Adaptive Queue Management wi Restaint on Non-Responsive Flows Lan Li and Gyungho Lee Depatment of Electical and Compute Engineeing Univesity of Illinois at Chicago 85 S. Mogan Steet Chicago, IL 667 {lli, ghlee}@ece.uic.edu ABSTRACT This pape poposes an adaptive queue management scheme (adaptive RED) to impove Random Ealy Detection (RED) on estaining non-esponsive flows. Due to a lack of flow contol mechanism, non-esponsive flows can stave esponsive flows fo buffe and bandwid at e gateway. In ode to solve e dispopotionate esouce poblem, RED famewok is modified in is way: on detecting when e non-esponsive flows stave e queue, packet-dop intensity (Max_p in RED) can be adaptively adusted to cub non-esponsive flows fo esouce fai-shaing, such as buffe and bandwid fai-shaing. Based on detection of taffic behavios, intentionally estaining nonesponsive flows is to incease e oughput and decease e dop ate of esponsive flows. Ou expeimental esults based on adaptive RED shows at e enhancement of esponsive taffic and e bette shaing of buffe and bandwid can be achieved unde a vaiety of taffic scenaios. Keywods Active queue management, dop pobability, non-esponsive flows, adaptive RED.. INTRODUCTION Congestion contol heteogeneity in e Intenet leads to fainess poblems of esouce shaing. Especially, non-esponsive flows do not back off ei tansmission ates in esponse to congestion indications of e netwok (i.e. packet loss) [5]. Consequently, non-esponsive flows tend to consume unfaily esouce high bandwid and buffe space [6] at congestion points. A lage numbe of UDP-based applications geneate is kind of flows. It is necessay to employ a flow management mechanism fo e fai utilization of netwok esouce. It may also be helpful to satisfy moe uses QOS equiement. Thee ae two kinds of flow management mechanisms at ty to achieve e esouce fai shaing: scheduling scheme and queue management scheme. Scheduling schemes have geneally too much complexity and low scalability to lage numbe of flows. If we plan to povide fai utilization on high-speed edge oute, queue management scheme could be a bette choice. Queue management scheme not only has less complexity, but also appoximates fainess bette. Route wi a queue management scheme maintains a single FIFO shaed by all flows [] and employs a dopping algoim to discad aiving packets when congestion builds up. Dopping pobability inceases wi e aise of congestion level. By adusting dopping pobability fo in-queue flows, it can achieve a bette fainess fo e esouce utilization between flows. One of classic queue management schemes is Random Ealy Detection (RED). The RED utilizes a andom dopping algoim [4]. In a RED gateway, incipient congestions ae detected by estimated aveage queue size. Aveage queue size is calculated via a low-pass filte. When e aveage queue size exceeds a peset eshold, e gateway dops packets wi a cetain pobability at is calculated based on e aveage queue size. In e RED gateway, occasional busts of packets in e queue ae allowed and aveage queue size is kept at low level [6]. Alough RED deceases loss-bias against esponsive taffic at exists in e Dop Tail queue, it does not achieve a faishae of e buffe and bandwid between esponsive flows and non-esponsive flows []. One possible eason is at given a cetain time slot e pecentage of packets dopped fom each flow is almost e same. As a esult, non-esponsive flows may consume a lage potion of e bandwid at congestion points because of its lack of flow contol mechanism [3]. Consequently, non-esponsive flows can stave out e esponsive flows. In ode to solve e dispopotionate esouce poblem, seveal vaiants of RED have been poposed, such as Flow Random Ealy Dop (FRED) [7], Refined Design of Random Ealy Detection Gateways [6] and uing RED Gateway [9]. Howeve, FRED maintains exta flow state which inceases implementation ovehead. As fo e Refined Design of RED, it dynamically adusts e Wq (queue weight) and Max_p (maximum dop pobability) wi espect to e vaiance of queue size change. Howeve, it divides e aea between minimal eshold and maximal eshold, called Yellow aea [4], into seveal sub-phases at inceases implementation ovehead as well. In addition, it still does not conside e fai shaing of e buffe and bandwid between esponsive flows and non-esponsive flows. uing RED Gateway also employs an adaptive paameteization at RED gateway. Wiout consideing non-esponsive taffic, adaptive paameteization is only based on e congestion level indicated by e aveage queue leng. It can effectively adapt to taffic scenaios wi diffeent congestion levels. Howeve, it does not distinguish esponsive and non-esponsive taffic. In some cases of medium congestion, esponsive taffic may get suffeed due to e unfiendly behavios of non-esponsive taffic. If we can distinguish non-esponsive flows fom esponsive flows, a esouce fai-shaing between em can be popely managed. Unfotunately, any meods based on pe-flow state must incease e implementation ovehead. Howeve, we ae awae of e appoaches at ty to appoximately discen nonesponsive flow and esponsive flow accoding to cetain taffic featues [5]. Such a featue could be an extaodinay high congestion o an impope eaction to congestion notification. 4 SYSTEMICS, CYBERNETICS AND INFORMATICS VOLUME - NUMBER 6
We en popose to detect flow s behavio based on e vaiance of instantaneous queue size. Based on e detection, we can penalize non-esponsive flows and impoves esponsive flows oughput via adaptive adustment of e dop pobability. The est of e pape is oganized as follows. Section pesents flow behavios in conunction wi e RED queue and notes some flaws of RED algoim. Section 3, descibes ou poposed queue management scheme Adaptive RED wi estaint on non-esponsive flows. In Section 4, we compae ou appoach wi oe schemes via simulation. Finally, we give discusses and conclusions in Section 5.. RED Scheme Oveview. BACKGROUND RED (Random Ealy Detection) is a congestion avoidance mechanism employed at gateways. Its basic idea is to use two peset esholds to detect incipient congestion and contol e aveage queue size wi e eshold []. Wi espect to e estimated aveage queue size, e gateway woks in one of ee woking states: ed, geen and yellow states [6]. When e aveage queue size is less an e Min_ (minimum eshold), e gateway woks in e geen state wiout packet dop. When e leng is between e Min_ and Max_ (maximum eshold), e state tansfes to yellow one in which aiving packets ae andomly dopped wi a pobability calculated on aveage queue leng. When e leng goes beyond e maximum eshold, e gateway eaches e ed state in which evey aiving packet is discaded [6]. The goal of RED scheme is to detect incipient congestion via aveage queue size and pefom andom packet dopping befoe e queue is full. Theefoe, it can implement congestion avoidance while allowing shot-tem busty. Alough RED pevents packets fom consecutive dopping (especially fo busty taffic) and emoves highe loss bias against busty taffic [4], it still has some flaws:. RED is unable to efain non-esponsive flows fom consuming most buffe space at heavy congestion [3].. Aveage queue size is not a good estimato of congestion. Sevee congestion should be elated to e amount of flows and e degee of bustiness []. 3. It is not clea how to select e RED paametes. RED equies selecting vaious paametes in diffeent congestion scenaios. That is why ecommended values had been changed ove time [].. Flow Behavios in e RED Queue Geneally speaking, most of TCP flows wi congestion contol ae egaded as esponsive flows, while UDP flows wiout congestion esponse ae consideed non-esponsive flows [5]. TCP congestion contol utilizes an Additive Incease/Multiplicative Decease mechanism to adust e flow ate [8]. This mechanism causes e bustiness of TCP flows. Since TCP congestion contol is based on end-to-end feedback, e contol esponse is elatively sluggish in compaison wi queue opeation. It implies at TCP would keep e sending ate fo a while afte packet dopping. The contol latency is equal to e TCP timeout, which is usually fou times as much as RRT (Round Tip Time) [8]. On e oe hand, nonesponsive flows have no congestion contol mechanism. They do not espond to congestion notification and us send packets as many as ey want. At e RED-suite gateways, TCP flows (esponsive flow) tend to decease sending ate afte aveage queue size goes beyond e minimal eshold. Howeve, RED does not penalize nonesponsive flow, because packet dop atio of each flow ove time is almost same. Alough UDP flows (non-esponsive flow) wi high bandwid may suffe moe packet dopping, ey obtain much moe oughput an TCP flows do. Consequently, non-esponsive flows may take up e most bandwid and stave out esponsive flows [3]. Since RED uses static paametes to deal wi dynamic taffic, e effectiveness of RED depends on e appopiate paameteization. It is known at no single set of RED paametes woks well unde diffeent congestion scenaios, including e case at TCP is mixed wi UDP..5..5..5 Total (including CBR UDP flows, 5k each).3.63..5.5.5 Max_p 4 (a) Two 5k UDP flows ae mixed wi TCP flows.5..5..5 Total (including CBR UDP flows, k each).3.6..3.5.5 Max_p 6 3 64 (b) Two k UDP flows ae mixed wi TCP flows Figue. TCP loss ate vs. Max_p Simulation esult in Figue shows at e total loss ate is petty high fo e small Max_p. As Max_p inceases, loss ate deceases since e RED queue is able to send congestion notification back to e souces in time to pevent continuous buffe oveflow. Finally, when Max_p becomes quite lage, e RED queue causes an incease in packet loss ates ove Dop Tail queues. As moe connections ae added, e optimal value of Max_p inceases. It clealy suppots at an adaptive scheme fo dop pobability is equied. Moe extensive and detailed discussion can be also found in [9]. The details of e simulation ae fully descibed in Section 4. 3. ADAPTIVE RED WITH RESTRAINT ON NON-RESPONSIVE FLOWS Though TCP flows ae dominant taffic in Intenet, moe and moe multimedia applications tend to geneate UDP flows, mostly belonging to non-esponsive flows. It is because ose SYSTEMICS, CYBERNETICS AND INFORMATICS VOLUME - NUMBER 6 4
applications mainly concen e delay constaint instead of data integity. It is necessay to deploy a flow management mechanism to implement a fai utilization of netwok esouce. Dop Pobability 3. Basic assumption Given at packet-pocessing ate is invaiable (sevice ate: S ) and non-esponsive flows have e aival ate R non (t), e instant queue size is detemined by: t Q inst ( t ) = ( R non ( t ) + R ep ( t ) S ) dt () t i, whee R ep (t) is e aival ate of esponsive flows; t i is e initialization time. The aveage ate of n TCP-fiendly flows (esponsive flow) can be appoximately calculated as []: n 3 packets ep = RTT p second (), whee RTT is e ound tip time and p is e loss pobability of flow. It is applicable only when e p is small. Accoding to e Eq. (), given a small time slot T wi e size of t, change in queue size is mainly detemined by e aival ate of flows: Qinst ( τ ) Rnon( τ ) + Rep( τ ) S t Consideing two consecutive slots T and T wi e same + size t, e vaiance of aival ate ( = non + ep ) can be appoximated by: Qinst ( τ + ) Qinst ( τ ) t, wheeτ T andτ + T (3) + Suppose ep is e aveage aival ates of esponsive flows at slot T and t RTT, extending Eq. () e vaiance of esponsive flows aival ate p ep ( ) ep p + (4), whee p and p + ae dop pobability at T and T. + 3. RED Queue Analysis The vey impotant featue of RED algoim is e use of a low-pass filte to estimate e aveage queue size [6]. The lowpass filte employs e EWMA (exponentially-weighted moving aveage) given by [4]: Q ave = ( wq ) Qave' + wqqinst,whee w is e weight facto. q Q ave and Q ave ' ae e cuent and pevious aveage queue size espectively. Figue shows e dop function of RED. Dop ate of RED is pesented by:, Qave < Min_ (5) Qave Min_ Ded = Max_ p, Min_ Qave < Max_ Max_ Min_, Max_ Qave Extending Eq. (4), we can have: τ Qave ( ) Min _ ep ( ) ep Qave ( τ + ) Min _ (6) Max_p Min_ Max_ Q ave Figue. Dop function of RED The tendency of esponsive flows can be detected via Qave ( Qave = Qave ( τ + ) Qave ( τ )). When Q ave inceases to e yellow state [], ep should decease accodingly. Howeve, if non gets invese tendency at at time, e aival ate of all flows ( ) inceases. It causes esponsive flows staved and esouce unfaily shaed. 3.3 Adaptive RED Scheme Adaptive RED is poposed to detect non-esponsive (unfiendly) flow behavios via monitoing e vaiation of queue leng, including instant and aveage queue leng. At congestion, esponsive flows decease its ate while non-esponsive flows may keep o incease e ate. Theefoe, adaptively adusting packet-dop intensity may estain non-esponsive flows to stave esponsive flows. We apply is adaptive scheme only if e aveage queue leng anging fom Mid_ ((Min_+Max_)/) to Max_. In at medium congestion aea, esponsive flows ust get essential suffeing and take deep congestion esponse. Based on ou assumption and analysis in Section 3. and 3., Q ave indicates e tendency of esponsive flows and Q ins ( ) eflects e vaiation of aival Q = Q ( τ ) Q ( τ ) inst inst + inst ate. If bo of em ae positive, must be positive non too. It implies non-esponsive flows take unfiendly esponse against esponsive flows and e atio of non-esponsive flows and esponsive flows gets inceased. At at time, unfai esouce shaing happens and e adustment of dop density is needed. In ode to avoid too aggessive of packet dopping, dop density needs to be adusted back once e condition is not satisfied ( Q inst < Qave < ). We en popose ou algoim as follows: Evey T (contol slot): While ( Max_ > Qave > Mid_) if Q inst > & Qave > Max_p = Max_p*α else Max_p =Min (Max_p, Max_p/α) Max_p : Default Max_p α: e scale facto of Max_p α is e scale facto to extend e Max_p. Max_p is e peset paamete of dop density. Based on e obsevation of Q ave and Q inst, e algoim detects unfiendly taffic 4 SYSTEMICS, CYBERNETICS AND INFORMATICS VOLUME - NUMBER 6
behavios and modulates e value of Max_p by e facto of α. The paametesα is expeimentally chosen accoding to e contol slot (T) which detemines e contol sensitivity. If contol slot is too small, ee is no enough time fo esponsive flows to esponse and e scheme may become moe aggessive and cause contol oscillation. On e oe hand, it may become insensitive and cause contol sluggishness. In e simulation, we chose contol time slot as.s (aveage RTT). This slot size can accommodate e shot bustiness and avoid e contol vibation. 4. SIMULATION RESULTS This section pesents e simulation esults of Adaptive RED s pefomance wi efaining of non-esponsive flows and enhancing e esponsive flows. The simulations ange ove netwok configuations and taffic pattens. 4. Basic netwok uation Netwok simulation topology is shown in Figue 3. Thee is a single congested link fom GW to GW in a dumbbell topology. The congested link is Mbps; oes ae Mbps each. TCP and UDP flows pass ough e congested link. The numbe of TCP flows anges fom 8 to 3, while at of UDP flows anges fom to 8. Thee active queue management schemes: RED, ue and ae applied on e bottleneck link. Node Mb 6ms Node 3 Mb 4ms Node Mb 3ms RED/ / SC GW Mb ms GW Mb ms Node 4 Mb 5ms Mb ms Mb 3ms Node 5 Mb 4ms Node 8 Node 6 Node 7 Figue 3. Topology of e Simulation To compae wi RED and schemes, a popula simulation envionment of RED-suite gateway is chosen and descibed in Table and Table. The physical queue capacity is set to 3 packets. Following [4], we set Min_ to packets, and set Max_ to be twice as e Min_. RED is simulated wi NS s default value of W q =. and.. Paametes fo ae efeed to [6] in which α=3 and β=. We expeimentally select T as.s and α as.5 to. In ode to ceate non-esponsive behavios, UDP flows fo all schemes ae chosen as exponential type wi a packet size of bytes. Queue Capacity 3 Packets Min_ Packets Max_ Packets αfo.5 T (Contol slot). s W q and Max_p fo RED. and. αβfo 3 and Table. Gateway Paametes TCP window size 5 UDP packet-size TCP packet-size UDP Type Exponential TCP Type UDP ate 4. Simulation Results FTP taffic.5 Mbps Table. Flow Paametes ) The oughput of RED, and Adaptive RED espectively 8 6 4 95 9 85 8 8 75 7 65 6 4 35 3 5 TCP Thoughput (Wiout UDP Taffic) (a) Pue TCP flows TCP Thoughput (Wi UDP flows, 5k each) (b) Mixing wi UDP flows TCP Thoughput (Wi 4 UDP flows, 5k each) (c) Mixing wi 4 UDP flows TCP Thoughput (Wi 8 UDP flows, 5k each) (d) Mixing wi 8 UDP flows Figue 4. TCP Thoughput Pefomance SYSTEMICS, CYBERNETICS AND INFORMATICS VOLUME - NUMBER 6 43
) TCP of ee schemes.6.4..5..5..5..5.3.. TCP (Wiout UDP Taffic) (a) Pue TCP flows TCP (Wi UDP flows, 5k each) (b) Mixing wi UDP flows TCP (Wi 4 UDP flows, 5k each) (c) Mixing wi 4 UDP flows T CP (Wi 8 UDP flows, 5k each) (d) Mixing wi 8 UDP flows Figue 5. TCP 3) Restaint on unfiendly non-esponsive taffic by. Aival ate atio.8.6.4. (8 TCP wi UDP).. 4. 6. 8. 4 6 8 (a) 8 TCP flows mixing wi UDP flows Map_p Map_p Aival ate atio Aival ate atio Aival ate atio..8.6.4. (8 TCP wi 4 UDP).. 44. 66. 88. 3 54 76 98 Map_p Map_p (b) 8 TCP flows mixing wi 4 UDP flows.5.5 (3 TCP wi UDP). 8. 56. 84. 4 68 96 Map_p Map_p (c) 3 TCP flows mixing wi UDP flows. (3 TCP wi 4 UDP) 8. 56. 84. 4 68 96 Map_p Map_p (d) 3 TCP flows mixing wi 4 UDP flows Figue 6. Aival Rate Ratio of UDP flows TCP oughput is shown in Figue 4. Compaed wi RED and, scheme has compehend-sively bette pefomance. In case of pue TCP (Figue 4(a)), scheme shows slightly bette. It is because adusts dop density only accoding to e congestion level. To do so can help avoid heavy and shot congestion caused by a lage numbe of TCP flows. When a lage numbe of TCP flows mix wi big UDP taffic (Figue 4(d)), scheme becomes e wost one. Accoding to e algoim, Max_p of may always be kept in a big value, if aveage queue leng fluctuates aound Max_ [9]. Such adustment to Max_p may be so aggessive at TCP suffes moe an UDP, since UDP flows always ovewhelm TCP flows so at e bandwid eleased fom TCP is taken ove by UDP. Coesponding TCP loss ate is shown in Figue 5. Basically, e decease of TCP dop ate indicates e impovement of TCP oughput. It implies at helps TCP flows scamble fo bandwid fom UDP flows. It makes esponsive flows be enhanced. Howeve, it can be found at a little bit lowe pefomance an RED is pesented when a lage numbe of TCP and UDP flows ae mixed (3 TCP flows in Figue 5(d)). The eason may be at wi sevee and continuous congestion, has less space to opeate so at it may misdetect nonesponsive behavios. The simila poblem would be found in e simulations of moe UDP flows. Howeve, ee is no need 44 SYSTEMICS, CYBERNETICS AND INFORMATICS VOLUME - NUMBER 6
to discuss ose cases because of vey sevee congestion and intoleable dop atio (beyond 3%). Those cases should be in system poblem and cannot be solved ust by queue management schemes. Figue 6 shows e effectiveness of. We sampled e aival ate atio of UDP flows ( non ) evey.5s. Most dop density adustments ae made when UDP flows occupy moe an half of e bandwid. At at time, moe UDP packets ae dopped via inceasing Max_p. In addition, adusts Max_p only when e aveage queue leng goes beyond Mid_. Theefoe, effectively efains non-esponsive taffic behavios and enhances esponsive ones as well. We would not list e simulation esult of Mixed taffic wi 6 and 8 UDP flows, because e esult does not mean much if UDP flows always ovewhelm TCP flows. As a conclusion, keeps esponsive flows moe adaptively an RED and do. It also has bette oughput pefomance and can adapt to a wide vaiety of taffic scenaios. Gateways fo Congestion Avoidance, IEEE/ACM Tansactions on Netwoking, (4), 993, 397-43. [5] S. Floyd, and K. Fall, Pomoting e Use of End-to-end Congestion Contol in e Intenet, IEEE/ACM Tansactions on Netwoking, 7(4), 999, 458-47. [6] H. Wang and K. G. Shin, Refined Design of Random Ealy Detection Gateways, Poc. IEEE GlobeCom, Rio de Janeio, Bazil, 999, 769-775. [7] D. Lin and R. Mois, Dynamics of Random Ealy Detection, Poc. ACM SIGCOMM, Cannes, Fance, 997, 7-37. [8] W. Stevens, TCP/IP illustated, Volume : e potocols (Reading, MA: Addison-Wesley, 994) [9] W. Feng, D. Kandlu, D. Saha, K. Shin, A configuing RED Gateway, Poc. IEEE INFOCOM, New Yok, NY, 999, 3-38. [] J. Padhye, V. Fioiu, D. Towsley and J. Kuose, Modeling TCP Thoughput: A Simple Model and Its Empiical Validation, Poc. ACM SIGCOMM, Vancouve, Canada, 998, 33-34. 5. DISCUSSIONS AND CONCLUSION We only applied to exponential UDP taffic in ou simulation because of its simplicity and typical non-esponsive featue. Wi egad to e basis of scheme, any unfiendly non-esponsive taffic can be detected and efained by, such as DDOS attack flows. Though DDOS attack may geneate TCP flows, e congestion mechanism does not wok. The attack taffic shows absolutely non-esponsive behavios. has geat chance to detect and estain it. We can also adaptively adust paamete of α so at can fit well wi e diffeent degee of non-esponsive behavios. It is expected at α is evised accoding to e diffeence between and. ep This pape has shown how adaptive active queue management can be used in conunction wi RED gateway to effectively enhance esponsive behavios. Peliminay simulation esults wi a popula envionment have shown e efficacy of e algoim. If e inteactions among diffeent flows ae fully undestood and contol paametes ae well optimized, moe pefomance impovement can be expected in an extensive eseach. Since e netwoks ae becoming moe heteogeneous, is scheme should adapt to moe extensive taffic. How to impove e scalability is ou futue wok. 6. REFERENCES [] W. Feng, K. G. Shin, D. Kandlu, and D. Saha, Stochastic Fai Blue: A Queue Management Algoim fo Enfocing Fainess, Poc. IEEE INFOCOM, Anchoage, AK,, 5-59. [] T. Bonald, M. May and J. Bolot, Analytic Evaluation of RED Pefomance, Poc. IEEE INFOCOM, Tel Aviv, Isael,, 45-44. [3] R. Pan, B. Pabhaka, K. Psounis, CHOKE: A Stateless Active Queue Management Scheme fo Appoximating Fai Bandwid Allocation, Poc. IEEE INFOCOM, Tel Aviv, Isael,, 94-95. [4] S. Floyd, and V. Jacobson, Random Ealy Detection SYSTEMICS, CYBERNETICS AND INFORMATICS VOLUME - NUMBER 6 45