre more analyss. In ths seton, we solate those omonents of the end-to-end delay, and dslose ther relatonsh wth the Load-Balanng mehansm to answer the questons before. 3. Comonents of delay n Load balanng network Network delay usually onssts of four omonents: roagaton delay, transmsson delay, queung delay, roessng delay [4]. When a aket traverses a sngle ho, t wll exerene eah delay omonent for one tme. D ho D D D D (6) ro D D D ro trans ro trans d vo Lk Lk r Eah delay omonent s ntrodued by dfferent reason. he roagaton delay s determned by the dstane aross hysal ho d and the roagaton seed of the sgnal n medum v. We assume that n the bakbone network, sgnal roagates n the same medum, and that the dstane aross a hysal ho does not dffer from eah other too muh, and dfferenes of eah logal ho s aused by the number of hysal hos t ontans. hus, we ould onsder the roagaton delay of a hysal ho as a onstant. Intutvely, Dret Mode would have a better erformane n roagaton delay beause n the best stuaton, traff traverse only half of the logal ho number than Load-Balanng Mode. However, n Dret Mode, often nvolved wth raff Engneerng, the result of ath seletng ould not always be the shortest-ath, n other word, the ath wth the least ho number. As the Load-Balanng Mode s not senstve to the varaton of traff matrx, dynam bandwdth alloaton an be avoded. Pre-onfguraton of lnk s aaty wll guarantee all logal hos to be the shortest ath, whh mles that Dret Mode may not erform better n roagaton delay at the level of hysal hos whh would reflet the real dstane between the ngress node and egress node. he transmsson delay, aused by the oeraton of sendng the akets nto medum, s determned by the akets length L k and lnks aaty. Although n Dret Mode, dfferent flows may be alloated wth dfferent bandwdth whh s less than the lnks aaty n most ases, a sngle aket does not suffer a worse transmsson delay beause akets are always transmtted as a whole nstead of sharng the bandwdth bt-by-bt. hus, under the same hysal ondton, the same aket exerenes the same transmsson delay, no matter whh mehansm s taken. Smlar wth the ro queue ro (7) Pro. of SPIE Vol. 6784 678428-3
transmsson delay, the roess delay s also ndeendent of the bakbone network mehansm, and t s usually less than 3s [5]. Note that all those delay above s fxed delay, and an be obtaned one the network toology and elements are known. Most mortantly, they are ndeendent of the traff haratersts, whle queung delay, the varable omonent of network delay, s greatly nfluened by the network loads, and s dffult to analyze dretly. hough the relatonsh between node loads and queung delay ould be useful, the omlex routng mehansm n Dret Mode revents us from estmatng the load of nodes. However, as the result of the smle routng mehansm of Load-Balanng Mode, we ould estmate the load of eah node exatly so that the queung delay analyss beomes avalable. 3.2 Queung delay vs. Sngle node load As queung delay s assoated wth a sef lnk, the utlzaton of lnks ould be an arorate ndator of node s lnk. he relatonsh between the utlzaton of lnk and queung delay s shown n Fg.2 [6]. Note that before lnk utlzaton reahes 7%, average queung delay nreases qute slowly, after 7% t nreases sharly. hs henomenon nsres us that the queung delay of hgh utlzed lnk would be greatly redued at the ost of less nrement n the one of low utlzed lnk, as llustrated by the dashed lne n Fg.2. For examle, akets traversng on the lnk wth utlzaton of % suffer.268s average queung delay, whle queung delay of akets on the lnk wth utlzaton of 9% s.683s. After balanng these two lnks load, both utlzatons wll mantan at 5% wth the average queung delay.86s..542s delay nrement of % utlzaton lnks leads to.5s delay reduton of the 9% lnk. he beneft of tradeoff s nearly tmes. Also note that the aket length s a onstant (5Bytes) n Fg.2, ndatng that ths urve should be the low bound the queung delay. Wth the nrement of varane of aket length, urve s nflexon wll move rghtwards, makng the balanng effet more obvous. :.2. :' boi!c) COJJLOIP D!L!PnI focj bc V\JJ!L.8.t- I.I:: 2 3 4 8 8 7 8 8 Lnk Ul Izalnn(%) Fg.2. Queung delay vs. Lnk Utlzaton Fg.3 Load-Balanng Swth Arhteture 3.3 Queung delay vs. Load-Balanng network load From the dsusson above, we ould fnd that queung delay erformane an be mroved by balanng the load. Now, we wll examne other fats of network feld that would also affet queung delay erformane, suh as the number of nodes n and the traff matrx. For smlty, we make some aroxmatons wthout runng the nature of loadbalanng. Load-Balanng mehansm does not erform muh better than Dret Mode wth the node wth all lnks at a hgh utlzaton, beause the average lnk utlzaton an not be redued, whereas Load-Balanng also does not erform better wth low utlzaton node. hus, we onsder the ase n whh a node has some hot lnks and some old lnks, then we ould assume that the average utlzaton of nodes s lose to /(n-), and we ould aroxmate the equaton (5): U After rearrangement and aroxmaton, equaton (6) would be n I k (6) ( n ) ( n ) k Pro. of SPIE Vol. 6784 678428-4
U I ( ) k n ) n n I n 2 n k ( n ) n n I k ( n ) n ( We ould fnd that the traff ntensty s not only affeted by tself, but also by the mean of nut traff ntensty. If the loal nut traff ntensty s omarable to the mean of nut traff ntensty, the requred node number for Load- Balanng mehansm should be at least 4. We should also note that f the traff matrx and lnk aaty s ertan, then wth the nrement of node number n, the U would redue followng the reroal urve, and lead to redung the queung delay aordng to 3.2. Note that n ould not be too muh large, as ) nrement of n does not affet queung delay obvously when n s enough large; 2) the number of lnks would nrease at the order of O(n 2 ); 3) f we use R, the rodut of lnk number and lnk utlzaton to denote the lnk resoures that have been used: R U N l I ( n 2) 2 I ( n )( n 2) ( n ) 2 hen we an fnd that the utlzaton of U dereases at the ost of establshng more lnk resoures. (7) (8) 4. SWICH ARCHIECURE FOR LOAD-BALANCING NEWORK Pratally, besdes the effet brought by load-balanng mehansm, swth arhteture would also nfluene queung delay n Load-balanng network. We desgn the swth arhteture wth the ablty of load-balanng shown as Fg.3. he swth has a loal ort where loal nut traff enters the swth. Wthout routng, aket wll be dstrbuted to outut sub-queues exet the loal ort, aordng to the status of eah outut lnk s outut sub-queue to whh the ort s onneted. Also, the traff enterng by Port to n- from other nodes wll be routed to the orresondng sub-queue. Sub-queues dreted to the same transmt ort are groued, and a sheduler s resonsble for arrangng the akets n sub-queue grou to transmt ort. he mehansm exeuted by sheduler and dstrbutor would also affet the queung delay. 4. Dstrbutng Mehansm Load-Balanng network s amed at mantanng a unform utlzaton on all network resoures, nludng nodes or lnks. It s mortant to study the nfluene of dstrbuton mehansm on utlzaton. he bas and the smlest mehansm s Round-Robn, whh just smly dstrbutes the akets to the transmt orts one by one. hs mehansm suffers several roblems. Frst, t an not adat to dfferent aaty lnks of the same node. Seond, even f all lnks are wth the same aaty, dfferent traff matrx would lead to uneven utlzaton. We roose a dynam dstrbutng mehansm to aheve the atual Load-Balanng based on status of eah ort s outut sub-queue. he deal mehansm would be always exludng the ort wth the largest queue sze out of the dstrbutng ort grou untl other orts take t lae. However, ths mehansm s mratal as t requres omarng the status of all orts for every tme when a aket needs to be dstrbuted, nreasng the oeraton omlexty. Our desgn ould be a more ratal method (shown n Fg.4): eah outut queue grou was montored, when the queue sze exeeds a threshold, montor wll send an nterrut to the dstrbutor, nformng t to omare the status of eah ort. he dstrbutor wll not dstrbute any akets to the ort wth the largest sze. Wth the akets resortng n the queue transmtted, the queue sze wll return below the threshold, whh trggers the dstrbutor to restart sendng akets nto the ort s outut queue. Pro. of SPIE Vol. 6784 678428-5
Montor ransmtter Poly Controllor Montor ransmtter Dstrbutor Montor ransmtter Fg.4. Dynam Dstrbutng Mehansm 4.2 Shedulng Mehansm Shedulng mehansm, another mortant omonent of the swth suortng the Load-Balanng network, would also affet the delay erformane. Note that there are two knds of traff n the swth s outut queue, one referred as dstrbuted traff, whle the other routed traff. A aket traversng a Load-Balanng network wll be treated as dstrbuted traff at the frst ho, and routed traff at the seond ho. As the aket s smly dstrbuted wthout onsderng ther own QoS requrement, we have to dfferentate these two knds of traff when dedng whh one goes frst and how muh bandwdth they ould take. Note that there s a art of traff that arrve the destnaton after only one ho, and we refer ths traff as one-ho traff. It seems more effent to set the dstrbuted traff wll hgher rorty. However, n our smulaton, f we onsder the routed traff wth hgher rorty, we ould fnd that the average end-to-end delay s obvously smaller than the dstrbuted-frst shedulng strategy. hs s beause the routed traff are all those traff have to arrve destnaton as quy as ossble, whle n the dstrbuted traff, the one-ho traff s not anxous to arrve destnaton. Dstrbuted-frst shedulng strategy mles that routed traff, the more urgent one, wll onede to some less urgent traff, and that s why ts delay erformane s worse than the routed-frst shedulng strategy. 5. SIMULAION AND RESUL Among the four omonents of delay erformane, queung delay s the one nfluened most by the Load-Balanng Mehansm. And our smulaton fouses on ths delay omonent to onfrm our evaluaton. 5. Load-Balanng vs. Dret Mode Our ntal smulaton onssts of 4 nodes, and the aket length s a onstant of 5 bytes. In order to faltate the queung delay omarson, we set all lnks wth the same aaty (Mbs) and hysal dstane to solate the fxed delay. Frst, we omare delay erformane of Dret Mode and Load-Balanng Mode (LB). Loal traff of node s.2mbs, and s dvded nto.8mbs,.2mbs,.2mbs for eah destnaton ort. Loal traff of other nodes s.9mbs. All these traff are assumed to be Posson Proess..3.3.3.8.3.3 We fous on.8mb art, whh wll lead to 8% utlzaton of lnk l 2 n Dret Mode. We also smulate the traff matrx wth whh the utlzaton of wll be u to 2% n Dret Mode. he results are shown n Fg.5..2.3.3.2.3.3 Pro. of SPIE Vol. 6784 678428-6
.2.2.2.8.8.8.6.6.6.4.4.4.2.2.2..8..8..8.6.6.6.4.4.4.2.2.2 2 3 4 5 6 7 8 9 he tme aket enterng network (s) 2 3 4 5 6 7 8 9 he tme aket enterng network (s) 2 3 4 5 6 7 8 9 he tme aket enterng network (s) 2 8.2.8 FR H.2.8 FBIH 6.6.6 4.4.4 2.2.2 8..8..8 6.6.6 4.4.4 2.2.2 2 3 4 5 6 7 8 9 he tme aket enterng network (s) q 2 3 4 5 6 7 8 9 he tme aket enterng network (s) 2 3 4 5 6 7 8 9 he tme aket enterng network (s) Fg.5. Delay of erformane: a).8mbs n Dret Mode; b).8mbs n one-ho of LB; ).8Mbs n two-ho of LB; d).2mbs n Dret Mode; e).2mbs n one-ho of LB; f).2mbs n two-ho of LB o observe the dfferene more learly, we dslay the delay s dstrbuton n Fg.6.a) and b). able.. shows ther statst haratersts. able.. Comarson of Delay erformane n dfferent modes ->traff Mode Mn(s) Max(s) Range(s) Average(s).8Mbs Dret.242.254.2262.497 Load-Balanng.484.44.92.582.2Mbs Dret.242 5.577 5.553 8.22 Load-Balanng.484.279.236.688.4.2.3 D!LGC oq rri Jq johj.8 FR DJJJ!C D!RL!G Perentage.2...6.4 Perentage.25.2.5..5.5..5.2.25.3 oq RIJC!J oq Perentage.5 Perentage.2..8.6.4.2.5..5.2.25.3..2.3.4.5 delay (s)..2.3.4.5 delay (s) Fg.6. Delay dstrbuton: a) dret mode; b) Load-Balanng mode; ) Load-Balanng wth Round-Robn dstrbutng; d) Load-Balanng wth Dynam dstrbutng. Whle the mnmum of Dret Mode s end-to-end delay s less than Load-Balanng mode, Load-Balanng exhbts a shorter range, mlyng a shorter queung delay. Note that the mnmum of the end-to-end delay n both modes s deded by the fxed delay omonent, eseally the number of hos. For the smlty of analyzng, our assumton that the Pro. of SPIE Vol. 6784 678428-7
numbers of hysal hos s equal wth the logal hos leads that a aket gong through a Load-Balanng network traverse double number of hos as the akets gong through the Dret Mode. However, n the real bakbone network, ths s not always true, and wth the nreasng hysal hos, the queung delay wll lay a more and more mortant role than the varaton brought by the fxed delay dervng from the number of hos. We should also note that although for some node, although ts loal ngress traff vares hugely, the end-to-end delay the traff exerene vares lttle as the result of the varaton of traff beng balaned to every other node. 5.2 Dstrbutng and Shedulng Mehansm We smulate dfferent dstrbutng mehansm. he referene traff matrx s 2.8.8.8.3.2.2 In ths traff matrx, utlzaton of lnks on the node 2 wll dffer huge after load-balanng at the seond ho. For examle, lnk l 2 wll transort.93mbs traff whle lnk l 23 only needs transort.47mbs traff, and the exeted traff on eah lnk should be.62 under an arorate dstrbutng mehansm. he delay dstrbuton s shown n Fg.6. ) and d). And able.2 shows the statsts of eah mehansm. able.2. Comarson of Delay erformane n dstrbutng mehansms Mehansm Mn(s) Max(s) Range(s) Average(s) Round-Robn.484.479.436.8 Dynam Dstrbutng.484.76.222.66 o omare shedulng mehansm wth dfferent rorty settngs, we adot followng matrx for smulaton: 3.4.4.4.8.4.4 We omute the tme average of end-to-end delay n the smulaton wth dfferent rorty settngs but same traff matrx. And the result s shown n Fg.7, n whh routed-frst strategy exhbt an obvously better delay erformane omared wth the dstrbuted-frst strategy..3.2.2.2.4.4.3.2.2.2.4.4 Il II 'I ' Dtthd R rnd.6 2 3 4 CC 6 CC ICC Fg.7. Delay erformane wth dfferent shedulng rortes. Pro. of SPIE Vol. 6784 678428-8
6. CONCLUSION From our analyss of the delay n load-balanng network, we an onlude that omared to the dret mode, loadbalanng redues the hgh utlzaton lnks queung delay, the varyng art of total delay, at the ost of the nrement on fxed delay and a tny nrement on other lght load lnks, so as to mrove the jtter erformane when network s not unformly utlzed. Whle balanng the queung delay, Load-Balanng mehansm also balanes the varaton of queung delay. Dstrbuted more evenly, all akets under load-balanng mehansm exerene smlar queung delay, whh mles the burst traff nfluene less on delay erformane under load-balanng mehansm. From our analyss, t ould be onluded that delay erformane wll be mroved wth the nreasng network sze. However, blndly nreasng number of nodes wll waste network resoures. We do not evaluate the delay erformane of load-balanng network from the rosetve of network marosoally, but also from a mroosm vew of swth arhteture. Both dstrbutng mehansm and shedulng mehansm ontrbute to the queung delay. We show that Round-Robn dstrbutng mehansm an not aheve the real loadbalanng, thus deterorate delay erformane, whle a dynam mehansm wll greatly mrove queung delay. We also exlore the mehansm n whh dfferent rorty settngs nfluene delay erformane, showng that routed-frst strategy would erform better than dstrbuted-frst strategy. hs work s suorted by NSFC and 863 Program. 7. ACKNOWLEDGEMEN REFERENCES S. Iyer et al., An aroah to allevate lnk overload as observed on an IP bakbone n Pro. INFOCOM 22,. 46-4 2 H. Nagesh et al., Load-Balaned Arhteture for Dynam raff n Pro OFC 25,.3 3 H. Lu et al., On Dret Routng n the Valant Load-Balanng Arhteture n Pro GLOBECOM 25,.6 4 B.Y. Cho etal., Analyss of ont-to-ont aket delay n an oeratonal network n Pro INFOCOM 24,. 797-87 5 K. Paagannak et al. Measurement and Analyss of Sngle-Ho Delay on an IP Bakbone Network. IEEE JSAC, 2,.98-92 6 W. Stallngs. Hgh-Seed Networks and Internets.48-56. Pro. of SPIE Vol. 6784 678428-9