A New Rate-based Active Queue Management: Adaptive Virtual Queue RED

Transcription

1 A New Rate-based Active Queue Management: Adaptive Virtual Queue RED Do J. Byun and John S. Baras, Fellow, IEEE Department of Computer Science, University of Maryland at College Park Abstract In an effort to improve performance of congested gateways, a new Active Queue Management (AQM) algorim, Adaptive Virtual Queue Random Early Detection (AVQRED), was developed by feeding virtual queue size to e RED algorim. The objective of e new algorim is to improve overall performance by keeping link utilization high, link utilization stable, queueing delay low and consecutive packet drop rate low. This paper shows e objective is met by comparing AVQRED wi six oer well known AQM meods in a realistic emulation environment. To provide fair comparisons, e AQM parameters are fine-tuned by exploring many different parameter settings via emulations. The emulation results conclude at AVQRED improves overall performance by 8 to 25%. To provide intuitive properties and validate e emulation results, a maematical model is proposed and fed to MATLAB, and e results from MATLAB and emulations are compared. 1. Introduction Due to exponential increases in internet traffic, congestion has been an unavoidable attribute to many Internet service providers for e last decade and improving performance of congested gateways has been an active research interest. This paper contributes an effort to improve overall performance of congested gateways by proposing a new rate-based AQM algorim, Active Virtual Queue Random Early Detection (AVQRED). Active Queue Management (AQM) is an algorim at detects and reacts to congestion to avoid queue overflows. There are generally two ways to react to congestion: signal congestion to traffic sources explicitly by setting Explicit Congestion Notification bits, or signal congestion to traffic sources implicitly by dropping packets. A new AQM algorim was developed by overlaying a virtual queue on top of e real queue and applying e RED algorim to e virtual queue instead of e real queue. The objective of e new algorim is to improve overall performance by keeping link utilization high, link utilization stable, queueing delay low and consecutive packet drop (mark) rate low. This paper compares AVQRED wi six well-known AQM meods, RED [4], ARED [7], BLUE [8], GKVQ [20], AVQ [18], and PI [17] in a realistic emulation environment. To provide fair comparisons, e parameters for all seven AQM meods are fine-tuned before e comparisons via emulations. The emulation results conclude at AVQRED improves overall performance by 8 to 25%. To provide intuitive properties and validate e emulation results, a maematical model is proposed and fed to MATLAB, and results from MATLAB and emulations are compared. This paper is organized as follows: Section 2 provides an overview of e six AQM meods. Section 3 discusses e new AQM algorim, AVQRED. Section 4 provides e emulation framework and meodologies. Section 5 provides e emulation results. Section 6 provides a maematical model. Section 7 validates e emulation results by feeding e model to MATLAB and comparing e MATLAB results wi e emulation results. Finally, section 8 concludes e paper. 2. Overview of oer AQM meods This section provides high-level understanding of each AQM meod which is compared wi AVQRED. Refer to e corresponding reference for more details about e algorim RED The RED [4] algorim computes e marking probability when e weighted queue size falls between min and max parameters. The marking probability becomes higher as e weighted queue size gets closer to max (becomes 1 if it s greater an max ), and it also becomes higher as e distance between each marking gets bigger. Parameter tuning is required for w q and max p. w q controls e weighted average queue size which en determines how quickly e algorim reacts to congestion. Reacting too quickly may result in oscillations in queue size, and reacting too slowly may result in queue overflows. max p is a scaling factor for e 389

2 marking probability which also controls how quickly e algorim reacts to congestion ARED The intuition behind e ARED [7] algorim is to selfadjust max p according to e target average queue size every half second. Emulation results revealed at e adjustment of max p does not find e value at best represents e desired packet loss rate of e gateway. i.e. max p reduction is so aggressive at it always stays at a lower value an e desired packet loss rate resulting in e average queue size to exceed max. Therefore, e ARED algorim was slightly modified to have a hysteresis requirement where decreasing max p by β requires e queue size to stay under e target queue size for n times PI The main idea behind e Proportional Integrator (PI) [17] algorim is at e queue leng slope determines e packet marking probability, and e queue is regulated by e desired queue leng. i.e. p[k+1] = p[k] + a(q[k+1] qref) b(q[k] qref). The algorim computes a new p every T seconds. 3. AVQRED algorim The motivations behind AVQRED are to fix e problems of AVQ: global synchronization (consecutive packet drops) and inaccurate virtual capacity which leads to high queueing delays BLUE Unlike RED and ARED, BLUE [8] relies on packet loss and idle events of e queue when adjusting its marking probability. It increases or decreases by a configured value, d1 or d2 respectively, when e queue size exceeds a reshold or it becomes empty. It does so no more an once every configured interval, freeze_time GKVQ and AVQ Gibbens-Kelly Virtual Queue (GKVQ) [20] maintains a virtual queue whose service rate is e desired link utilization. When an incoming packet exceeds e virtual queue limit, it drops or marks e packet. Adaptive Virtual Queue (AVQ) [18] maintains e same virtual queue whose capacity is dynamically adjusted. The virtual capacity is adjusted by adding e number of bytes at could have been serviced between e last and e current packets minus e bytes at were just received. The problems wi ese two schemes are global synchronization and queue overflows when virtual capacity exceeds e actual processing capacity. Alough AVQ seems to solve e latter by adjusting e virtual capacity, ere is an assumption at e actual processing capacity is static which could lead to high queueing delays and queue overflows if e processing allocation shifts a common phenomenon in real networks. However, as long as e virtual capacity does not exceed e actual processing capacity, GKVQ and AVQ yield low queueing delays and stable link utilizations. Figure 1. AVQRED flow diagram The AVQRED algorim constructs a virtual queue and feeds e virtual queue sizes to e RED algorim instead of feeding e weighted average queue sizes to it. AVQRED reshapes e incoming traffic according to e desired link utilization because e RED algorim reacts to e congestion level of e virtual queue which is serviced by e desired link utilization. The AVQRED Algorim highlights e AVQRED parameters in bold. Note at w q and max p are no longer in e algorim because eir functionalities are replaced by e desired link utilization in AVQRED. alpha is a low-pass filter for e actual capacity calculation and choosing a small number such as 0.05 is recommended since changes in processing capacity are not drastic. min_capacity and max_capacity define e range of processing capacity. Choosing a big value for max_capacity is recommended if e maximum processing capacity of e gateway is not known and one wishes to maximize e capacity. Choosing a value too low for min_capacity will cause a delay in finding e actual capacity at e beginning but it will stay wiin e right range once it is found. Note at min and max no longer control e real queue leng directly. They define e RED region of e virtual queue which indirectly controls e real queue leng. One big advantage of applying RED to e virtual queue is at e algorim works well even wi a small queue (because 390

3 e RED region for e virtual queue can still be big) whereas having a small queue makes RED resemble a tail-drop queue. transmits e packets. The seven AQM meods were implemented at e receive queue and drop was chosen for e marking meod. AQM was implemented only on one queue which is assumed to be e only congestion queue in our network scenarios. To eliminate e unnecessary system complexity, satellite specific features such as TCP spoofing, per-user queueing and output rate limiting were removed. Alough e software serves as a gateway in satellite networks, it was modified to be a generic IP Gateway. The emulation environment was constructed using two IP Gateways (one serves as a router to e internet and e oer serves as modems for N different users) and a traffic generator called Spirent. Figure 2. Emulation flow diagram Figure 2 illustrates e connectivity of e emulation setup. To best emulate latencies, a delay simulator was inserted between e two gateways. The round trip time (RTT) between e client IP Gateway and e Spirent is 4 msec, e RTT between e client IP Gateway and e AQM IP Gateway is 20~40 msec, and e RTT between e AQM IP Gateway and e Spirent is 40~80 msec resulting in an end to end RTT of 104~204 msec HTTP connections were generated between 200 clients and 60 servers wi e following attributes. AVQRED Algorim 4. Emulation framework The actual gateway code from Hughes Network Systems was used to evaluate e AQM meods. The software was simplified to have only one queue for receiving packets and one read at dequeues and 1. At e startup, ere are 200 new HTTP connections every 5 seconds wi 5 second sleep time between each ramp up until 4000 HTTP connections are established. 2. When a connection is closed, a new connection is created to fill e gap to maintain 4000 HTTP connections. 3. Each web page contains 250Kbytes ~ 550Kbytes of data. 4. The maximum download speed of each TCP connection is 5Mbps. 5. The average bir and dea rate of e connections is about 200 connections per second (approximately 5% of e total population). 391

4 4.1. Evaluation meodologies Table 1. Final parameter settings The following performance metrics were used: 1. Link utilization Mean and standard deviation of link utilizations are compared. 2. Queueing delay Mean and standard deviation of queueing delays are compared. 3. Packet loss Mean and standard deviation of packet losses are compared Parameter settings The processing capacity ranges between 13 to 25 1 Mbps. Throughout e experiments, each AQM meod has e same target queue size or queueing delay, 22 2 packets or 10 msec respectively. In oer words, e queue size based AQM meods will use e target queue size of 22 packets, and e rate based AQM meods will use e target queueing delay of 10 msec. For each AQM meod, results from ree 15 minute emulation runs for 10 to 27 different parameter settings were used to finetune e parameters. Details on how e tuning was done are described in e next subsection, The parameter permutations and e results are not shown due to e space limitation. 5. Emulation results Table 2. Final results (mean of ree iterations) Parameter selection rule. Parameter settings are ranked for each performance metric. Based on e rank, it is given a weight, x = 2 rank. Each setting is en given a weight which is e sum of e weights (x s) of all its metrics. The settings wi e ree lowest weights are chosen assuming each metric is equally important. Only two settings are chosen for ARED due to its adaptive adjustments for max p Final parameter settings. Table 1 shows e parameter settings at produced e best emulation results which were used roughout e final emulations. 1 This can be a hypoetical number at may or may not be achievable by e gateway due to dynamics of e gateway resource allocation is derived from e desired queueing delay of 10 msec wi 25 Mbps maximum processing capacity and 1400 bytes average packet size. Each setting in Table 1 was run for 15 minutes, ree times. The ree results are averaged and eir standard deviations were also calculated. Three iterations of emulations were averaged to produce Table 2. Standard deviations (not shown due to e space limitation) of e ree iterations were relatively small, us validating e use of e mean values. Using e results in Table 2, e following were done: 1. All AQM meods in Table 2 were ranked for each 392

5 performance metric and e same weighted ranking rule discussed in e previous section was applied to produce e rank weights, column G in Table 2. This measurement indicates at e AVQRED ranked first and second. 2. For each AQM meod, e parameter setting at scored e best rank weight was picked and e results are compared wi AVQRED in Figure 3, Figure 4, Figure 5, Table 3, Table 4 and Table 5. Figure 4 and Table 4 show at AVQRED s queueing delays are at least 25% lower an e oer AQM meods while delay jitters are about 5 to 15% higher an RED and ARED. However, it is also important to note at having low delay jitters is not always good because it means at short-lived and bursty traffic could be dropped. Figure 3. Link utilization Table 3. Link utilization improvement by AVQRED Figure 3 and Table 3 show at AVQRED s link utilization degrades by less an 5%, and 9% or more link utilization stability is gained. GKVQ and AVQ are exceptional cases because eir link utilizations are well below e actual gateway capacity which resulted in lower variations an AVQRED. Figure 5. Packet loss Table 5. Packet loss improvement by AVQRED Figure 5 and Table 5 show at AVQRED s packet loss rate is up to 15% lower and e standard deviation is 28 to 50% lower an e oer AQM meods except for BLUE 3. It is important to note at a low drop standard deviation implies at e distance between drops is more uniform resulting in less consecutive drops. About 30 second time period was magnified in Figure 5 to better visualize e packet loss behavior. Figure 4. Queueing delay Table 4. Queueing delay improvement AVQRED 6. Model and analysis This section discusses e maematical model used for validating e emulation results. Alough details on how 3 Higher queueing delays and higher queue dep resulted in lower packet drop rates and variations. 393

6 each equation is derived are not discussed here due to e space limitation, ey can be found in [1], [2] and [5]. For simplicity, RED, ARED and AVQRED (which produced e best emulation results) were modeled wi slight modifications to e models from [1], [2] and [5]. Note at e actual queue size ODE was modeled wi M/M/1 assumption and e AVQRED queue size ODE was modeled wi M/D/1 assumption to expose e main characteristic of e AVQRED algorim which is deterministic service time. Our model does not capture some of e TCP dynamics such as TCP slow start, timeouts, variable packet size and variable propagation delay which caused a slight discrepancy between e emulation results and e MATLAB results. Even wi such discrepancy, e MATLAB results are consistent wi e emulation results wi respect to e important characteristics (utilization, stability, unbiased drops against transient traffic, global synchronization level and queueing delay) of e AQM meods. The following ODEs are used to model RED, ARED and AVQRED, and e details on how ey are derived are in [1], [2] and [5]. dx ln( 1 α ) ln( 1 α ) = x ( t ) q ( t ) δ δ dq q ( t ) = λ ( t )( 1 p ( t )) u 1 + q ( t ) 2 d λ m λ = (1 p ( t )) p ( t ) 2 R 2 m ds = τ u + λ ( 1 p ) dl q ( t ) = u ( ) l ( t 1) 1 + q ( t ) Variables: dx = Weighted average queue size. dx from [5]. dq = Actual Q size. Equation (6) from [1]. dλ = Arrival rate. Equation (3) from [2]. ds = AVQRED queue size from e Lindley equation. dl = Link utilization. Parameters: α = Queue weight for weighted average δ = Delta time between each packet arrival λ = Aggregate arrival rate p = AQM packet loss probability u = Service rate wi uniform distribution between [80%, 100%] m = Number of concurrent TCP connections R = Round trip delay τ = Target link utilization % Eier q (when RED or ARED is used) or s (when AVQRED is used) is fed to e following equation from [5] to compute e marking probability: 0, x min p( x) = max min 1, max, p 0 x < min min x max max < x One modification made to e models proposed in [1], [2] and [5] is e variation factor in e service rate. This could be due to many reasons but e most common reason is a single processing resource shared by multiple processes and reads. Emulation results also confirm at e variation is a valid assumption. From e emulation results, it was also confirmed at e servers we used for e emulations have a uniform distribution between 80% and 100% of e maximum service rate 7. Validation Three AQM meods (AVQRED wi setting 3, ARED wi setting 1 and RED wi setting 2) at have e best emulation results are fed to MATLAB using e equations provided in section 6. Then e validation is done by comparing e MATLAB results wi e emulation results. There are 200 TCP connections at e beginning and e traffic volume is increased by 200 connections every 5 seconds until 4000 connections are established. 5% of e population goes in and out of e system every second to resemble e traffic characteristics used for e emulations. The propagation delay is set to 200 msec and e average packet size is set to 1400 bytes. The comparison metrics are queue size, arrival rate and link utilization. Alough we may not see perfectly matching graphs between e MATLAB results and e emulation results due to dynamics of traffic such as TCP slow start, timeouts, variable packet size and variable propagation delay, we are still able to show at e maematical model captures e important characteristics of e ree AQM meods such as utilization, stability, unbiased drops against transient traffic, global synchronization level and queueing delay. 394

7 degradations in TCP window evolution. The MATLAB results for link utilization are also consistent wi e emulation results, Figure 3 and Table 3, because bo show at AVQRED has more stable utilization wi slightly less utilization alough e difference in stability is smaller in e MATLAB results. We believe at is smaller difference in stability is due to e fact at some of e traffic dynamics such as TCP slow start, timeouts, variable packet size and variable propagation delay are not captured by e model. 8. Conclusion Figure 6. Queue size Figure 7. Arrival rate A new rate-based AQM meod, AVQRED, was developed to improve performance of congested gateways. AVQRED essentially combines AVQ and RED and enhances e way virtual capacity is adjusted to adapt to dynamics of gateway resources. AVQRED was compared wi six oer AQM meods, RED, ARED, BLUE, GKVQ, AVQ and PI. Comparisons between e seven AQM meods were carried out using a realistic emulation environment created by a traffic generator, Spirent. For fair comparisons, all AQM parameters were fine-tuned via emulations. The performance metrics used for e comparisons were link utilization, queueing delay and packet loss. Table 6 shows e overall improvement made by AVQRED wi an assumption at each metric is equally important. The overall improvement was computed by averaging all e improvement measurements in Table 3, Table 4 and Table 5. Table 6. Overall improvement by AVQRED Comparisons between AVQRED and e oer six AQM meods can be concluded as: Figure 8. Link utilization The MATLAB results for queue size, Figure 6, are consistent wi e emulation results for queueing delay, Figure 4 and Table 4, because bo show at e AVQRED has lower queueing delay wi wider variation (which is due to unbiased drops against transient traffic). The MATLAB results for arrival rate, Figure 7, are consistent wi e emulation results for packet loss variation, Figure 5 and Table 5, because a higher arrival rate implies less number of consecutive drops which cause significant 1. AVQRED has slightly lower link utilization wi higher stability. 2. AVQRED has lower queueing delay wi higher delay jitter but it is less biased against short-lived traffic. 3. AVQRED has slightly fewer packet drops wi lower global synchronization level. To provide intuitive properties and validate e emulation results, a maematical model was proposed and fed to MATLAB. The MATLAB results were compared wi e emulation results and ey were 395

8 consistent each oer. We are currently evaluating e use of AVQRED in satellite networks where queue size based AQM meods do not fit well due to high buffering. 10. References [1] P. Lassila and J. Virtamo, Modeling e dynamics of e RED algorim, in Proceedings of QofIS 00, pp , September [2] P. Kuusela, P. Lassila, J. Virtamo and P. Key, Modeling RED wi idealized TCP sources, df, [3] S. Auraliya, V. H. Li, S. H. Low, and Q. Yin, REM: Active queue management, IEEE Network, vol. 15, no, 3 pp , May/June [4] S. Floyd and V. Jacobson, Random early detection gateways in congestion avoidance, IEEE/ACM Transactions on Network, vol. 1 no. 3, pp , [5] V. Misra, V. Gong, and D. Towsley, A fluid-based analysis of a network of AQM routers supporting TCP flows wi an application to RED, in Proceedings of ACM SIGCOMM, August 2000, pp [6] S. Floyd et al., Discussions on setting parameters, November [7] S. Floyd, R. Gummadi, S. Shenker, and ICSI. Adaptive RED: An algorim for increasing e robustness of RED s active queue management, Berkeley, CA. [8] W. Feng. D. Kandlur, D. Saha, and K. Shin, Blue: A new class of active queue management algorims, Tech. Rep., UM CSE-TR , [9] W. Feng, D. Dilip D. Kandlur, Debanjan Saha, and Kang G Shin, A self-configuring RED gateway, in Proceedings of IEEE/INFOCOM, [10] D. Clark and W. Fang, Explicit allocation of best-effort packet delivery service, IEEE/ACM Transactions of Networking, vol. 6 no. 4 pp , August [11] J. Padhye, V. Firoiu, D. Towsley, and J. Kurose. Modeling TCP roughput: A simple model and its empirical validation, in Proceedings of ACM/SIGCOMM, [12] T. J. Ott, T. V. Lakshman, and L. H. Wong, SRED: stabilized RED, in Proceedings of IEEE INFOCOM, March [13] W. Stevens. TCP/IP Illustrated, Vol. 1 The Protocols. Addison-Wesley, [14] W. Stevens, TCP slow start, congestion avoidance, fast retransmit, and fast recovery algorims, RFC2001, Jan [15] M. May, T. Bonald, and J. Bolot, Analytic evaluation of RED performance, in Proceedings of IEEE IFOCOM, March [16] C. Hollot, V. Misra, D. Towsley, and W-B. Gong, A control eoretical analysis of RED, in Proceedings of IEEE INFOCOM, [17] C. Hollot, V. Misra, D. Towsley, and Wei-Bo Gong, On designing improved controllers for AQM routers supporting TCP flows, in Proceedings of IEEE/INFOCOM, April [18] S. Kunniyur and R. Srikant, Analysis and design of an adaptive virtual (AVQ) algorim for active queue management, in Proceedings of ACM/SIGCOMM, August [19] K. K. Ramakrishnan and S. Floyd, A proposal to add explicit congestion notification (ECN) to IP, RFC 2481, Jan [20] R. J. Gibbens and F. P. Kelly, Distributed connection acceptance control for a connectionless network, in Proceedings of e 16 Intl. Teletraffic Congress, June