Fair and Scalable Load Distribution in the Internet

Transcription

1 Proc. 3rd Internatonal Conerence on Internet Computng, Las Vegas, NV (June 2002), pp Far and Scalable Load Dstrbuton n the Internet Vasl Hnatyshn and Adarshpal S. Seth Department o Computer and Inormaton Scences, Unversty o Delaware, Newar, DE ABSTRACT The current best eort approach to qualty o servce (QoS) n the Internet can no longer satsy a dverse varety o customer servce requrements, and that s why there s a need or alternatve strateges. We beleve that the Internet needs means or provdng a ne-graned per-low QoS that does not cause networ congeston and eeps overall ln utlzaton hgh. In ths paper we ntroduce an ecent, ast and scalable load dstrbuton mechansm, whch arly dstrbutes avalable resources among the lows based on ther resource requrements. The load dstrbuton scheme (LDS) s mplemented va a message exchange protocol whch mantans hgh ln utlzaton whle ncurrng low overhead. We study the LDS and compare two arness mechansms ntroduced wthn the LDS ramewor usng smulatons n OPNET. Keywords: Qualty o servce, dynamc admsson control, load dstrbuton, networ eedbac, ar rate dstrbuton 1. Introducton. The one servce or all model used by the current Internet can no longer satsy the multtude o customer requrements. We beleve that the Internet needs s a scalable mechansm or provdng a ne-graned per-low qualty o servce that does not cause congeston n the networ and eeps overall utlzaton o the lns hgh. In author s opnon, one o the best ways to acheve ths goal s to ntroduce a load dstrbuton scheme (LDS) at the networ boundares. Such a scheme would requre the ngress nodes to lmt sendng rates o the lows based on the congeston level n the networ. In ths paper we ntroduce a ast, ecent, and scalable load dstrbuton mechansm that arly dstrbutes avalable resources among the lows based on ther resource requrements. The LDS s based on the dea that durng congeston, an overloaded nterace notes the ngress nodes that contrbute to the congeston asng them to reduce transmsson rates o the lows. Upon recevng congeston notcaton message, the ngress router computes ts new share o the bandwdth on the overloaded nterace and dstrbutes t among ndvdual lows. The proposed LDS guarantees that each low receves at least ts mnmum requested amount o bandwdth, whle any addtonal resources are shared among all lows n a ar manner. The LDS admts a new low nto the networ only there s enough resources to satsy the low s request. The LDS s mplemented usng a message exchange protocol that retans hgh ln utlzaton and ncurs low overhead. The LDS s most benecal to applcatons that can tolerate varaton o ther transmsson rate whle stll provdng the end-user wth an adequate level o servce qualty. FTP s an example o such an applcaton; t tolerates varaton n the loadng speed but can use as much bandwdth as the networ can provde. Multmeda applcatons can also benet rom the LDS. Such applcatons cannot operate below certan transmsson rates because the qualty o the pcture or sound becomes unrecognzable; however, they can tolerate varaton o the allocated bandwdth as long as the qualty o arrvng data s acceptable to the recever. The rest o ths paper s organzed as ollows. Secton 2 ntroduces dentons o arness, whle secton 3 descrbes the message exchange protocol used n the LDS. Secton 4 presents smulaton results and n Secton 5 we dscuss scalablty ssues and examne the problem o provdng per-low QoS wthn the LDS ramewor. Secton 6 provdes an overvew o the related. Fnally, we conclude n Secton 7 wth possble uture drectons. 2. Farness Dentons In order to determne a permssble sendng rate o a low, each boundary node mantans a Requested Load Range, RLR = (b, B ), or the lows that enter the networ doman through t. A low s RLR conssts o two values: a mnmum rate, b, below whch the low cannot operate normally and the maxmum rate, B, that the low can utlze. The low s sendng rate, R, s lmted by ts RLR and les wthn ths requested range. Throughout the paper we wll oten reer to numerous dentons o the RLR aggregates. To avod potental conuson we dene these aggregates as ollows. In addton to the low RLRs, each ngress node eeps trac o the path RLRs. The path RLR, (b P, B P ), or the RLR o the ngress node on the path P s a load range where b P corresponds to the sum o the mnmum requested rates o the lows that orgnate rom the ngress node and traverse the path P, whle B P s the sum o the correspondng requested maxmum rates. Usng P to denote that low traverses path P, we dene the path RLR as ollows:

2 P P b = b = P B B (1) P Smlarly, we dene nterace and aggregated nterace RLRs or the core router nteraces. The nterace RLR o nterace or the ngress node, (b, B ), s the sum o the path RLRs o the ngress node, subject to the condton that the paths nclude nterace. P P b = b B = B (2) P P To avod conuson, we wll use an upper-case letter (e.g. P) or a path and a lower-case letter (e.g. ) or an nterace. Fnally, the aggregated nterace RLR o nterace, (b, B ), s the sum o ts nterace RLRs. b = b B = B (3) Ingress nodes obtan the low RLRs rom the servce level agreements establshed wth the user, and they compute the path RLRs based on these values. Each core nterace obtans nterace RLRs rom the ngress node s advertsements and computes an aggregated nterace RLR. Ingress nodes mantan normaton about ndvdual lows (e.g. low s RLR) and ther correspondng paths (e.g. path RLRs), whle the core routers mantan only per-ngress node normaton (e.g. nterace RLRs). A more detaled overvew o the data structures mantaned n the ngress and core nodes s provded n [4]. Congeston notcaton messages carry nterace and aggregated nterace RLRs. These values allow ngress nodes to arly dstrbute avalable resource among ndvdual lows. In ths paper we examne two dentons or computng an ngress node s ar share. The rst denton, whch we call proportonal, computes the ar share, FS, o the ngress node proportonally to the mnmum value o ts nterace RLR on the congested nterace : b b ( ) FS = = + mn b C b, B mn C, B (4) b b where C s the capacty o the outgong ln on nterace. Accordng to equaton (4), each ngress node receves ts share o the nterace s capacty proportonally to ts mnmum requested rate on the congested nterace. The ar share o each low that contrbutes to the congeston s computed n a smlar way. Furthermore, the ar share o the low, FS, s also proportonal to ts mnmum requested rate on the congested nterace: b FS = mn C, B (5) b The second denton o arness computes the ar share o an ngress node proportonally to the derence between the maxmum and the mnmum requested rates, whch s the amount o bandwdth the low needs to maxmze ts perormance. We call ths arness denton a maxmzng utlty arness. Thus, the ar share on congested nterace o ngress node and o low s computed as ollows: B b ( ) FS = mn b + C b, B (6) B b ( ) B b FS = mn b + FS b, B B b ( ) B b = mn b + C b, B (7) B b 3. The message exchange protocol The message exchange protocol conssts o three dstnct phases. Durng the rst phase, called path probng, the ngress node attempts to learn about the current state o a path or to learn the path tsel the route to the low s destnaton s unnown. The probe messages collect the current arrval rate o the trac and the aggregated nterace RLR or each traversed ln. The probe messages are generated ether perodcally or when a new low s actvated. Perodc probng s used to determne the ngress node can ncrease ts sendng rate on the path, whle the probng caused by the low actvaton determnes the new low can be admtted nto the networ. Admsson o a newly actvated low nto the networ or a low termnaton ntates the second phase called the RLR change phase. The purpose o ths phase s to update the nterace RLRs along the low s path. I the admsson o the new low causes congeston anywhere along the path, then the ngress node ntates the thrd phase, called the Rate Reducton Phase. Durng the thrd phase congested nteraces noty ngress nodes to slow. The message exchange protocol uses the ollowng message types. In the rst phase, ngress nodes generate PROBE pacets and receve results o the path probng va PROBE_REPLY messages. In the second phase, ngress nodes advertse changes usng RLR_CNG pacets. CN and CN_CORE messages are used durng the rate reducton phase to convey normaton about congested nteraces to the ngress and core nodes respectvely. 3.1 The Path Probng Phase The purpose o the path probng s to learn the current arrval rate and the aggregated nterace RLR at each ln on the path o nterest. The PROBE messages are generated by an ngress node ether perodcally or due to the actvaton o a new low, and are processed by all core routers on the path to the speced egress node. When a core router receves a PROBE message, t retreves normaton about the nterace on whch the probe message wll depart. Ths normaton conssts o the IP address o the nterace, estmated total arrval rate, capacty, and an aggregated nterace RLR. The nterace normaton s

3 stored n the body o the PROBE, and the message tsel s orwarded urther along the path. The PROBE_REPLY message that carres collected path normaton rom the egress node bac to the ngress node does not requre any addtonal processng by the core routers. Inormaton collected by the PROBE s used to update the path state normaton at the ngress node. Each ngress node mantans three tables: a low table, a path table, and an nterace table. The low table contans normaton such as the low RLR, the current sendng rate, and the destnaton address o each low. The path table contans normaton about each path that s beng traversed by the trac orgnatng rom ths ngress node. The path table entry contans the ordered lst o the nteraces that belong to the path and the lst o the lows that travel on ths path. Inormaton about each nterace o the path s mantaned n the nterace table. An nterace table entry contans the IP address o the nterace, the total arrval rate, capacty, and an aggregated nterace RLR. Once the nternal tables are updated, the ngress router examnes the cause o the PROBE generaton. I t was a perodc probe, then the ngress node examnes the path normaton n order to determne there s excess bandwdth avalable on the path. I the probe was generated due to low actvaton, then the ngress node calculates the new low s permssble rate and ntates the RLR change phase. Detals regardng computaton o the avalable resources on the path and new low s rate can be ound n [5]. They were omtted due to space lmtatons. 3.2 The RLR change phase The ngress node ntates the RLR change phase upon low actvaton or termnaton. The RLR_CNG message generated due to low actvaton s always preceded by the path probng phase, durng whch the ngress node computes the sendng rate o the new low and determnes such a rate ncrease wll cause congeston anywhere on the path. An nterace consdered to be congested the new low s rate causes the total arrval rate on the nterace to be larger then ts capacty: A + R > C. I addton o a new low causes one or more nteraces on the path to become congested, the ngress node dentes an nterace that wll ntate the rate reducton process. Such nterace should satsy two condtons: t should be congested and t should be located the smallest number o hops away rom the destnaton along the low s path. The second condton allows aggregaton o congeston notcatons because the rate reducton process wors opposte to the low o trac on the path. Subsequently, the ngress node generates and orwards the RLR_CNG message whch contans the RLR and the sendng rate o the new low and the dentty o the nterace that wll ntate the rate reducton process. The nterace dentty eld s empty there s no congeston on the path or the RLR_CNG s generated upon low termnaton. When a core router on the low s path receves ths message, t updates ts nterace RLR and the estmated arrval rate on the outgong nterace. The core router ntates the rate reducton process the dentty o ts outgong nterace s ncluded n the RLR_CNG message. The RLR change message generated due to low deactvaton causes a decrease o the nterace RLR on each nterace o the path. Ths n turn causes an ncrease n the ar share or each ngress node that sends trac through the nterace. However, snce none o the ngress nodes nows about the RLR decrease, they contnue sendng trac at the rate below ther new ar share, whch may result n temporary underutlzaton. The LDS reles on the perodc probng nstead o explct advertsement o avalable bandwdth to deal wth such stuaton. A low may only ncrease ts sendng rate each nterace on the path has excess bandwdth avalable. Ths normaton can be obtaned only by probng the path. 3.3 The Rate Reducton Phase The Rate Reducton Phase begns when the core router nterace chosen to ntate the rate reducton process receves the RLR_CNG message. We wll use to denote ths nterace. The ngress node that requested the RLR change wll be called the ntator. The Rate Reducton Phase conssts o two steps: dentyng the ngress nodes to throttle, and generatng congeston notcatons Identyng ngress nodes to throttle The core router dentes the set U that conssts o ngress nodes that send trac through nterace. The set ndrect U s then dvded nto two subsets U and U drect, called ndrect and drect notcaton sets, respectvely. Ingress nodes whose trac arrves on the same ncomng ln to nterace as trac rom the ntator belong to U ndrect, whle the remanng ngress nodes belong to the set U drect = U - U ndrect. Ingress A C1 C2 C3 Egress E Intator Congested nterace Ingress B Ingress C Fgure 1. Selecton o the drect and ndrect notcaton sets. Fgure 1 shows an example o how these sets are beng selected. The ntator A causes congeston on the nterace o the core router C2. The set U conssts o A, B, and C, because ther trac traverses the nterace. The ndrect notcaton set, U ndrect, contans ngress nodes A and B and the drect notcaton set, U drect, contans only ngress node C. Snce ntator may cause congeston on multple lns, the ndrect notcaton set contans ngress nodes that

4 contrbute to congeston not only on the current nterace but possbly on other upstream nteraces. Thus, to avod dealng wth multple congeston notcatons and to reduce the overall number o control messages, the core routers aggregate normaton about the ngress nodes that belong to the ndrect notcaton set. Ths normaton s carred n a CN_CORE message to the upstream nodes. By contrast, trac rom ngress nodes that belong to the drect notcaton set does not nluence the congeston stuaton upstream but t does contrbute to the congeston n the current nterace. Thus, these ngress nodes are drectly noted through the CN message Generatng congeston notcatons The CN message carres normaton about the congested nteraces that requested a rate reducton to a sngle ngress node n the drect notcaton set. The CN_CORE message carres such normaton to multple (all) ngress nodes n the ndrect notcaton set va other upstream nodes. Both CN and CN_CORE messages carry the ollowng normaton about the congested nterace(s): IP address, nterace and aggregated nterace RLRs, and the capacty. Ingress nodes that receve CN dstrbute the load accordng to the algorthm presented n Appendx A. Upstream nodes that receve CN_CORE may need to update the payload o the orwarded CN_CORE and newly created CN messages. Upstream nodes that receve the CN_CORE message update notcaton pacets only the outgong nterace j correspondng to the ln on whch the CN_CORE message arrved s congested and satses one o the ollowng rules. The core routers determne congeston status o the nterace j by comparng the ln capacty wth the derence between the total arrval rate on the nterace j and the rate reducton requested stream. For more detals see Appendx B. Rule 1: I the rate reducton on the congested nterace j s larger than the rate reducton on the stream nteraces, then the normaton about these stream nteraces s replaced wth the data o the nterace j. Rule 2: Otherwse, the nterace RLR on the congested nterace j s larger than that on the next-hop stream nterace, then the normaton about the nterace j s added to the correspondng ngress node entry n the control messages. In rule 1, we dene the rate reducton on nterace j to be larger than that on nterace j or the same RLR values, the nterace j causes larger rate decrease then the stream nterace j. These rules apply to each ngress node entry separately because ndvdual ngress nodes may contrbute to congeston on derent sets o nteraces. 4. Smulaton Results To test and evaluate the perormance o the load dstrbuton scheme, we perormed a smulaton study usng the OPNET networ smulator [11]. The goal o our smulaton was to show that the load dstrbuton scheme prevents congeston n the networ, mantans hgh ln utlzaton, arly dstrbutes avalable bandwdth among ndvdual lows, and ncurs a very small load overhead. We also compare the proposed dentons o arness and show that the maxmzng utlty arness acheves hgher throughput n the networ. [400, 1200] T=185 s DEST 4 Ingress 12 [200, 1000] T=110 s DEST 3 SRC 1 SRC 2 [800, 2000] T = [170 s, 260 s] DEST 1 DEST 1 DEST 2 Egress 1 Egress 2 Core 2 Core 5 Core 3 Ingress 3 Ingress 4 SRC 3 SRC 4 [500, 1300] T = [60 s, 360 s] DEST 2 Fgure 2. Smulaton Topology. DEST 3 DEST 4 In our smulaton we consder our sources sendng multmeda trac through the networ o Fgure 2, whch shows the topology used n the smulaton as well as the destnaton, RLR, and the tmerame o operaton or each source. For example, SRC 1 starts sendng ts trac to DEST 4 at tme 185 seconds and nshes at the end o the smulaton. Its RLR s [400 Kbps, 1200 Kbps]. Each ln n the networ s provsoned wth 1544 Kbps. We mplemented and tested the b-drectonal verson o the LDS but n the paper we consder only a undrectonal case n order reduce the complexty o dscusson. 4.1 Evaluaton o the congeston preventon I the probe message ndcates that addtonal trac wll not cause congeston on the path, then the new low can start transmttng beore the RLR change message s sent. However, addtonal trac wll cause congeston, then the ngress node generates an RLR change message and wll allow the new low to transmt ts trac only when the CN message arrves. Thus, upon a new low s actvaton, congeston can stll occur only the ngress node wll ncrease ts sendng rate beore other ngress nodes that contrbute to congeston wll slow. However, congeston wll only last untl all partcpatng ngress nodes receve the congeston notcatons and adjust ther sendng rates. Ingress node Flow actvaton Probe Probe Reply RLR change Congeston Notcaton Ingress 12 Source sec sec sec Ingress 3 Source sec sec sec Table 1.Control message exchange tmng Egress 3

5 Table 1 shows the arrval and departure tmes o the control pacets at the ngress nodes upon the low s actvaton. These results were collected usng the smulaton scenaro shown n Fgure 2. One could observe that Ingress 3 needed sgncantly less tme to complete the message exchange as compared to Ingress 12. Ths phenomenon s explaned by the act that Ingress 12 probes a longer path than Ingress 3. Smlarly, both ngress nodes wated longer or the probe reply than or the arrval o the congeston notcaton ater the RLR change. Ths happens because the probe has to traverse a complete round trp path, whle an ntermedate core router, and not an egress node, may generate the congeston notcaton. Thus, n general, a complete control message exchange ntated by a low actvaton wll last not longer than two round trp tmes (RTT). Ingress Node Congeston Notcaton (CN) arrval tme Ingress seconds Ingress seconds Ingress seconds Table 2. Arrval tmes o CN messages due to the low actvaton To better understand how long the congeston may last, let us examne what happens when SRC 1 starts sendng trac at T=185 seconds. Table 2 provdes a lst o the congeston notcaton arrval tmes caused by the actvaton o SRC 1. Ingress 12, whch requested the RLR change, receves a congeston notcaton and adjusts ts sendng rate at tme seconds, whle ngress nodes 3 and 4 receve ther congeston notcatons at tmes and seconds, respectvely. Thus, n ths stuaton, congeston was avoded, because ngress nodes 3 and 4 were able to reduce ther sendng rates beore Ingress 12 njected addtonal trac. I Ingress nodes 3 and 4 would receve the congeston notcaton ater Ingress 12, then congeston would not be avoded. However, even n the worst case, congeston would last only rom the tme o the sendng rate ncrease untl the tme the last ngress node that contrbutes to congeston receves ts notcaton. Ths amount should be lmted by the longest RTT. 4.2 Farness schemes and ln utlzaton Fgures 3 5 show how the ngress nodes adjust ther sendng rates or the scenaro descrbed n Fgure 2. These results were collected usng the maxmzng utlty denton o arness. Note that throughout the smulaton the ngress nodes share avalable resources arly. For example, durng the tme perod [185 sec, 260 sec] the ln core 2 core 5 s a bottlenec or ngress nodes 12 and 3, whle ln core 5 core 3 s the bottlenec or the ngress nodes 12 and 4. As a result, avalable resources are dstrbuted as ollows: sendng rate or the Ingress 12 s 682 Kbps, 862 Kbps or Ingress 3, and 648 Kbps or Ingress 4; ther ar shares at the respectve bottlenecs. However, snce Ingress 12 cannot ully utlze ts ar share at the ln core 5 core 3 due to the bottlenec at core 2 core 5, Ingress 4 s able to use the excess bandwdth and gradually ncrease ts sendng rate. Fgure 3. Load dstrbuton by the Ingress Router 12 Fgure 4. Load dstrbuton by the Ingress Router 3 Fgure 5. Load dstrbuton by the Ingress Router 4 Fgure 6 shows utlzaton o the lns core 2 core 5 and core 5 core 3, whch ndcates that or the duraton o the experment at least one o the lns was 100% utlzed. Furthermore, even when the Ingress 12 was not able to send trac at ts ar share on the ln core 5 core 3, the ln utlzaton oscllated around 95%.

6 To compare perormance o the proportonal and maxmzng utlty dentons o arness we slghtly moded our scenaro. In the new scenaro, we change RLR o SRC 4 to [800 Kbps, 900 Kbps]. Fgures 7 and 8 show ln utlzaton or the derent dentons o the arness. As Fgure 8 shows, proportonal arness does not utlze ln resources completely durng the tme perod [110 sec, 170 sec], whle the maxmzng utlty arness does, as shown n Fgure 7. B b B b ( C b ) > B ( C b ) > B b C B + (9) b > B b B b Inequalty (9) shows that usng maxmzng utlty denton o arness the ln s underutlzed only the maxmum requested rate on the nterace s lower than the ln s capacty. Otherwse the ln bandwdth s completely utlzed. Our smulaton results support these observatons. Fgure 6. Ln Utlzaton usng Maxmzng Utlty Farness Fgure 7. Ln Utlzaton usng Maxmzng Utlty arness The proportonal arness suers rom ths decency because the ar share o the ngress node s larger than ts maxmum requested rate, then the ngress node sends trac at ts maxmum requested rate and letover bandwdth s not dstrbuted among the rest o the ngress nodes. I FS > B then the ln wll be underutlzed because Σ (FS ) = C. Let us examne when proportonal arness causes ln underutlzaton. b C B > (8) C B > b b b Thus, usng proportonal denton o arness ln wll be underutlzed whenever nequalty (8) holds. However, maxmzng utlty arness always utlzes the ln the maxmum value o the nterace RLR s larger than the nterace s capacty. Fgure 8. Ln Utlzaton usng Proportonal arness 4.3 Examnng the overhead We dene the control load overhead caused by the LDS as the rato between the total number o the data and control pacets generated. Fgure 9 shows how the requency o the perodc path probng nluences the load overhead. As expected, control pacet overhead ncreases as the probe generaton rate grows. Stll, even when we rase the probe generaton rate to 1 pacet every 4 seconds, the overall overhead dd not reach 0.8%. However, the probe generaton rate s not the only actor that aects the overhead. In a networ wth a lot o short-lved lows, RLR change and congeston notcaton messages could sgncantly ncrease the overall load. We address ths problem n the next secton. Fgure 9. Control pacet overhead vs. Probe Generaton Rate 5. Scalablty and per-low QoS In the prevous secton, we have shown that the LDS ncurs nsgncant load overhead. However, n our smu-

7 laton we used a small networ wth arly large lows. Large networs usually have a lot o short-lved, small lows, whch actvate and termnate very requently. As a result, LDS may ncur a notceable amount o overhead due to control message exchange. Furthermore, requent advertsements o RLR change and subsequent rate adjustments may cause extreme varaton o the congeston level n the networ, whch would ncrease the number o control messages beng exchanged. Thus, ntatng the RLR change phase each tme a small low actvates or termnates could be detrmental to the scalablty o LDS. To elmnate ths problem, we propose that the ngress nodes request large chuns o the resources to accommodate requent actvaton and de-actvaton o the small lows. For example, or large long-term lows, the ngress nodes would request RLR change each tme they are actvated or deactvated, but or small short-term lows, ngress nodes would generate RLR change requests only when the aggregated RLR o the small lows goes beyond a certan threshold. Ths approach would reduce the overall number o control messages and thus would mprove scalablty o the load dstrbuton scheme. Requested mnrlr Bandwdth Actve mnrlr Requested RLR ar share Actve RLR ar share Compared wth threshold RLR allocated on the path Requested maxrlr RLR requested by the currently actve lows Fgure 11. Requested and actve RLRs Actve maxrlr To mplement ths approach, we consder when the ngress node should advertse the RLR change o the low aggregate. Let the RLR that the ngress node requested on the path be the Requested RLR, and the RLR that the lows ased or rom the ngress node be the Actve RLR. The ngress node would generate an RLR change message only the derence between the Requested RLR ar share and the Actve RLR ar share goes beyond some threshold value. Fgure 11 shows a relatonshp between the Actve and Requested RLR values. Introduced LDS also provdes the possblty o mplementng a varety o servces on a per-low bass. In order to preserve such LDS propertes as arness among ngress nodes, hgh ln utlzaton, and congeston preventon, the ngress nodes should use the same denton o arness throughout the networ when computng ther bandwdth share. However, the ngress nodes may use derent and more complex polces when dstrbutng resources among the lows. Each ngress node may mplement ts own set o load dstrbuton polces wthout ntererng wth the polces o other ngress nodes, whch maes LDS extremely lexble n terms o provdng perlow QoS. 6. Related wor overvew Ths paper s a drect extenson and mprovement o the wor done n [4], whch addresses the same problem through the approxmaton o the ngress node ar share based on networ eedbac. The approach proposed n [4] s less accurate n computng the ar share o ngress nodes, taes longer tme to converge, and does not guarantee ar load dstrbuton among the edge nodes under all networ condtons. In [6], Kar et al provded an excellent denton o the dynamc rate control problem and ntroduced an teratve algorthm that solves t. In [6], ndvdual sources adjust ther sendng rates based on the utlty uncton and the networ eedbac whch conssts o normaton about the number o congested lns on the path. However, the algorthm proposed n [6] converges to the optmal values slowly, operates on a per-low bass, requres sources to communcate ther sendng rates to the core routers, and reles on the ACK pacets to carry the eedbac. In certan stuatons, the soluton proposed n [6] becomes unacceptable because o these eatures. Mrhaa et al ntroduced a somewhat related dea n [10]. Ther goal was to mody the resource reservaton protocol RSVP or supportng dynamcally changng QoS requrements n moble ad hoc networs. The proposed drsvp mechansm also assumes that each low requests resources n a range. When a new low enters the networ and there are not enough resources to accommodate t, the congested ln wll adjust the reservatons o other lows n order to accept the new low s reservaton. Unortunately drsvp also wor on a per-low bass and thus does not scale well. Furthermore, t does not guarantee that the lns n the networ wll be ully utlzed and t allows perods o QoS degradaton. The Explct Congeston Notcaton (ECN) model [12] requres that the sources wll reduce ther rates upon recepton o the CE mared pacets. Both Explct Congeston Notcaton approach and smple rate control algorthm [6] assume that the sources are well behaved and would reduce ther sendng rate upon congeston notcaton arrval. Unortunately n the dverse Internet envronment, we cannot be sure that all the sources wll behave as requested. Thus nether o these approaches provdes protecton aganst denal-o-servce attacs. On the contrary, the load dstrbuton scheme that we have ntroduced deals wth trustworthy boundary nodes that would adjust sendng rates regardless o the user behavor and thus mtgates the possblty o msbehavng sources launchng a denal-o-servce attac. The problem o admsson control [3, 6-7] and controlled-load servces [13] s somewhat related to the ssues dscussed n ths paper. However they address a slghtly derent problem o determnng when a new low could

8 be accepted nto the networ, whle the LDS examnes the problem o how to arly dstrbute resources among the sources n order to accommodate the new low s request. 7. Conclusons and uture wor In ths paper we ntroduced a new load dstrbuton scheme that allows ast and ar rate adjustment at the edge nodes. Our scheme requres a core node to mantan normaton about the edge nodes that send trac through ts nteraces and t uses a message exchange protocol to dstrbute ths normaton. However, the core nodes do not eep any per-low normaton and the message exchange protocol does not cause too much overhead. In the worst case, the proposed load dstrbuton scheme requres two RTTs to adjust the sendng rate among the ngress nodes. However, t requres at most the length o the longest RTT to elmnate congeston n the networ. Currently, we are urther nvestgatng the characterstcs o the ntroduced load dstrbuton scheme. In partcular, snce each ngress node probes the path and requests RLR change ndependently o other nodes, we are examnng the possblty o race condtons. In the current verson o the LDS mplementaton, we employ tmers to determne the reshness o the collected normaton; however we beleve that ths may not be enough to solve the problem o race condtons. In addton, we are examnng possbltes o mprovng the scalablty o the LDS, buldng servce polces or provdng per-low QoS, and expandng the LDS to a moble envronment. 8. Reerences [1] H. Chow, A. Leon-Garca, "A Feedbac Control Extenson to Derentated Servces", March I-D: drat-chow-dserv-bctrl.txt. [2] D. Clar, W. Fang, "Explct Allocaton o Best Eort Pacet Delvery Servce," IEEE/ACM Transactons on Networng, vol. 6, no. 4, pp , August [3] R.J. Gbbens, F.P. Kelly, "Dstrbuted connecton acceptance control or a connectonless networ", ITC 16. Elsever, [4] V. Hnatyshn and A.S. Seth, Avodng Congeston Through Dynamc Load Control,'' SPIE' s Int l Symposum on The Convergence o Inorm a- ton Technologes and Communcatons, Aug. 2001, pp [5] V. Hnatyshn and A.S. Seth, Provdng per-low QoS usng load dstrbuton scheme, TR , Department o Computer and Inormaton Scence, Unversty o Delaware, February [6] K. Kar, S. Sarar, L. Tassulas, "A Smple Rate Control Algorthm or Maxmzng Total User Utlty," Proc. o INFOCOM 2001, pp [7] F.P. Kelly, P.B. Key, S. Zachary, "Dstrbuted Admsson Control", IEEE Journal on Selected Areas n Comm., 18 (2000) [8] T. Kelly, An ECN Probe-Based Connecton Acceptance Control, Computer Communcaton Revew 31(3), July [9] A.F. Lobo and A.S. Seth, '' A ooperatve c congeston management scheme or swtched hgh-speed networs,' ' Proc. ICNP -96, Internatonal Conerence on Networ Protocols, Oct.-Nov. 1996, pp [10] M. Mrhaa, N. Schult, D. Thomson, "Dynamc Qualty-o-Servce or Moble Ad Hoc Networs," MobHOC Frst Annual Worshop on Moble and Ad Hoc Networng and Computng, 2000, pp [11] OPNET Modeler. OPNET Technologes Inc. [12] K. K. Ramarshnan, Sally Floyd, D. Blac, "The Addton o Explct Congeston Notcaton (ECN) to IP", March Internet Drat: dratet-tsvwg-ecn-03.txt. [13] J. Wroclaws, Speccaton o the Controlled-Load Networ Element Servce, September IETF RFC Appendces Appendx A. The rate reducton at the edges The congeston notcaton message that arrves at the ngress node contans the lst o the nteraces that experence congeston. Based on the IP address o each congested nterace n the lst, the ngress node dentes correspondng entres n the router table and updates them wth the new normaton. Ater that, the ngress node dentes the set o lows that travels through each congested nterace. We dene the set o lows that vst nterace as vst Φ = { }. Subsequently, the ngress node computes a new sendng rate or each low that belongs to the set Φ startng rom the last congested nterace n the lst, whch s arranged n the order rom the closest to the most dstant nterace rom the ngress node. Snce each low may vst multple nteraces n the lst, we adjust the sendng rate o the low only once accordng to the normaton o the nterace that t vsts last, the nterace that s located the closest to the end o the lst. I there are n congested nteraces n the lst then the set o lows that reduce ther sendng rate based on the congeston normaton o the nterace s computed as ollows: Φ = Φ Φ (10) j < j n Then the new rate or each low that belongs to the set Φ s computed accordng to equatons (5) or (7) presented n Secton 2. Appendx B. Determnng the congeston status The nterace s congested the ollowng nequalty holds: R R > C (11) U We use symbol U to denote the set o all the ngress nodes that adjust ther sendng rates because o the congeston stream and symbol R to denote the amount o bandwdth released by the ngress node U due to congeston. Interace computes the amount o bandwdth released by the ngress node due to the congeston stream, whch we call rate decrease, as the derence between the ar shares o the ngress node beore and ater the RLR change phase. Equatons (13) and (14) show how the rate decrease by the ngress node s computed or the proportonal and maxmzng utlty dentons o arness respectvely. beore ater R = FS FS (12) R = + C b ( b b ) K b b b ( B b ) ( B b ) K ( B b ) ( B b ) 1 (13) ( ) B b C b B b R = b + (14) We used superscrpt to denote normaton assocated wth the stream nteraces. We use notaton K to denote the set o the stream nteraces that cause ngress nodes to reduce ther sendng rates. A more detaled descrpton o the rate decrease computaton can be ound n [5].