Reducing Load Distribution Overhead With Message Aggregation

Transcription

1 roc. ICCC-2003, 22 nd IEEE Internatonal erormance Computng and Communcatons Conerence, hoenx, AZ (Aprl 2003), pp Reducng Load Dstrbuton Overhead Wth Message Aggregaton Vasl Hnatyshn and Adarshpal S. Seth Department o Computer and Inormaton Scences, Unversty o Delaware, Newar, DE {vasl, seth}@cs.udel.edu Abstract The current best eort approach to Qualty o Servce n the Internet can no longer satsy a dverse varety o customer servce requrements, because o whch there s a need or alternatve strateges. A promsng approach or dealng wth ths problem s a method called Load Dstrbuton Scheme (LDS) whch dynamcally adjusts trac load at the networ boundary based on eedbac rom the networ. In order to arly share avalable resources among ndvdual lows, the load dstrbuton scheme reles on a message exchange protocol whch n certan cases may cause sgncant overhead n the system. In ths paper, we examne the ssues related to the problem o the message overhead n the LDS, propose solutons to the problem, and evaluate these solutons through smulaton n ONET. 1. Introducton The current approach to provdng Qualty o Servce n the Internet s no longer adequate because o the ncreasng emergence o applcatons wth dverse customer servce requrements. As people become wllng to pay more or servces that satsy ther applcaton needs, the one-servceor-all approach o today s Internet wll become obsolete, creatng a need or alternatve strateges. In order to tacle ths problem, a number o servce derentaton models have been proposed. Integrated [3] and Derentated [2] Servce archtectures ntroduced by IETF s IntServ and DServ worng groups, core-stateless ar queung [24], and proportonal servce derentaton ramewor [6-8] are currently among the most popular approaches. Unortunately, these schemes oten al to provde proper servce derentaton or may not be applcable to current networs. For example, Derentated Servces model may al to provde ar resource allocaton and ar servce degradaton durng perods o congeston [9, 19, 21, and 23] because o statc resource allocaton. Integrated Servce approach, on the other hand, guarantees per-low QoS but does not scale well to large networs due to per-low normaton stored n the networ core. The proportonal servce derentaton model does not volate relatve guarantees under any networ condton. However n ths model, the lac o mechansms or lmtng the amount o data njected nto the networ can reduce the absolute level o QoS below user expectatons [6-8]. We beleve that one way to deal wth ths problem s to ntroduce a load dstrbuton scheme (LDS) [11 13] at the networ boundary. The man objectves o the load dstrbuton scheme are to satsy mnmum per-low QoS guarantees and to arly dstrbute excess resources among the lows. In partcular, the LDS guarantees that each actve low n the networ wll receve at least ts mnmum requested amount o bandwdth. In ths paper, we extend wor publshed n [11 13] and urther study the perormance o the LDS under more realstc condtons. In partcular, we examne the behavor o the LDS n a networ wth a large number o small lows. Snce the LDS reles on a message exchange protocol to dstrbute QoS requrements among ndvdual nodes, t s expected that n a networ wth a large number o small lows the control message overhead wll be proportonal to the number o actve lows. Subsequently, n such cases the message exchange protocol o the LDS mght cause too much overhead n the system and would rase scalablty concerns. In ths paper, we ntroduce a technque or reducng the total number o control message generated upon actvaton or termnaton o a low. Ths approach, whch we call message aggregaton, merges multple control messages nto a sngle pacet. We examne how eectvely message aggregaton reduces the overhead o the control message exchange and study ts nluence on the resource allocaton by the LDS. The rest o the paper s organzed as ollows. Secton 2 provdes a denton o arness and an overvew o the message exchange protocol o the LDS. In Secton 3 we ntroduce the dea o message aggregaton and we present ts evaluaton through smulatons wth ONET n Secton 4. Secton 5 provdes a related wor overvew and nally we conclude n Secton The LDS Overvew and Denton o Farness In order to satsy mnmum per-low guarantees and to provde ar resource allocaton the load dstrbuton scheme reles on the networ eedbac and admsson control. When a new low actvates, the boundary node probes the networ

2 to determne the new low could be admtted nto the networ wthout volaton o the mnmum QoS requrements o the currently actve lows. In ths paper we wll not dscuss admsson control. However, we mply that a new low can enter the networ only there are enough resources to satsy mnmal QoS requrements o all the lows n the doman. Fgure 1 llustrates the dea o the LDS. As the gure shows, trac enters the networ doman at the boundary router 1 B1, traverses the networ n some ashon, and then exts ths networ doman at the boundary router B2. When a new low actvates or termnates, the boundary node advertses the change n the QoS requrements on the path. I congeston arses, the core routers dstrbute the aggregated QoS requrements through the congeston notcatons sent to the networ boundares. Based on ths eedbac, the boundary routers arly adjust the amount o trac admtted nto the doman. Incomng trac LDS B1 Fgure 1. Scenaro or Load Dstrbuton Scheme 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, whch we dene as ollows. In addton to the low RLRs, each ngress node eeps trac o the path RLRs. The path RLR, [ b, B ], or the RLR o the ngress node on the path, s a load range where b corresponds to the sum o the mnmum requested rates o the lows that orgnate rom the ngress node and traverse the path, whle s the sum o the correspondng requested maxmum rates. Usng to denote that low traverses path, we dene the path RLR as ollows: b = b B = B (1) C1 Feedbac Change n QoS Requrements B 1 Throughout ths paper we wll also reer to the boundary nodes at whch trac enters a doman as ngress routers and we wll call the boundary nodes where trac leaves a doman as egress routers. C2 C3 B2 Outgong Trac where only lows orgnatng at boundary node are consdered. 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. b B B (2) b = = To avod conuson, we wll use an upper-case letter (e.g. ) 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 [12, 13]. Congeston notcaton messages, whch are sent by a congested core nterace to ngress nodes, carry nterace and aggregated nterace RLRs. These values allow ngress nodes to arly dstrbute avalable resources among ndvdual lows. The ar shares on congested nterace o ngress node and o low are computed as ollows: B ( ) b C b B B b = mn b +, (4) ( ) B b = mn b + b, B (5) B b where C s the capacty o the outgong ln on nterace, whle and denote the ar shares o ngress node and o low on congested nterace, respectvely. Other alternatve dentons o arness wthn the ramewor o LDS were examned n [12, 13]. 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. erodc probng s used to determne the ngress node can ncrease ts sendng rate on the path. In the

3 presence o excess bandwdth, the ngress node ncreases transmsson rates o the lows that travel on the correspondng path proportonally to the ndvdual low s RLR. The path probng ntated due to 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 hase. Durng the thrd phase congested nteraces noty ngress nodes to slow down. Upon arrval o the congeston notcaton message the ngress nodes compute ther correspondng ar shares and adjust transmsson rates o ndvdual lows accordngly. A specal case occurs when the ngress node that transmts data at the rate hgher than ts ar share due to the presence o the excess bandwdth on the path, receves a congeston notcaton. In ths case, the ngress node mght not need to reduce the transmsson rate to ts ar share. Instead, the ngress node reduces ts rate proportonally to the RLR change on the path. The message exchange protocol uses the ollowng message types. In the rst phase, ngress nodes generate ROBE pacets and receve results o the path probng va ROBE_RELY 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. Message Aggregaton The load dstrbuton scheme reles heavly on the message exchange protocol to dstrbute the nterace and the aggregated nterace RLR among the boundary nodes. In a networ wth a small number o lows, the overhead due to the message exchange protocol o the LDS s neglgble as reported n [12 and 13]. In such networs, the major cause o the overhead s the perodc path probng. Snce the probe messages are nrequent and ther szes are sgncantly smaller than the average sze o the data pacet, the total overhead due to the control messages s very small. However, n a networ wth a large number o small lows that actvate and termnate very requently, the LDS scheme wll constantly reman n the RLR change phase o the message exchange protocol. As a result, the ngress nodes would generate the RLR_CNG pacets or each low actvaton or termnaton, whch may cause a sgncant overhead. Furthermore, numerous RLR_CNG messages may cause requent changes o the congeston status n the networ, whch would subsequently result n addtonal overhead due to congeston notcaton messages. The ey to reducton o the message exchange overhead s to lmt the number o the RLR_CNG messages. To do that, the boundary nodes should combne requent updates o the RLR normaton by carryng the normaton about the multple requests or low actvaton or termnaton n a sngle RLR_CNG message. We wll call ths technque message aggregaton. It should be noted that the boundary nodes do not generate an RLR_CNG message upon each low actvaton or termnaton, then the nterace and aggregated nterace normaton stored n the networ core would not be accurate, whch n turn may nluence the arness o the LDS. In act, the message aggregaton technque reduces the overall overhead due to the control messages at the cost o volatng strct guarantees o the ar resource dstrbuton o the LDS. Because o that, we conclude that the ngress node s ar share on that path s one o the parameters that determne the low s request or actvaton or termnaton could be aggregated n subsequent RLR_CNG messages. Let us consder the stuaton when the boundary node receves a request rom the low to be actvated on the path, wth the bottlenec ln. In ths case, the boundary node should compute and compare ts ar shares on the path, or the case when the RLR_CNG message was generated and when the low s request was aggregated (e.g. RLR_CNG message was not generated): noaggr = ( ) ( ( ) ) ( ( ) ) ( ) B + b B b C b b B + B B b + B b mn b + b +, (6) AGGR B ( ) b = mn b + C b, B (7) B b where, noaggr s the ar share o the ngress node on the path n the case when the ngress node advertses RLR change on the path and AGGR s the ar share o the ngress node on the path when the low s request was aggregated. It should be noted that the request or the low actvaton was aggregated then the lows that ollow the path receve the amount o resources slghtly below ther correspondng ar shares. On the contrary, the request or the low termnaton was aggregated, then the lows that ollow the path receve the amount o resources slghtly above ther correspondng ar shares. I the devaton rom the low s ar share s wthn an acceptable range, then the node should aggregate the low s request n the subsequent RLR_CNG messages. Otherwse the ngress node should ntate the RLR change phase. Ths s an example o the smplest message aggregaton polcy that reles only on the ngress node s ar shares to

4 determne the low s request could be aggregated. However, the boundary nodes are allowed to mplement more complex message aggregaton polces that would nclude the networ status or other parameters n the decson mang process. 4. Evaluaton o the Message Aggregaton 4.1 Smulaton Setup To study and evaluate the perormance o message aggregaton, we perormed a smulaton study usng the ONET networ smulator [17]. The goal o the smulaton study was to examne how the message aggregaton technque nluences the overhead n the system as well as to nvestgate ts eects on the arness o the LDS. In order to study how the message aggregaton nluences the arness o the resource dstrbuton n the networ, we ntroduce a new term called degree o arness dened as ollows. FT 1 FT 12 Edge 1 FT Flows [30s, 500s] DEST 5 DEST 1 DEST 2 Edge 6 Edge 5 Core 2 Core 5 Core 3 Edge 2 Edge 3 Edge 4 DEST 3 DEST 4 DEST 5 termnates at tme 250 seconds, and travels to Destnaton 1. The duraton o the smulaton was 500 seconds. Flow Numbers Flow s Actvaton/ Termnaton Schedule Flow s RLR (Kbps) Ingress Node FT 1-3 [30 s, 500 s] [20, 50] Edge 1 FT 4-6 [30 s, 500 s] [40, 50] Edge 1 FT 7-10 [30 s, 500 s] [10, 30] Edge 1 Vdeo 1 [185 s, 500 s] [400, 1200] Edge 1 Vdeo 2 [60 s, 500 s] [200, 1000] Edge 1 Vdeo 3 [160 s, 250 s] [800, 2000] Edge 2 Vdeo 4 [80 s, 350 s] [500, 1300] Edge 3 Table 1. Flow speccaton In the smulaton we used two types o applcatons: FT and vdeo trac. FT lows are small, short lved lows that actvate and termnate very requently. Each FT low randomly actvates multple tmes durng the smulaton and remans actve or a random duraton not longer than 30 seconds. FT lows use TC as ther transport protocol. Vdeo trac conssts o the large, longlved lows that use UD as the transport protocol. Each vdeo low actvates only once durng the smulaton and remans actve accordng to the schedule shown n Table 1. All the vdeo trac n the smulaton s b-drectonal; however, n order to avod unnecessary conuson we wll not dscuss vdeo trac that travels rom the boundary nodes Edge 4, Edge 5, and Edge 6 to ther correspondng destnatons. SRC 1 SRC 2 SRC 3 SRC 4 [185 s, 500s] [60 s, 500s] [160 s, 250 s] [80 s, 350 s] DEST 4 DEST 3 DEST 1 DEST 2 Fgure 2. Smulaton Topology We wll reer to the ar shares o the boundary node on the path at the tme τ wth and wthout the message aggregaton as ma ( τ ) and ( τ ), respectvely. Then, the degree o arness o the boundary node on the path at the tme τ s dened as: DF ( τ ) = 1 ma ( τ ) ( τ ) 1 (7) It should be noted that wth ths denton, a hgh degree o arness (e.g. more than 95%) ndcates that the resources are shared arly among the lows, whle a low degree o arness ndcates unar sharng. To study the perormance o the message aggregaton we used the topology o Fgure 2 that shows the low s actvaton/termnaton schedule and the pont o destnaton or each low. For example, the low o Source 1 actvates at tme 185 seconds and travels to node Destnaton 4, whle the low o source 3 actvates at tme 160 seconds, Fgure 3. Load Dstrbuton at Edge Load Dstrbuton Usng LDS Let us examne the bandwdth allocaton by the LDS n greater detal. The load dstrbuton among ngress nodes or the scenaro dened by Fgure 2 and Table 1 s shown n Fgures 3 5.

5 At tme 30 seconds the FT lows begn to actvate n a random ashon. At tme 60 seconds low Vdeo 2 actvates ollowed by actvaton o low Vdeo 4 at tme 80 seconds. At ths pont the ln between the nodes Core 5 and Core 3 becomes congested whch orces the edge nodes 1 and 3 to throttle transmsson rates o ther correspondng lows. At tme 160 seconds low Vdeo 3 actvates whch shts the bottlenec or ngress node Edge 1 to ln Core 2 Core 5. As a result, ngress nodes Edge 1 and Edge 2 adjust transmsson rates o ther lows accordng to aggregated nterace RLR on ln Core 2 Core 5, whle ngress node Edge 3 benets rom the excess bandwdth created by throttlng the lows o ngress node Edge 1. All the actve lows n the networ adjust ther transmsson rates accordng to ther correspondng bottlenec lns upon actvaton o the low Vdeo 1 at tme 185 seconds. However, ater all the lows have adjusted ther transmsson rates, ln Core 5 Core 3 becomes underutlzed whch enables low Vdeo 4 to benet rom the excess bandwdth. 5 Core 3, resultng n unnecessary load luctuatons. Thus, upon CN message arrval, the boundary node may reduce ts transmsson rate proportonally to the RLR change on the congested nterace, nstead o sendng trac at ts ar share. Fgure 5. Load Dstrbuton at Edge 3 Fgure 4. Load Dstrbuton at Edge 2 In should be noted that requent actvatons o the FT lows oten cause congeston on both bottlenec lns Core 2 Core 5 and Core 5 Core 3. As a result, all the actve lows n the networ adjust ther transmsson rates upon actvaton o the FT lows. However, snce ln Core 5 Core 3 does not lmt transmsson rate o the FT lows, low Vdeo 4 need not reduce ts transmsson rate to ts ar share on the ln Core 5 Core 3. Instead, low Vdeo 4 reduces ts transmsson rate proportonally to the RLR o the newly actvate FT low. As long as ln Core 2 Core 5 remans the bottlenec or the trac rom ngress node Edge 1, ln Core 5 Core 3 wll contan excess bandwdth and low Vdeo 4 would utlze t. I upon recepton o every CN message, low Vdeo 4 adjusts ts transmsson rate to the correspondng ar share then later t would ncrease ts sendng rate because o the excess bandwdth avalable on the ln Core At tme 250 seconds low Vdeo 3 termnates whch causes the bottlenec or the trac rom Edge 1 to sht bac to ln Core 5 Core3. Consequently, all the actve lows adjust ther transmsson rates accordngly. In partcular, low Vdeo 4 adjusts ts sendng rate to ts ar share n order to accommodate trac rom Edge 1. Fnally, at tme 350 seconds, low Vdeo 4 termnates and trac rom Edge 1 utlzes all the avalable bandwdth on the path. It should be noted that throughout the smulaton each actve low receves an amount o bandwdth that s wthn ts requested load range. Furthermore, the per-low bandwdth allocaton n the networ satses (wthn small error lmts) the arness crtera dened by equaton 5. In the next secton we examne the nluence o the message aggregaton on the arness o the load dstrbuton by the LDS. 4.3 Evaluaton O The Message Aggregaton To evaluate the message aggregaton technque and ts nluence on the arness o the load dstrbuton, we mplemented the ollowng aggregaton polces or the scenaro o Fgure 2. As mentoned beore, the goal o the message aggregaton s to reduce the total number o control messages (RLR_CNG) generated upon the low actvaton or termnaton. The ollowng message aggregaton rules specy under what condtons the RLR_CNG message should be generated and when t should be aggregated. Rule 1. Always generate RLR_CNG message upon actvaton or termnaton o the vdeo low. Rule 2. Always generate RLR_CNG message the low s actvaton does not cause congeston.

6 Rule 3. Generate RLR_CNG message upon the actvaton/termnaton o the FT low, the rato between noaggr AGGR and s larger than the aggregaton noaggr threshold, where, s the ar share o the FT lows on the path when the RLR_CNG message was sent AGGR and s the ar share o the FT lows on the path when the low s request was aggregated. Rule 4. Otherwse do not generate control message. Reducton n the RLR_CNG generatons Average Degree o Farness 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 30% 40% 50% 60% 70% 80% Aggregaton Threshold Fgure 6. Reducton n RLR_CNG messages 99.8% 99.6% 99.4% 99.2% 99.0% 98.8% 98.6% 98.4% 98.2% 98.0% 30% 40% 50% 60% 70% 80% Aggregaton Threshold Fgure 7. Average Degree o Farness We examned the reducton n the total number o RLR_CNG messages and the varaton o the degree o arness at the ngress node Edge 1 by changng the value o the aggregaton threshold rom 30% to 80%. Each scenaro was executed 10 tmes and the averaged results presented n Fgures 6 9. As expected, the reducton n the total o the RLR_CNG messages ncreases whle the degree o arness decreases as the aggregaton threshold becomes larger. When the aggregaton threshold ncreases, a larger number o low requests are beng aggregated, as a result o whch ewer RLR_CNG messages are generated. Subsequently, the core nodes contan and advertse a less accurate value o the aggregated nterace RLR, whch nluences the accuracy o the load dstrbuton and causes the degree o arness to decrease (mplyng less arness). Control Load Reducton Control Load (Kbytes) 9% 8% 7% 6% 5% 4% 3% 2% 1% 0% % 40% 50% 60% 70% 80% Aggregaton Threshold Fgure 8. Control Load Reducton 30% 40% 50% 60% 70% 80% Aggregaton Threshold Fgure 9. Total Control Load Smulaton results collected or the smulaton topology o Fgure 2 showed that the message aggregaton sgncantly reduces the total number o the RLR_CNG messages generated. In the best case, when the aggregaton threshold was set to 80%, the message aggregaton technque reduced the total number o RLR_CNG messages by almost 80%. At the same tme, the average degree o arness durng smulaton was only 98.4%, whch means the allocaton bandwdth values dd not devate much rom the optmally ar load dstrbuton. However, the reducton o the total control load due to message aggregaton was sgncantly lower as shown n Fgure 8. In the best case, when the aggregaton threshold was set to 80%, the total load reducton was only 8%. Nevertheless, as Fgure 9 shows, the total load o the control messages consstently decreases as the aggregaton threshold s ncreased. We observed such a small reducton n the amount o the control trac because the smulaton o Fgure 2 was congured wth a relatvely small number o FT lows. On average, durng the smulaton there were only 300 low actvaton requests by the FT sources. As a result, the ROBE messages were the domnant contrbutor to the total control load n the networ. Snce the goal o the message

7 aggregaton s to reduce the total number o RLR_CNG messages, ts eects on the reducton o the total control load are very small, only 4% -- 8%. In the scenaro where the total number o the low actvaton requests s sgncantly larger, the RLR_CNG messages would domnate the control load and thus the message aggregaton would sgncantly reduce the control load overhead. It should also be noted that although the average degree o arness vared between 99.6% and 98.4%, n certan nstances durng the smulaton ts value reached as low as 87%. Fgure 10 shows the requency dstrbuton o the degree o arness values durng the smulaton wth the aggregaton threshold set to 80%. Fgure 10 shows that the degree o arness reaches low values very nrequently and remans n that state or a very short tme. For example, value o the degree o arness s lower than 90% only 0.66% o the tme, whle t vares between 90% and 95% only 8.47% o the tme. As a result, we beleve that such behavor o the LDS due to message aggregaton s acceptable. Frequency o Occurance 35% 30% 25% 20% 15% 10% 5% 0% 86% - 87% Degree o Farness Hstorgam 88% - 89% 90% - 91% 92% - 93% 94% - 95% 96% - 97% 98% - 99% = 100% Degree o Farness Fgure 10. Dstrbuton o the Degree o Farness values 5. Related Wor Overvew Ths paper ntroduces a new dea or reducng the total overhead due to control message exchange wthn the ramewor o the LDS ntroduced n [11-13]. Ths wor s a drect extenson and mprovement o [13] where a more detaled descrpton o the LDS may be ound. In [11] Hnatyshn et al provde an alternatve approach to load dstrbuton n the Internet. The LDS proposed n [11] does not requre the core nodes to mantan aggregated nterace RLR and nstead reles on an approxmaton mechansm or computng the ar share o the ngress nodes. However, ths approach cannot accurately compute the ar shares, taes a long tme to converge, and does not guarantee ar load dstrbuton among the edge nodes under all networ condtons. In [14], Kar et al provded an excellent denton o the dynamc rate control problem and ntroduced an teratve algorthm that solves t. In [14], 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 [14] 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 [14] becomes unacceptable because o these eatures. Mrhaa et al ntroduced a somewhat related dea n [18]. Ther goal was to mody the resource reservaton protocol RSV or supportng dynamcally changng QoS requrements n moble ad hoc networs. The proposed drsv 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 drsv also wors 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 [22] 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 [14] 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-oservce attac. The problem o admsson control [10, 15-16] and controlled-load servces [24] s somewhat related to the load dstrbuton ssues dscussed n ths paper. However they address a slghtly derent problem o determnng when a new low could 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. 6. Conclusons In ths paper we examned perormance o the LDS n a networ that contans a large number o small lows and ntroduced a message aggregaton technque or reducng the control load overhead. The message aggregaton reduces the total number o RLR_CNG control messages at the cost o volatng the arness requrements o the resource

8 dstrbuton. Smulaton results reported that the LDS provdes a ar and ecent resource allocaton n the networ wth a large number o small lows. Furthermore the message aggregaton s capable o sgncantly reducng the total number o RLR_CNG messages at the small cost o less than 2% o average devaton rom the optmally ar load dstrbuton. Although the message aggregaton showed very promsng results, t should not be used n networs where the domnant part o the message exchange overhead comes rom the perodc path probng. In ths case, the message aggregaton technque would be neectve. Furthermore, smulaton results showed that the overhead due to the perodc path probng s very nsgncant, less than 0.1% o the total load. As a result, the message aggregaton should be used only n networs where the RLR_CNG messages due to low actvaton/ termnaton domnate control load overhead. 7. Reerences 1. M. Allman, A Web Server s Vew o the Transport Layer, Computer Communcatons Revew, vol. 30, no. 5, October 2000, pp S. Blae, D. Blac, M. Carlson, E. Daves, Z. Wang, and W. Wess. "An Archtecture or Derentated Servces", December IETF RFC R. Braden, D. Clar, and S. Shener, "Integrated Servces n the Internet Archtecture: an Overvew", June IETF RFC H. Chow and A. Leon-Garca, "A Feedbac Control Extenson to Derentated Servces", March Internet Drat: drat-chow-dserv-bctrl.txt. 5. D. Clar and W. Fang, "Explct Allocaton o Best Eort acet Delvery Servce," IEEE/ACM Transactons on Networng, vol. 6, no. 4, August 1998, pp C. Dovrols and. Ramanathan, "A Case or Relatve Derentated Servces and roportonal Derentaton Model," IEEE Networ, vol. 13, no. 5, Sep. 1999, pp C. Dovrols and D. Stllads, "Relatve Derentated Servces n the Internet: Issues and Mechansms," roceedngs o ACM SIGMETRICS, May C. Dovrols, D. Stllads, and. Ramanathan, "roportonal Derentated Servces: Delay Derentaton and acet Schedulng," roceedngs o ACM SIGCOMM 99 Conerence, Cambrdge, MA, Sep. 1999, pp W. Feng, D. Kandlur, D. Saha, and K. Shn "Understandng and Improvng TC erormance over Networs wth Mnmum Rate Guarantees," IEEE/ACM Transactons on Networng, vol. 7, no. 2, Aprl 1999, pp R. Gbbens and F.. Kelly, "Dstrbuted connecton acceptance control or a connectonless networ," roceedngs o ITC 99, Ednburgh, UK, June V. Hnatyshn and A.S. Seth, Avodng Congeston Through Dynamc Load Control, SIE's Int l Symp. on The Convergence o Inormaton Technologes and Communcatons, Aug. 2001, pp V. Hnatyshn and A.S. Seth, rovdng per-low QoS usng load dstrbuton scheme, TR , Department o Computer and Inormaton Scence, Unversty o Delaware, February V. Hnatyshn and A.S. Seth, Far and Scalable Load Dstrbuton n the Internet, roceedngs o the Internatonal Conerence on Internet Computng, Las Vegas, NV, June 24-27, 2002, pp K. Kar, S. Sarar, and L. Tassulas, "A Smple Rate Control Algorthm or Maxmzng Total User Utlty," roceedngs o INFOCOM 2001, Anchorage, USA, Aprl F.. Kelly, A. K. Maulloo, and D. K. H. Tan, Rate Control or communcaton networs: shadow prces, proportonal arness and stablty, Journal o the Operatonal Research Socety, vol. 49, no 3, pp , March F. Kelly,.B. Key, and S. Zachary, "Dstrbuted Admsson Control," IEEE Journal on Selected Areas n Communcatons, vol. 18, no. 12, December 2000, pp A.F. Lobo and A.S. Seth, ''A cooperatve congeston management scheme or swtched hgh-speed networs,'' roceedngs o ICN-96, Internatonal Conerence on Networ rotocols, Columbus, Oho (Oct.-Nov. 1996), pp M. Mrhaa, N. Schult, and D. Thomson, "Dynamc Qualty-o-Servce or Moble Ad Hoc Networs," Frst Annual Worshop on Moble and Ad Hoc Networng and Computng, 2000, pp B. Nandy, N. Seddgh,. eda, and J. Ethrdge, "Intellgent Trac Condtoners or Assured Forwardng Based Derentated Servces Networs, roceedngs o IFI Hgh erormance Networng (HN 2000), June ONET Modeler. ONET Technologes Inc eda, N. Seddgh, and B. Nandy, "The Dynamcs o TC and UD Interacton n I-QOS Derentated Servces Networs," roceedngs o the 3rd Canadan Conerence on Broadband Research, November K. Ramarshnan, S. Floyd, and D. Blac, "The Addton o Explct Congeston Notcaton (ECN) to I", March Internet Drat: drat-et-tsvwg-ecn-03.txt. 23. J. Rezende, "Assured Servce Evaluaton", roceedngs o Globecom 99, March J. Wroclaws, Speccaton o the Controlled-Load Networ Element Servce, September IETF RFC 2211.