Computer Communications 28 (2005) 111 www.elsevier.

com/locate/comcom

A study of robust active queue management schemes for correlated trafc

Sai S. Orugantia, Michael Devetsikiotisb,*,1
a

Operations Research, North Carolina State University, 207 Daniels, Raleigh, NC 27695-7913, USA b ECE Department, North Carolina State University, Raleigh, NC 27695-7911, USA Received 4 May 2003; revised 8 May 2004; accepted 18 May 2004 Available online 25 June 2004

Abstract Random Early Detection detects the presence of congestion by measuring the exponentially weighted moving average (EWMA) of the queue length against two thresholds. Recently, rate-based packet marking schemes, like Adaptive Virtual Queue, have shown that they are more exible in detecting and dealing with congestion as compared to schemes based on EWMA of queue length. In this paper we shed light on the operation of distributed local Active Queue Management (AQM) schemes by decomposing them based on their logical functions and by studying them as a combination of a measurement module and a control module. We also exploit the presence of strong correlation in network trafc in order to formulate a generalized Predictive AQM (PAQM) scheme. For this, we extend the calculation of the EWMA by including a term that represents the future trafc intensity. We compare the performance of our generalized predictive AQM scheme with a pure rate-based packet marking scheme. We study the performance of PAQM for various combinations of predictors and controllers with respect to robustness, end-to-end delay jitter, goodput and linkutilization. We present the goodput results purely with the intention that increased robustness is not at the expense of decreased goodput. Based on the simulation results, we have the following observations: (i) for proper selection of the weight of future observations, PAQM schemes achieve better delay bounds than the traditional (non-PAQM) ones. (ii) AQM schemes with stochastic based trafc prediction are usually more robust when the control is based on EWMA of queue length. Similarly, AQM schemes with trafc prediction based on nite impulse response lters are more robust when the control is based on xed queue occupancy.(iii) More robust AQM schemes usually have higher goodput for comparable average queueing delay values, (iv) excessive dependence on predicted future arrivals always results in deterioration of performance. This sets a bound on the optimistic performance of AQM schemes even when using prediction. q 2004 Elsevier B.V. All rights reserved.
Keywords: Predictors; Controllers; Correlation; Robustness

1. Introduction With an increasing amount of network trafc running on cooperative transport protocols, it has become evident that the performance of a simple DropTail scheme is not adequate to ensure low buffer occupancy and high goodput for sustained periods of time. Global synchronization is a phenomenon which occurs due to the combination of cooperative transport protocols and DropTail. In order to avoid uctuations in arrival rates the packets arriving at the buffer need to be pro-actively dropped. It is for this reason
* Corresponding author. Tel.: 1-919-515-5253; fax: 1-919-513-5108. E-mail addresses: mdevets@ncsu.edu (M. Devetsikiotis), ssoruganti@ ieee.org (S.S. Oruganti). 1 Tel.: 1-919-515-5253; fax: 1-919-515-5601; URL: http://www4. ncsu.edu/~mdevets 0140-3664/$ - see front matter q 2004 Elsevier B.V. All rights reserved. doi:10.1016/j.comcom.2004.05.016

that active queue management (AQM) schemes in the Internet have attracted lot of attention since the last decade. The motivation behind implementing an AQM scheme is driven by the following goals: low buffer occupancy resulting in small queueing delays, low packet losses and high link utilization. A good AQM scheme would, therefore, make a good choice of packet drops at each stage so that the above mentioned goals are realized. The second desirable feature of a good AQM scheme is its robustness. A robust AQM scheme would require little tuning by a network operator and would remain inherently stable to trafc uctuations. The third feature of a good AQM scheme is its ability to take into account the closed-loop behavior exhibited by the cooperative transport protocols. Random Early Detection (RED) [1] and its several variants [2 9] do not account for both the robustness of AQM schemes and the closed-loop effect of TCP on packet

S.S. Oruganti, M. Devetsikiotis / Computer Communications 28 (2005) 111

marking. In Ref. [10] the authors present an excellent mathematical framework to model the closed-loop of TCP, however, they still do not address the issue of robustness. Among the various proposed AQM schemes there has been a relative scarcity of systematic and statistically complete performance study of the workings of AQM and the effects of its actions on its robustness. Our aim in this paper is threefold. First, we shed more light on the operation of a distributed local AQM scheme. Although, along with its variants, it is distributed throughout the Internet, AQM actions are local in nature. We carry out a systematic study of the working of such an AQM scheme by decomposing it into two logical modules, namely, a measurement module, one that measures the incipient network congestion at every interval, and a control module that makes decisions based on the information provided to it by the measurement module. Secondly, we investigate the effect of the observation that trafc generated by TCP sources is (positively) correlated in nature. This property of strong correlation can be used to predict the future trafc intensity. We also describe ways to exploit its correlation using different measurement modules. Finally, we carry out a detailed analysis of the efciency of AQM schemes for various module combinations by employing robustness and delay jitter as performance metrics. We also record the goodput and link utilization values of these combinations and present them here purely with the intention of asserting that increased robustness of such schemes is not at the expense of degraded goodput. In this part, we especially concentrate on the study of robustness of AQM schemes that exploit the strong correlation that exists in successive measurements of network trafc. It has been seen that to extract good performance from an AQM scheme, say RED, the mechanism has to be properly and carefully tuned [11,12], failing which it tends to achieve worse goodput than a DropTail scheme. It is for these reasons we prefer robustness and delay jitter as our metrics of comparison rather than the traditional metric of goodput. We have also conducted exhaustive statistical analysis of our output by carefully removing the observations made during warm-up periods and presenting the 95% condence intervals for all cases. We strongly believe that our approach towards network analysis will help identify the essence of working of a distributed local AQM scheme and also lay groundwork to build efcient and robust ones. Our motivation for this work is partly inspired from the efforts of Park and Willinger [13]. In this paper we describe two existing predictors and one rate-based controller. We also propose modications to an existing predictor and an existing controller. The new predictor not only captures the long range correlation but also accounts for the short term burstiness of network trafc. The modied controller is designed to achieve new control actions. Ref. [14] contains earlier, partial results of our

work. In this paper we present the detailed results for all predictor-controller combinations. The paper is organized as follows: Section 2 discusses the characteristics and importance of a robust AQM scheme. Section 3 explains the chaotic characteristics of TCP trafc. Section 4 provides our denition of robustness. Section 5 starts with a discussion on the need for heuristic methods of prediction rather than analytical ones. Section 5.1 discusses the two logical modules of an AQM technique. In this section, we refer to some of the existing predictors and controllers and propose new, modied ones. Here we also discuss the implications of using an ideal predictor for which prediction error is zero. Section 6 details the network conguration and the data collection technique. Section 7 collates the results. And nally we conclude Section 8 by summarizing the work done and the inferences we draw from the results.

2. Features of a Robust AQM scheme All AQM schemes constantly adjust their parameters to extract the maximum performance from a router. The amount and frequency of ne tuning generally varies depending on the amount of trafc encountered and the buffer space available. It is usually desired that the routers adjust the parameters themselves without much assistance from a network operator. Robust AQM schemes take into account such uctuations and make necessary adjustments to the parameter values whenever the trafc load or available buffer space changes. With the above arguments in mind and the discussions with Blake and Haraszti [15], we believe that robust AQM schemes should take precedence over strictly optimal ones. Our notion of a robust scheme is that which is capable of switching from one trafc environment to another without sustained loss of performance . Blake and Haraszti suggested a few such trafc environments, which are: small number of high-bandwidth ows large number of low-bandwidth ows heterogeneous round trip times heterogeneous bandwidths due to upstream constraints (broadband vs. modem access rate constraints) heterogeneous ow durations non-stationary ow duration distribution Fig. 1 illustrates the difference in performance between a hypothetical DropTail scheme and an AQM scheme, AQM1, under the same network trafc conditions. It also shows the difference in performance between the same AQM scheme but under different trafc conditions, namely, AQM1 and AQM2. For a given queueing delay, an AQM scheme should yield better goodput than a DropTail scheme.

S.S. Oruganti, M. Devetsikiotis / Computer Communications 28 (2005) 111

the strongly correlated trafc generated by TCP. The trafc values at the present time contain a signicant amount of information about the trafc values in the near future. This property of dependence can be used to predict the future trafc intensity. We believe that by combining more accurate trafc information with rate-based adaptive controllers it is possible to design AQM schemes that are more robust and perform better than the currently implemented ones.

4. Denition and quantication of robustness Robustness signies insensitivity against small deviations in the assumptions [19]. In this paper we have quantied AQM robustness as the ability of a router to maintain its performance even under harsh changes in its environment. Referring to Fig. 1, the robustness of an AQM scheme, under two different network trafc conditions (AQM1 and AQM2) and with an average queueing delay of d1 is g 2 g1 1 Rg 2 g2 With this denition it is clear that more robust schemes have smaller robustness index values. We would also like to make clear that this value gives only a comparison metric and does not give a normalized value of robustness.

Fig. 1. Plot of Goodput vs. average queueing delay for a DropTail scheme and an AQM scheme, under different trafc conditions, AQM1 and AQM2. (Source: authors impression of the scheme).

3. Understanding TCP trafc 3.1. Behavior of TCP trafc It has been observed in Refs. [16,17] that TCP trafc can be interpreted as pseudo self-similar in nature, which means that TCP displays self-similarity or scaling only on short time scales. It is the nature of TCP to saturate the outgoing link of a router and it achieves this saturation by ramping up its packets sending rate until it reaches its maximum window size or experiences congestion, whichever occurs earlier. Similarly, packet losses (or delayed acknowledgements) lead to a decrease in packets sending rate. This window scaling phenomenon reects the adaptive nature of TCP. Also TCP trafc shows a bursty nature. A stream of packets arrive from a source which is then followed by a period of silence. This behavior is quite different from the behavior of a UDP trafc where a steady ow of packets is guaranteed by the UDP source. The third characteristic of TCP trafc is that it is chaotic [18]. Although periodicity is induced due to round trip time and retransmission timeout calculations, in the event of continued packet losses the exponential back-off behavior of TCP changes this periodicity making the packet sending behavior chaotic. Therefore, for predicting the aggregate behavior of TCP trafc at a router, we have to take into account its adaptive, bursty and chaotic nature. 3.2. Motivation for a predictive AQM scheme The discussions in Sections 2 and 3.1 highlight the fact that the next generation AQM schemes should be not only robust in their performance but also should take into account

5. Modules of a predictive AQM scheme The logical diagram of our predictive AQM framework is illustrated in Fig. 2. As mentioned in previous discussions, network trafc, due to TCP connections, was shown to have strong short-range dependence, which means that based on recent observations of network trafc it is possible to make a reasonable judgement of the future trafc levels. However, network trafc shows nonlinear variations and includes probabilistic elements due to the universal deployment of RED in almost all routers. Hence, an accurate TCP-RED like system would need to be modeled using principles of nonlinear stochastic control theory.

Fig. 2. The logical diagram of our predictive AQM framework.

S.S. Oruganti, M. Devetsikiotis / Computer Communications 28 (2005) 111

The analysis of a predictively enhanced version of such a system is complicated and rather intractable in closed form as it involves several sub-systems operating in a closed loop, and receiving correlated (short-range and possibly long-range dependent) random input. In Ref. [20], the authors present a nonlinear and non-stochastic model of the trafc rate. To make it analytically tractable, they linearize the system around the equilibrium point. Since TCP-RED is a difcult system to analyze in closed form, the closed-loop, global optimization of the choices for the predictor becomes equally intractable. Instead, we explore the use of heuristics for the prediction and control components of the AQM system. Such approaches been proved to work effectively in their original applications. Furthermore, we utilize extensive simulations under realistic trafc models in order to evaluate its performance. In this section we present a discussion on the need for using heuristic methods of prediction rather than analytical ones. In Section 5.1 we dissect the working of an AQM scheme into two logical modules: Measurement module, or Predictor Control module, or Controller 5.1. Measurement module or predictor The purpose of predictors is to capture the trafc intensity and estimate the future arrivals over a prediction interval. The accuracy of prediction depends a lot on the length of the sampling interval [21]. For our project the sampling interval was kept at 10 ms. We focus on the four following predictors. 5.1.1. Predictive congestion control (PCC) The detailed operation of this predictor is mentioned in Ref. [13]. In this paper we have used the same notations as mentioned in Ref. [13]. The trafc arrival rates are divided into eight intervals based on its past mean m and standard deviation s values.2 These intervals (or quantized levels) from l 1 to l 8; are, respectively 21; m 2 3s; m 2 3s; m 2 2s; m 2 2s; m 2 s; m 2 s; m; m; m s; m s; m 2s; m 2s; m 3s; m 3s; 1: Based on these intervals an 8 8 conditional probability matrix Mlt1 ; lt2 P{Xt2 lXt1 } is created corresponding to each of the eight current quantized levels mapping into the same eight quantized levels into the future. For every slot in the training period (that consists of the sub-slots lt1 ; lt2 ), the corresponding value of Mlt1 ; lt2 is increased by one.
2 For the purpose of simulation eight divisions were found to be sufciently granular to capture the trafc rates.

Fig. 3. Flow diagram of DTMW predictor. (Source: reproduced from Ref. [22]).

Finally the values of M are normalized across rows to give the conditional probability P{Xt2 lXt1 }: After the training period the probability matrix M should be updated at suitable intervals. For our simulations we updated the matrix at the end of every slot. 5.1.2. Double threshold moving window (DTMW) This predictor is a type of low-pass but nonlinear nite impulse response (FIR) lter. The details of its working and its original application are mentioned in Ref. [22]. It consists of a moving window of size N where each slot in the window records the trafc rate for that slot. DTMW maintains two thresholds, T 1 and T 2: At each measurement interval the arrival trafc is measured and compared with the rst threshold T 1: If the arrival rate equals or exceeds the threshold T 1 then a bit is set in the moving window corresponding to that slot. A control action for the next slot can then be based on the second threshold T 2 and the summation of all bits in the N slots. The schematic diagram of a DTMW predictor is redrawn from Ref. [22] and shown in Fig. 3. 5.1.3. Weighted double threshold moving window (wDTMW) This is a modied version of the DTMW predictor. Instead of assigning equal weightage to the observations over the N slots in a moving window, higher priority is assigned to the more recent observations and lower priority to the older ones. This type of measurement captures the continuous occurrence of high trafc rates and is relatively unbiased towards the window size N : Occurrences of high trafc rates at close proximity to each other is an indication that future trafc will follow similar trends. Similarly, for large window sizes, all observations in the past affect the current packet marking decision. This method of marking may not be optimal especially for bursty sources, which show high correlation to past data. This implies that the window size N should be large enough to capture the correlation of the past data and at the same time small enough to not include data of no importance to the present packet marking decision. In order to address these problems, observations are assigned weights such that these weights decrease as the observations become older.

S.S. Oruganti, M. Devetsikiotis / Computer Communications 28 (2005) 111

The weights of observations are assigned according to a polynomial of binary power. Let n be an integer with values between 1 and N where the oldest slot in the window has value 1 and the latest slot has value N : The weight polynomial is constructed as: P
1 X iN

Finally, we design a rate-based controller that xes the average queue occupancy, as calculated by EWMA scheme, to a certain desired level. The details of each controller are mentioned in the following sections. 5.2.1. Estimated future average queue length This controller is an extension of the EWMA of queue length. Although it has been used both as a predictor and as a controller in the traditional RED, we implement it here strictly as a controller. It makes a packet marking decision based on the past information of the average queue length and the estimated future arrival rates. Its control action is similar to the one in Ref. [1] and is mentioned below 8 > 0 QAvg t n , Minth > > > > > Q t n 2 Minth > < Avg Maxp Minth # QAvg t n Maxth 2 Minth p > > > , Maxth > > > > :1 Maxth , QAvg t n p drop probability for the next interval, n 1 Maxth ; Minth maximum and minimum threshold values for average queue size Maxp maximum drop probability QAvg t n average queue size at the interval, n: EFAQL is calculated as: QAvg t n w1 QAvg t n 2 1 w2 QAvg t n M 1 2 w1 2 w2 QInst t n 3 In Eq. (3) w2 is the weight attached to the future average queue length obtained from the estimate of future arrival rates in the measurement intervals n 1; n M : The future arrival rates are estimated from the predictors. Qavg t n M is calculated recursively as: QAvg t n M 1 2 u1 QAvg t n M 2 1 u1 QInst t n M where u1 0:002 and, c t n M QInst t n M QInst t n M 2 1 Ar 2 Set n M 2 Dr t n M 4

2i21 Xi 2

2N 21 XN 2N 22 XN 21 20 X1 where Xn is bit status of slot n and is dened as: ( 1 if arrival rate at slot n $ T 1 Xn 0 otherwise

The packet marking decision is taken based on P and T 2: 5.1.4. Phase lag (PL) This is a representation of an ideal predictor, where instead of predicting the future trafc rate its value is obtained from a lookahead into the simulation model. It works as follows: for a phase lag predictor with prediction interval length h; trafc generation is started at a time t 2h but is not fed into the simulation system until at time t 0: This way a router can see the future trafc rates that are going to arrive in the next h intervals. Since such a scheme discounts for prediction errors it can be used as benchmark to compare the performance of other predictors. Although this is not realizable in practice, we use it to set bounds on performance based on the exact knowledge of trafc arrivals. The schematic diagram of a Phase Lag predictor is shown in Fig. 4. 5.2. Control module or Controller The information collected by the predictors is used by the control module to adjust the aggressiveness of packet drops in the next interval. We study the following controllers: Estimated Future Average Queue Length Least Mean Square Fixed Queue Occupancy The Least Mean Square Fixed Queue Occupancy (LMSFQO) is an existing rate-based controller and the details of its operation are cited in Ref. [23]. We propose a modication to the EWMA of queue length calculation to achieve a different control action exhibited by Estimated Future Average Queue Length controller (EFAQL).

c t n M is the predicted arrival at time t where Ar n M and QInst t U Max0; QInst t: For DTMW and ct n M was calculated as: wDTMW Ar 8 < T 1 if sum of bits $ T 2 c t n 1 T 1 Ar : otherwise 2 Eq. (5) represents the buffer dynamics at time t n M : A special case of Eq. (3) is when w2 becomes equal to zero. For this value it reduces to the ordinary RED.

Fig. 4. Figure of an idealized prediction scheme enforced by a phase lag predictor. (Source: authors impression of the scheme).

S.S. Oruganti, M. Devetsikiotis / Computer Communications 28 (2005) 111

5.2.2. Least mean square xed queue occupancy This controller works to keep the queue occupancy close to a target fraction of the maximum queue size, Qfixed : Since the queue occupancy is not allowed to exceed the predetermined value, implementing such controllers help in achieving low buffer variations and low end-to-end delay jitter of packets. This controller is represented as: inf J
M X j1

Router service capacity20 Mbps Buffer capacity100 packets (1 packet 1000 Bytes) The OFF period duration and File size distribution are obtained from [24]. OFF period , Pareto(1.21,2) s File size , Pareto(1.21,2)p10,000 Bytes Round Trip Time (RTT) , U(80, 120) ms Sampling time10 ms The simulation is carried out on a TCP stack which models the widely used TCP Reno with delayed ACKs. It is assumed that ACKs are never lost. The trafc being modeled is based on the dominant network trafc which is http trafc [25]. The document sizes and user think times are assumed heavy-tailed. At the start of the simulation all the sources remain in the OFF state which is distributed as mentioned above. After the source switches to the ON state it receives a le size request which is distributed as mentioned above. During the ON state the sources engage in continuous page transfers. After the le transfer is over each source switches back to the OFF state. This cycle continues till the end of the simulation. 6.2. Data collection and analysis The statistics that were collected during the simulation were goodput, average queueing delay and end to end delay jitter. A robustness graph like the one shown in Fig. 1 was created by varying the buffer capacity from 100 packets to 1000 packets and recording the corresponding goodput. For each predictor-controller combination several runs of simulation were conducted in order to remove bias in measurements during the warm-up period and to calculate the 95th percentile condence intervals during the steady-state period.3 Detailed analysis of the detection of the warm-up period by the use of the moving window method and the calculation of 95th percentile condence interval is mentioned in Ref. [26].

Qt n j 2 Qfixed 2 sampling interval is: c Q , Se Qfixed 2 Ar ^ Se Qfixed 2 Ar , Q , Se Qfixed Q . Se Qfixed

The control action for the next 8 > 0 > > > > > c 2 Se 2 Qfixed > < Q Ar p c Ar > > > > > > > : 1 p Q Se c Ar

drop probability for the next interval, k 1 instantaneous queue size at current interval, k service rate at current interval, k estimated arrival rate of next interval, k 1:

The detailed operation of this controller is described in [23].

6. Experimental study 6.1. Network topology For our experiments we used a simple network as shown in Fig. 5. Each source sends its packets to the same destination through a single router. The output link of this gateway acts as a bottleneck. The network has the following specications: Number of sources30 to 100 All links of capacity20 Mbps

7. Results In this section, we present and discuss the results for robustness, delay jitter, goodput and link utilization. Exhaustive results for the various predictor-controller combinations are presented in a tabular fashion, which we admit, gives a clearer idea of their relative merits and tradeoffs. To maintain the brevity and to display the Goodput variation with respect to w2 and Average queueing delay, we have attached only the plots for schemes based on DTMW. Plots for other combinations display similar trends.
Fig. 5. Network topology consisting of 30 independent sources connected to the same destination through a single congested gateway.
3 Similarly several runs were conducted for each value of w2 mentioned in Eq. (3).

S.S. Oruganti, M. Devetsikiotis / Computer Communications 28 (2005) 111

Fig. 6. Plot of Goodput vs. average queueing delay of DTMW predictor with (a) EFAQL and, (b) LMSFQO controller for a xed RTT setup.

Fig. 7. Plot of Goodput vs. average queueing delay of DTMW predictor with (a) EFAQL and, (b) LMSFQO controller for a variable RTT setup.

Also, we refer to EFAQL and LMSFQO as EFAQL and LMSFQO, respectively throughout the rest of this paper. We calculate the robustness of different predictor-controller combinations for all the scenarios (except 4 and 6) mentioned in Section 2. Figs. 6 8 show the plots obtained for DropTail scheme and schemes based on various controller combinations with DTMW for a small number of high-bandwidth ows and a large number of low-bandwidth ows with a xed and variable RTT setup. Small number of high-bandwidth ows were simulated by 30 sources with maximum window size of 128 and large number of low-bandwidth ows were simulated by 100 sources with a maximum window size of 16. The value of

robustness, between two trafc scenarios, for an average queueing delay of 0.2 s4, is presented in Table 1. Since a lower robustness index indicates better performance, the results show that realistic trafc scenarios, under variable RTT sources, are less robust than their ideal counterparts. This is because of a high degree of variability inherent in the real trafc. Table 1 shows that AQM schemes with stochastic based trafc prediction (PCC) are usually more robust when the control is based on EWMA of queue
In some cases the simulation results did not go beyond 0.2 s due to bounds in programming, in those cases robustness index values for 0.1 s was quoted.
4

S.S. Oruganti, M. Devetsikiotis / Computer Communications 28 (2005) 111

Fig. 8. Plot of Goodput vs. average queueing delay of DropTail for (a) xed and, (b) variable RTT setup.

length. Similarly, AQM schemes with trafc prediction based on nite impulse response lters (DTMW) are more robust when the control is based on xed queue occupancy. Along the same lines we observe that for most part the robustness results in Table 1, between xed and variable RTT setup for the same combination of predictor and controller, are strongly and positively correlated with goodput values in Tables 3 and 4 (with the exception of the Phase-lag predictor with an EFAQL controller). Also, for the same RTT setup, robustness is usually higher for AQM schemes with higher goodput. Delay jitter was obtained for parameters displaying the best performance in robustness. The average queue occupancy with EFAQL is between 5 and 15 packets and with LMSFQO is 90 packets. LMSFQO helps in controlling the delay variations more effectively than EFAQL which is evident from the fact that even with high average queue occupancy LMSFQO displays comparable delay jitter values. This can be explained by the fact that the maximum waiting time in the queue in bounded if the controller is LMSFQO, which results in smaller delay variations. Whereas the EFAQL controller allows the instantaneous buffer occupancy to vary widely resulting in degraded performance of delay-jitter.
Table 1 Robustness, 1023, for a xed RTT setup (in parenthesis results for a variable RTT setup) Robustness PCC DTMW mDTMW PL EFAQL 1.16 (83) 5.32 (2.86) 5.32 (9.12) 5.75 (0.388) LMSFQO

The results for different predictor-controller combinations are presented in Table 2. Delay jitter values for RED and DropTail (in units of 1026) are 85.3 ^ 9548.5 1026 and 138.1 ^ 7934.8 1026 s, respectively. The goodput obtained for different predictor-controller combinations, for a xed RTT of 100 ms, is presented in Table 3 and for variable RTTs is presented in Table 4. The rst row lists all predictors and the rst column lists all controllers. Goodput of RED and DropTail, in units of 1000 Bytes, for a xed RTT setup, is 123,483.4 ^ 185 and 121,446.4 ^ 224 packets, respectively. Similar values for RED and DropTail for a variable RTT setup is 85,900.1 ^ 2012 and 82,627.1 ^ 541 packets, respectively. The values in parenthesis show percentage improvement over RED. EFAQL with w2 0 becomes RED. The graphs corresponding to these tables, for the DTMW predictor, are shown in Figs. 9 and 10. As stated earlier, the results of goodput are presented here purely with the intention of asserting our view that the robustness of AQM schemes is not at the expense of degraded goodput. The results of goodput for various predictor controller combinations show that the performance is usually better when the trafc characteristics are exploited. However, the percentage improvement of goodput, over RED, with PAQM schemes is mild in most of the cases.
Table 2 Delay jitter, in units of 1026 s, for various predictorcontroller combinations for a xed RTT setup Delay jitter EFAQL 21.83 ^ 13.37 1026 41.51 ^ 548.08 1026 41.63 ^ 1207.6 1026 53.76 ^ 783.13 1026 LMSFQO 56.4 ^ 286.18 1026 36.9 ^ 1606.27 1026 57.5 ^ 675.04 1026 95.0 ^ 3349.7 1026

18.62 (140) 3.27 (2.56) 3.23 (4.1) 6.67 (30.3)

The robustness of a DropTail scheme is 4 1023 (8 1023).

PCC DTMW wDTMW PL

S.S. Oruganti, M. Devetsikiotis / Computer Communications 28 (2005) 111 Table 3 Comparison of goodput, in units of 1000 Bytes, for a xed RTT setup Goodput PCC DTMW wDTMW PL EFAQL 2543.5 ^ 257 (0.05) 2976.5 ^ 67.96 (0.4) 2884.2 ^ 90 (0.32) 2218.35 ^ 213 (2 0.21) LMSFQO 1038.8 ^ 132.88 (-1.17) 2919.5 ^ 65.89 (0.35) 2811.5 ^ 121 (0.27) 9044.5 ^ 217 (5.31)

The rst term in each cell shows packets in excess of 121,000, that is, for PCC-EFAQL combination the throughput is (121,00 2543.5) 123,543.5 packets.

mentioned in Ref. [23]. Although we have been conservative in using a relatively small buffer size of 100 packets, we know that for larger buffer sizes the results are expected to be better. To conclude our analysis, we present the results of link utilization for various combinations of predictors and controllers. We measure the link utilization values to assess the efciency of robust AQM schemes. Link utilization is dened as: total_bytes_transmitted elapsed_time link_utilization link_capacity

Table 4 Comparison of goodput, in units of 1000 Bytes, for a variable RTT setup Goodput PCC DTMW wDTMW PL EFAQL 32 979.8 ^ 2166 (19.8) 16 224.7 ^ 1872 (0.38) 16 800.3 ^ 1690 (1.05) 4101 ^ 229.4 (2 13.74) LMSFQO 7415.8 ^ 1600 (2 9.88) 16 942.2 ^ 2547 (1.21) 16 686.3 ^ 1729 (0.92) 12 936.3 ^ 1683 (2 3.45)

The rst term in each cell shows packets in excess of 70,000, that is, for PCC-EFAQL combination the throughput is (70,000 32,979.8) 102,979.8 packets.

Moreover, Figs. 9 and 10 show that although there is gain in goodput with increasing values of w2 in Eq. (3), the goodput is quite sensitive to the weight of predicted future arrivals. These observations impose a bound on the optimistic performance of an AQM scheme as they expose the limitations in greatly improving the goodput of AQM schemes even if the underlying trafc structure is exploited to predict the future arrivals. We have used the minimum and maximum threshold values and buffer size as

Eq. (7), which gives the rate of data transfer as a percentage of link capacity, gives an indication of the amount of drops that are triggered. Packet size is an important factor in determining link utilization. Small packet sizes (64 Bytes) will produce low utilization values as high processor overhead is associated with examining many packets [27]. The utilization values for RED and Predictive AQM schemes are mentioned in Table 5. These utilization values are for a buffer size of 100 packets. Link utilization for a DropTail scheme is 97.2% (66.1%). The result in parenthesis is utilization value for a variable RTT setup and the value outside parenthesis is for a xed RTT setup. Referring to Table 5 it can be seen that the utilization values for Predictive AQM schemes are almost always greater than either RED or DropTail for any RTT setup. Since the measurement of utilization does not make any assumption about the conguration of network and the end hosts, this performance metric can be used quite reliably in calculating the efciency of an AQM scheme.

Fig. 9. Plot of Goodput vs. w2 of DTMW predictor with (a) EFAQL and, (b) LMSFQO (at 1) controller for a xed RTT setup. The dotted lines represent the 95% condence interval of the solid line.

10

S.S. Oruganti, M. Devetsikiotis / Computer Communications 28 (2005) 111

Fig. 10. Plot of Goodput vs. w2 of DTMW predictor with (a) EFAQL and, (b) LMSFQO (at 1) controller for a variable RTT scheme. The dotted lines represent the 95% condence interval of the solid line.

8. Conclusions Our efforts in this paper can be divided into three parts: In the rst part we have analyzed AQM schemes as a combination of a measurement module and a control module. This method of analysis allows us to observe carefully the effect of each module on the performance of the AQM scheme. For each of the modules we implemented different predictors and controllers and compared the performance of an AQM scheme for various combinations. In our simulations we saw that careful control decisions based on prediction of future packet arrivals always lead to an increase in performance. To summarize our simulation results: With a proper selection of the weight of future observations Predictive AQM schemes achieve better delay bounds than the traditional (non-PAQM) ones. AQM schemes with stochastic based trafc prediction (PCC) are usually more robust when the control is based on EWMA of queue length. Similarly, AQM schemes with trafc prediction based on nite impulse response lters (DTMW) are more robust when the control is based on xed queue occupancy. More robust AQM schemes usually have higher goodput for comparable average queueing delay values. We also believe that Explicit Congestion Notication (ECN) based packet marking, rather than packet dropping, would help boost the goodput performance of an AQM scheme. Excessive dependence on predicted future arrivals always results in deterioration of performance. This sets bound on the optimistic performance of AQM schemes even using prediction.

It is inappropriate to evaluate an AQM scheme based only on the goodput performance, which is highly sensitive to the duration of ows (long-lived ftp versus short-lived http). Robustness, on the other hand, is relatively independent of the trafc ow structure and so can be used reliably for AQM performance evaluation. Delay jitter is another important yardstick of performance. Large delay jitter values imply large queue length variations which renders the AQM scheme less robust to accommodate delay sensitive trafc like telnet. Goodput and link utilization were used strictly as additional measures of performance. Predictive AQM schemes, for most cases, display higher link utilization values than the traditional ones. It is known that the precision of prediction degrades as the prediction interval increases. For the same prediction interval, Figs. 9 and 10 show that even though the prediction of future packet arrivals helps in achieving better performance, it performs worse than RED when the current packet marking decisions are too heavily relied on future prediction. In the second part we studied the effects of the fact that the trafc generated by TCP sources is correlated in nature. Dependence of the future trafc measurements on
Table 5 Link utilization values, in %, for a xed RTT setup (in parenthesis results for a variable RTT setup) % Utilization PCC DTMW wDTMW PL EFAQL 98.8 (82.7) 99.1 (69.0) 99.1 (69.4) 98.4 (59.3) LMSFQO 95.4 (62.0) 99.1 (70.0) 99.0 (69.3) 98.5 (62.0) RED (no prediction) 98.7 (81.7) 99.1 (67.1) 99.1 (68.1) 98.3 (58.0)

S.S. Oruganti, M. Devetsikiotis / Computer Communications 28 (2005) 111

11

the current value is a result of (positively) correlated trafc and we have used this property of strong and even long-range dependence to predict the future trafc intensity. We have also suggested ways to exploit the strong (if not long-range) dependence using different predictors. In the nal part we have emphasized the relevance of using robustness and delay jitter as reliable performance metrics of an AQM scheme rather than goodput. In this regard, we have dened our idea of robustness. We have concentrated our attention on the performance of robust AQM schemes that exploited the strong correlation that exists in the network trafc. We have shown that, with our denition of robustness, a positive correlation exists between robustness, goodput and link utilization. We have considered delay jitter as another important yardstick of performance. Large delay jitter values imply large queue length variations which renders the AQM scheme not so robust to accommodate delay sensitive trafc like telnet. We have used goodput and link utilization strictly as additional measures of performance. We have used link utilization to evaluate the AQM scheme efciency as it remains independent of the network conguration and hence a reliable performance metric. We have measured the link utilization values of predictive schemes over non-predictive ones. Our results indicate that predictive schemes usually have higher link utilization values than the non-predictive schemes. We strongly support the use of robustness and delay jitter as performance metrics as they tend to provide better evaluation of an AQM scheme. Extensive simulations over varied trafc environments in this paper have supported our propositions.

Acknowledgements This research is supported by the Center for Advanced Computing and CommunicationNorth Carolina State University, as a Core project. The authors thank Steve Blake and Zsolt Haraszti from Ericsson IPI, Raleigh, for their comments and assistance. Finally the authors acknowledge the assistance of Chih-Chieh Hsu in the preparation of the illustrations.

References
[1] S. Floyd, V. Jacobson, Random early detection gateways for congestion avoidance, IEEE/ACM Transactions on Networking 1 (4) (1993) 397 413. [2] Cisco technical papers: weighted red, congestion avoidance overview, 2000. [3] W. Feng, D.D. Kandlur, D. Saha, K.G. Shin, A self-conguring RED gateway, in: Proceedings of INFOCOM 99, vol. 3, 1999, pp. 13201328.

[4] S. Floyd, R. Gummadi, S. Shenker, Adaptive RED: an algorithm for increasing the robustness of red, 2001. [5] D. Lin, R. Morris, Dynamics of random early detection, in: SIGCOMM 97, Cannes, France, 1997, p. 127137. [6] T.J. Ott, T.V. Lakshman, L.H. Wong, SRED: stabilized RED, in: Proceedings of INFOCOM, vol. 3, 1999, pp. 13461355. [7] R. Pan, B. Prabhakar, K. Psounis, CHOKE, a stateless active queue management scheme for approximating fair bandwidth allocation, in: INFOCOM (2), 2000, pp. 942951. [8] K. Psounis, R. Pan, B. Prabhakar, Approximate fair dropping for variable-length packets, IEEE Micro 21 (1) (2001) 4856. [9] B. Zheng, M. Atiquzzaman, DSRED: improving performance of active queue management over heterogeneous networks, IEEE International Conference on Communications, ICC 8 (2001) 2375 2379. [10] V. Firoiu, M. Borden, A study of active queue management for congestion control, in: INFOCOM (3), 2000, pp. 14351444. [11] M. Christiansen, K. Jaffay, D. Ott, F.D. Smith, Tuning RED for web trafc, in: SIGCOMM, 2000, pp. 139150. URL citeseer.nj.nec.com/ christiansen00tuning.html [12] M. May, J. Bolot, C. Diot, B. Lyles, Reasons not to deploy RED, in: Proceedings of Seventh International Workshop on Quality of Service (IWQoS99), London, 1999, pp. 260262. URL citeseer.nj.nec.com/ may99reasons.html [13] K. Park, W. Willinger, Congestion Control for Self-similar Network Trafc, in: Self-Similar Network Trafc and Performance Evaluation, 2000, pp. 446480. [14] S.S. Oruganti, M. Devetsikiotis, Analyzing robust AQM schemes: a comparative study of predictors and controllers, in: Proceedings in IEEE International Conference on Communications, ICC, Anchorage, Alaska, http://www4.ncsu.edu/~ssorugan [15] S.L. Blake, Z. Haraszti, Personal Communication, in: Ericsson IPI, Centennial Campus, NC State University, 2002. [16] D. Figueiredo, B. Liu, V. Misra, D. Towsley, The autocorrelation structure of TCP trafc, Computer networks: The International Journal of Computer and Telecommunications networking, 40(3) 339361, 22 October 2002. [17] S. Manthorpe, I. Norris, J. Boudec, The second-order characteristics of TCP, October 1996. [18] A. Veres, M. Boda, The Chaotic Nature of TCP Congestion Control, in: INFOCOM (3), 2000, pp. 1715 1723. [19] P.J. Huber, Robust Statistical Procedures, Regional Conference Series in Applied Mathematics, second ed., vol. 27, Society for Industrial and Applied Mathematics, Philadelphia, 1996. [20] C.V. Hollot, V. Misra, D.F. Towsley, W. Gong, A control theoretic analysis of RED, in: INFOCOM, 2001, pp. 15101519. URL citeseer. nj.nec.com/hollot01control.html [21] A. Sang, S. Li, A predictability analysis of network trafc, in: INFOCOM (1), 2000, pp. 342351. [22] C. Huang, Long range dependent trafc: modeling, simulation and congestion control, in: PhD Thesis: Carleton University, 1997. [23] Y. Gao, G. He, J.C. Hou, On exploiting trafc predictability in active queue management, IEEE/INFOCOM 21 (1) June 2002, 16301639. [24] M. Crovella, M. Taqqu, Heavy-tailed probability distributions in the World Wide Web, in: A Practical Guide To Heavy Tails, Chapman & Hall, London, 1998, Chapter 1. [25] S. Deng, Empirical model of WWW document arrivals at access link, IEEE International Conference on Communications 3 (1996) 71 86. [26] A.M. Law, W.D. Kelton, Modeling and Analysis, third ed., McGraw Hill, New York, 2000. [27] Cisco white papers: performance measurements of advanced queuing techniques in the Cisco IOS, 2000, http://www.cisco.com/warp/ public/614/15.html.1