: An Active Queue Management Agorithm for a Sef Managed Internet Bartek Wydrowski and Moshe Zukerman ARC Specia Research Centre for Utra-Broadband Information Networks, EEE Department, The University of Mebourne, Parkvie, Vic. 300, Austraia Abstract - In this paper we introduce a new active queue management (AQM) agorithm caed. provides high ink utiization whist maintaining ow deay and packet oss. enabes ow atency interactive appications such as teephony and network games. is shown to outperform the current AQM agorithms. Certain performance probems with current AQMs are discussed.. INTRODUCTION The Internet is essentiay a network of interconnected queues. The two most fundamenta experiences of a packet whist traversing this network are deay and oss. The Internet is facing increasing packet oss rates and queuing deays. Lost packets waste resources, and may resut in congestion coapse [2], whereas queuing causes packet deay which reduces the quaity of interactive appications. Active queue management (AQM) agorithms were introduced to aeviate the probems of network congestion. In genera, AQM schemes contro congestion by controing fow. Congestion is measured and a contro action is taken. There are two approaches for measuring congestion: () queue based, and (2) fow based. In queue based AQMs congestion is observed by queue size. The drawback of this is that a backog of packets is inherenty necessitated by the contro mechanism, as congestion is observed when the queue is aready positive. This creates unnecessary deay and jitter. Fow based AQMs, on the other hand, determine congestion and take action based on the packet arriva rate. For such schemes, backog, and a its adverse impications, is not necessary for the contro mechanism. The tota Network deay is essentiay the sum of queuing deay (in routers and switches) and propagation deay. Currenty queuing deay dominates most round trip times (RTT). The goa shoud be to reduce the network deay to just the propagation deay. Currenty packet oss is both a signa of congestion and resut of overfowing queues. The Internet engineering task force (IETF) are introducing expicit congestion notification (ECN) to feedback congestion by packet marking instead. (We wi refer to congestion notification to mean either packet dropping or ECN). With ECN, AQM queues can operate with minima packet oss. The next generation of network paradigm wi have AQMs that maintain high ink utiisation whist achieving the QoS requirements of imited packet oss and deay. In this paper, we introduce a new AQM agorithm caed that achieves this objective. The paper is organised as foows: in Chapter 2 we describe current AQMs, pointing out their drawbacks. In Chapter 3 we demonstrate the performance of versus the other AQMs. In Chapters 5 and 6 we discuss the impications of depoying or in on the Internet. 2. CURRENT AQM ALGORITHMS In this section, we describe the currenty proposed AQMs and discuss performance issues. Droptai: Under Droptai when queue overfow occurs, arriving packets are dropped. This eads to high queue sizes and high oss rates at congested inks. Droptai is effectivey no management. Random Eary Drop (RED) [2] [3] [5]: RED is a first generation AQM. The rate of congestion notification is a function of the queue size. As discussed in depth in [5] RED suffers from severe shortcomings. The queue size is not a good indicator of the severity of the congestion, and the eve of congestion notifications issued may be too great and bursty, eading to excessive packet oss. RED is prone to periods of high oss foowed by ink underutiization. BLUE [5]: BLUE is a hybrid fow and queue based congestion contro scheme. It uses packet oss (queue) and ink under-utiization (fow) events to adjust the rate of congestion notification. The congestion notification rate p m is increased at a set rate if the queue size exceeds a threshod L, and it is decreased if the ink is ide. The amount of increase is d, and decrease d2, at a rate of /freezetime. We uncovered a subte drawback with BLUE that may imit its practicabiity. BLUE behaves we if operated within a region of RTT and number of connections N for which the parameters d and d2 were set. However, changes in the dominant RTT of the connections going through the queue, or a significant change in N can invaidate the parameter settings and ead to queue backog osciation between oss and under-utiization. The amount of change of notification rate p m during a queue fu or ink ide event, is the product of the time spent in this event mutipied by the rate of change d(,2)/freezetime. This time is reated to the deay in the response of the TCP sources to the changed notification rate (2 x RTT). The greater the RTT, the greater wi be the p m adjustment. If the RTT increases, so does the change in p m and this may resut in backog osciation. This was observed in simuations using the recommended parameter settings of [5]. Another cause of instabiity is a arge change in the number of connections. It is not our intention to expore BLUE in depth here, but this instabiity is the resut of the adjustment of congestion notification rate p m by a constant d or d2, despite the non inear reation of p m and N. Reca
that based on the TCP Friendy Equation [2] the function of p m versus N requires arger changes of p m for arger N. Random Exponentia Marking () [6]: is a framework for communicating congestion information from inks to sources by exponentia marking. A compete impementation consists of marking at the ink and decoding at the source, however operation on the current Internet with ony the ink agorithm and TCP has been suggested [6]. Three different aternative pricing agorithms PC-PC3 constitute. PC3 is evauated here Eq (), because of its superiority over PC and PC2 [6]: + PC3: p t + ) = [ p ( t) + γ ( α b ( t) + x ( t) c )] () ( where x is the aggregate arriva rate, c ink capacity, b backog, and α and γ are pricing constants. An exponentia function determines the marking probabiity from the ink price: p ( t) m = φ (2) where φ contros marking. Athough works very we in a steady state situation, the behaviour in transient conditions, and with reaisticay constrained buffer sizes is not necessariy optima. The experimenta resuts detai how in an environment with a wide variation in N and finite buffers the performance suffers. 3. The agorithm is a feedback contro function which adjusts the rate of congestion notification in response to the fow based congestion measure, x est, the estimated data arriva rate. is based on a threshod function. If the ink s estimated data arriva rate x est is above the target ink capacity c, the rate of congestion notification, P, is incremented by P at a rate of / T. Conversey, if x est is beow c, P is decremented by P at a rate of / T. The agorithm appies probabiistic marking of incoming packets at the rate P, either by dropping packets, or setting the ECN. Let the step function U(x) be defined by: + x 0 U ( x) = (3) x < 0. Therefore, P = P + P U x est c ). (4) ( t The target ink capacity, c t is assigned a vaue just beow the actua ink capacity c, typicay 0.97 c, so that the queue size converges to 0. Incoming data rate estimation is performed using exponentia averaging: x est = (-exp(-de/k))*(b/de)+exp(-de/k)* x est (5) where De is the inter-packet deay, B the packet size and K the time constant. Other arriva rate estimation techniques coud aso be used successfuy. In [4] Eq (5) is shown to converge to the actua data rate under a wide set of conditions. There is a reationship between PC3 and. If Eq (2) is inearised, m = P, the exponentia marking is eiminated. Furthermore if the buffer term α = 0, and the inear constant γ is repaced with the step function (3), s congestion notification rate P becomes equivaent to s Price p. 4. PERFORMANCE EVALUATION A number of simuations were performed to compare the performance of,, BLUE, RED and Drop- Tai using the NS-2 simuator [8]. The network topoogy (Fig ) and parameters common to a trias are described here. The nodes A-J are the sources and Node 2 is the sink. The AQM agorithm is paced in Node, which represents the ingress of an edge router and governs the 2 ink. Each tria asted 20 seconds with measurements being taken after a 20 second warm up period. Mutipe TCP connections can originate from each source. The active TCP connections were eveny distributed among the A-J sources. Connections were started at a uniformy distributed time over the test period with a uniformy distributed duration between 3 and 23 seconds. The packet size was 000 bytes. Figure. Simuation Scenario B C D E F G H I 4ms 00Mb 9ms 00Mb ms 00Mb 9ms 00Mb 37ms 00Mb 73ms 00Mb 0ms 00Mb 20ms 00Mb A J 3ms 00Mb 20ms 0Mb 8ms 00Mb The configuration parameters for each of the AQMs foowed recommended settings: RED : min th =20% and max th = 80% of the queue size. BLUE: Threshod L is 66% of the queue size. d=0.00, d2=0.0002, freezetime=0ms. 2
Packet Loss Probabiity Deay (ms) 0.4 0.2 0. 0.08 0.06 0.04 0.02 20 8 6 4 2 0 8 6 4 2 Fig a: Packet Loss Probabiity vs Number of TCP Sessions 0 0 50 00 50 200 Average Aggregate Number of TCP sessions DropTai RED BLUE Fig 2a: Mean Queueing Deay vs Max Buffer Size 0 50 00 50 200 250 Max Buffer Size (ms Deay) DropTai RED BLUE : (Parameters adapted from Interworking with Reno [6]) γ=0.0 φ=.003 α=0. updatetime=0.0 bo=55. (Pubic domain rem-mark.cc, was used). : T=0ms, P=0.00,C =0.97% of Capacity, K=0.. A agorithms were configured to use ECN. Tria : The AQMs were tested under different traffic oads, by varying the number of TCP sessions. The tota number of TCP sessions activated in a 20 s simuation period was varied from 00 to 300 ( to 4 concurrent sessions on average), in increments of 00 sessions. The buffer size was 9 packets (72.8ms), making some of the connections bandwidth deay greater than this. Tabe summarizes the utiisation, oss and deay of the 3 simuation periods (00 to 300 sessions) from each AQM. RED and Droptai suffered unacceptabe oss and deay across the whoe range of resuts. BLUE performed we in packet oss and deay, at the price of utiisation. To match the utiisation of the other agorithms, the buffer size woud have to be increased significanty resuting in arger average backog and oss. sighty outperformed. The traffic in an edge router may vary from tens to hundreds of connections, which represents a vast change in the aggregation of the traffic. An AQM agorithm must cope with this range of traffic conditions without reconfiguration. performs better across a wide range of sessions for reasons detaied in Trai 3. Tabe. Tria Resut Summary Droptai RED BLUE Uti % 0.98 0.98 0.95 0.96 0.97 Loss % 6.85 5.43 0.4 0.23 0.05 Deay(ms) 52.45 52.4 3.95 2.68 8.79 Tria 2: The buffer size was varied to determine the performance of AQMs under the reaistic constrained of finite buffer, which is cose to the bandwidth deay product of the network. 000 TCP sessions over 20 seconds were initiated for each queue size. Utiisation, packet oss and deay were measured. For brevity, ony the deay performance is shown Fig Ta. This tria shows ceary how queue based AQMs maintain a higher average backog. Figure Ta demonstrates that the backog for RED, Droptai and BLUE is reated to the buffer size. The Fow based AQMs, and can maintain the same utiisation with ow backog even with high buffer sizes. This means it is possibe to have a arge buffer, to reduce packet oss, without the drawback of high packet deay common to queue based fow contro agorithms. Indeed with and, it is possibe to seect a ow target utiisation, and operate amost entirey in the zero queue region, eiminating queuing deay. A agorithms except BLUE maintained high utiisation as BLUE sacrificed utiisation for ower backog and oss. There is a singe sweet-spot for BLUE at around 00 ms queue size, where greater maximum buffer sizes increase the queuing deay, and a ower buffer size decreases the utiisation. In contrast, has consistent performance across a variety of buffer sizes. By operating with a sufficient buffer size, is in a very robust region where queuing deay and packet oss probabiity are ow. This makes much easier to depoy, and more ikey to perform better with unpredictabe traffic patterns. has a very simiar performance to in this tria.
Tria 3: Tria 3 expores further the behaviour of and in an edge router. The deay of the botteneck ink -2 was varied from 0 to 80 ms in this tria. 400 TCP sources were started during the 20 second simuation period (43 concurrent fows average). A buffer size of 9 packets was used. Any AQM depoyed must be robust to variations in RTT, which resut from the wide use of RED and drop-tai. For an AQM to be robust, there must be a gracefu degradation in queuing performance if the actua bandwidth deay product becomes greater than the buffer size. Figures 3a and 3b show how the AQMs performance degrades when the botteneck ink s deay is increased, making the bandwidth deay product greater than the queue size (around 30 ms ink -2 deay). s oss rate increased significanty beyond this threshod, whereas s performance degraded more gracefuy. The difference in performance between and originates in s inear price adjustment (x-c) in Eq (). exhibits the probem that when the traffic drops, the price P is decreased drasticay because of (x-c). A drop in arriva rate does not aways represent a drop in demand. PC3 misinterprets the instantaneous fow decrease to a decrease in demand and decreases its price in two situations: ) Finite queues mean there is a possibiity of overfow. Upon overfow, the fow rate decreases drasticay. misinterprets this as a decrease in demand and drops the price heaviy. By dropping the ink price, invites another overfow. remains cam upon overfow by ignoring the magnitude of x-c. 2) With ony tens of fows active, the traffic is bursty and refects TCPs ramp behaviour. Smooth Gaussian packet arrivas begin to occur with ony hundreds of fows. The parameters of the AQM must be vaid for this range of traffic without reconfiguration, as the edge router ives in such an environment. As is more aggressive in price adjustment, the bursty packet arrivas with tens of sources can be misinterpreted as a drop in demand. maintains a more steady dropping probabiity in face of ow number of sources which resuts in a ower packet oss probabiity as seen in Tria 2. Decreasing γ in can achieve this for one set of operating conditions, however at the cost of performance at higher fow eves. wi not aways outperform. There is a sacrifice in the convergence rate by not having s x-c term from Eq, however this is margina. By definition, the arriva rate must be matched to the ink capacity for a system to maintain high utiisation. With finite buffers, this variation in arriva rate must be tighty bounded, otherwise overfow wi occur. So ony sma deviations in x away from c occur in a stabe system with ow oss and high utiisation, and most of the contro action is done in this region. Large differences in x-c are more ikey due to buffer overfow or TCP ramp behaviour in the ow aggregation case. By using a step function as a marking probabiity adjustment, ignores arge differences which are more ikey discontinuous conditions, such as buffer overfow, than changes in bandwidth demand. This makes robust in a wide set of reaistic constrained conditions. Utiisation Packet Loss Probabiity Fig 3a: Packet Loss Probabiity vs 0.04 RTT 0.035 0.03 0.025 0.02 0.05 0.0 0.005 0.95 0.9 0.85 0.8 0.75 0.7 0 0 20 40 60 80 00 botteneck ink deay (ms) Fig 3b: Link Utiisation vs RTT 0 20 40 60 80 00 botteneck ink deay (ms) 5. MULTI-SERVICE IMPLICATIONS Diffserv is being proposed as a means for providing QoS to different services. Packets are abeed with a priority and a scheduing decision is made at each node based on this priority. With a simper soution is possibe, where capacity distribution is sef-managed according to the utiity functions of the sources [6]. Capacity aocation is performed on a macro-economic mode; the congestion notification rate can be viewed as the price of sending a byte. If capacity is oversubscribed, the price goes up and sources receive more congestion notifications. Sources that pace a higher vaue on having more capacity, simpy
react ess to the price increase, and do not reduce their rate as much despite the congestion notification rate increase. In the simpest case, et us assume there are two types of appications, ) rea-time (RT) and 2) non-rea-time (NRT). RT appications require ow deay and may not impement congestion contro. NRT appications are insensitive to deay, and impement congestion contro, such as TCP, ie: HTTP, FTP and Tenet. A ink using wi deiver ow deay and oss to RT appications, as s congestion contro forces NRT appications to buffer at the source (sow send rate) in face of increased RT demand. Assuming that the tota capacity is greater than the RT peak demand, s congestion notification contros the responsive NRT arriva rate so that the aggregate arriva rate (NRT+RT) is beow the capacity. Depoying enabes both casses of traffic to coexist, without the extra infrastructure of packet cassification, or cass capacity reservations. 6. DEPLOYMENT WITH NO ECN Currenty the Internet has few ECN enabed TCP sources, so the significant congestion notification mechanism is packet dropping. A side effect of AQM agorithms which reduce the network atency, ( and ), is that a high congestion notification rate is required to contro TCP sessions which become very aggressive with ow RTT. With ow RTT and the resuting high packet dropping rates for non-ecn sources, the effective throughput suffers and TCP experiences significant timeouts. In a network that is not fuy ECN capabe, a guaranteed minimum packet atency is required for non-ecn TCP sources to keep the packet oss beow a maximum toerabe amount max. The key concept is that an artificia packet deay is induced to minimise packet oss. This deay, D, can be the minimum deay required to maintain a packet oss requirement. ECN enabed packets need not suffer this deay. The fu technique is discussed in [9]. 7. CONCLUSION This paper has presented some deficiencies in we known queue management agorithms and provided an aternative agorithm,. was shown to maintain a ow packet oss rate as we as ow average queue size. The agorithm is simpe, robust, ow in computationa compexity, easiy configured, and sefcontained to a singe router, making it easy to depoy. Depoyment of ow deay, ow oss agorithms such as wi improve Internet performance and enabe rea-time appications. ACKNOWLEDGEMENTS This work was financiay supported by Agient Technoogies. REFERENCES [] Ion Stoica, Scott Shenker, Hui Zhang, "Core-Stateess Fair Queueing: A Scaabe Architecture to Approximate Fair Bandwidth Aocations in High Speed Networks", SIGCOMM, 998. [2] Foyd, S. and Fa, K, Router mechanisms to support end-to-end congestion contro., Tech.rep, LBL, 997. (http://wwwnrg.ee.b.gov/nrg-papers.htm.) [3] S. Foyd and V. Jacobson, Random eary detection gateways for congestion avoidance IEEE/ACM Transactions on Networking, vo., Number 4, pp397-43, August 993. [4] I. Stoica, S. Shenker, and H. Zhang. Core-stateess fair queueing: Achieving approximatey fair bandwidth aocations in high speed networks, Technica Report CMU-CS-98-36, Carnegie Meon University. June 998. [5] W. Feng, D. Kandur, D. Saha, K. Shin, "Bue: A New Cass of Active Queue Management Agorithms" U. Michigan CSE-TR-387-99, Apri 999. [6] S. Athuraiya and S.H. Low. "Optimization Fow Contro, II: Random Exponentia Marking", Submitted for pubication, http://www.ee.mu.oz.au/staff/sow/research/, May 2000. [7] W. Feng, D. Kandur, D. Saha, K. Shin, "Stochastic Fair Bue: A Queue Management Agorithm for Enforcing Fairness", in Proc. of INFOCOM 200, Apri 200. [8] The Network Simuator - ns-2 homepage: http://www.isi.edu/nsnam/ns/ [9] B. Wydrowski and M.Zukerman, On the Transition to the Next Generation ow atency TCP/IP Internet, (submitted for pubication, avaiabe http://www.ee.mu.oz.au/pgrad/bpw).