DESIGN OF ACTIVE QUEUE MANAGEMENT BASED ON THE CORRELATIONS IN INTERNET TRAFFIC

Transcription

1 DESIGN OF ACTIVE QUEUE MANAGEMENT BASED ON THE CORRELATIONS IN INTERNET TRAFFIC KHALID S. AL-AWFI AND MICHAEL E. WOODWARD { k.s.r.alawf, m.e.woodward }@bradford.ac.uk Department of Computing, University of Bradford Richmond road, Bradford, BD7 1DP, UK Abstract. The implementation of Active Queue Management (AQM) in Internet routers is a recommendation of e Internet Engineering Task Force. However, none of e current AQM algorims directly makes use of e correlations present in Internet traffic, which are related to Long Range Dependence (LRD) and Self-Similarity. This is somewhat surprising in at it is well known at LRD is exhibited by Internet traffic, particularly at e edges of e network. In is paper, an algorim called Modified Adaptive Random Early Detection is presented. This algorim makes use of e correlations in e mean queue leng in order to predict likely congestion scenarios in e not too distant future. These can en be acted on anticipatively. The results are extremely encouraging and demonstrate a superior performance when benchmarked against current mechanisms. The proposed algorim could also be non-intrusively implemented in current routers. Keywords. Active Queue Management, Random Early Detection, Long Range Dependence, Internet Traffic. 1. INTRODUCTION Increased congestion has been experienced in various links in e Internet wi loss and delay increasing as e degree of congestion increases [Paxson,1997]. During e last decade, e use of Active Queue Management or AQM systems based on Random Early Detection or RED has led to a substantial increase in e efficiency of using e network resources. As a result, e Internet Engineering Task Force (IETF) has recommended e use of AQM in Internet routers [Barden et al, 1998; Mankin and Ramakrishnan, 1991]. There are different algorims of AQM systems and e recommended one is RED, which shows many advantages over e oer AQM systems. For instance, RED prevents global synchronization by having a random marking or dropping probability, no bias against burst traffic and reduced packet losses [Ohsaki and Murata, 2002]. RED is designed to accompany transport layer protocols, such as TCP, and enhances TCP capability by reporting congestion at e gateway. It was first proposed in e early 1990s by Floyd and Van Jacobson [Floyd and Jacobson, 1993]. Gentle RED (GRED) is e revised version of e original RED [Floyd, 2000]. GRED is proposed by Floyd in response to e analysis of Firoiu and Borden [Firoiu and Borden, 2000]. However, e main problem wi RED and Gentle RED is e mean queue leng (MQL) variations. The algorims compute e MQL wi every arriving packet to e queue but, unfortunately, e MQL is very sensitive to e level of congestion and parameter configurations. In an attempt to overcome ese problems, adaptive RED (ARED) is proposed by Feng et al. [Feng et al, 1997; Feng et al, 1999] and a suggested modification is presented by Floyd [Floyd et al,2001]. In is paper, we refer to ARED as e development by Floyd [Floyd et al,2001]. The shape of Internet Traffic is found to follow Self-Similarity and Long Range Dependent (LRD) processes [Alawfi and Woodward, 2005]. However, none of current AQM mechanisms embed ese Internet traffic characteristics. The second order Self Similarity is referred to as LRD. ARED increases e dropping probability when e MQL exceeds e target. The proposed algorim is called modified ARED. Raer an maintaining MQL target, e I. J. of SIMULATION Vol. 8 No 3 26 ISSN x online, print

2 modified ARED uses e continuous increase of e MQL as an indication at sources are increasing e sending rate and consequently increased loss rate and delay. The basic concept of e proposed algorim is explained in e following section. Then, e results and analysis of e proposed algorim are compared wi ARED and original RED. Finally, conclusions and future work are presented. no dropping during e decreasing interval. The modified ARED algorim differs from e ARED algorim in computing, max p which is computed using e following algorim. 2. MODIFIED ARED 2.1 The Algorim The ARED algorim computes e maximum drop probability ( max p ) based on e target MQL. The proposed modified ARED computes max p using e fact at Internet traffic is following an LRD process and any LRD process is more likely to continue, ON, for e near future, when a process is ON for a long time [Feng et al, 1999]. The max p continues to increment as e ON time get longer. In oer words, as e MQL increases, e max p increases. Continuous increase of e MQL for a long time indicate an increasing correlation. It has been recommended at e AQM incorporate and LRD correlation structure [Ostring and Sirisena, 2001]. When e MQL is increasing, e AQM should use is as an indication at sources are sending more traffic. As LRD process, is indicates at is moment at sources will continue to be ON. This is directly related to e correlation of e traffic which follow LRD process. Increase of correlation results in increase of e delay and loss rate. Sources, which are TCP in is research, will suffer longer round trip time or RTT. Hence increasing e max p wi correlation will results in dynamically configured parameters to target specific quality of service. The algorim uses e Additive-Increase Multiplicative-Decrease (AIMD) policy of setting, max p which is incremented wi α every time a packet arrives to e queue to a a maximum value of a range as shown in e algorim. When e MQL is decreasing, max p is set back to 0 which means The dropping probability P b is linearly computed by e equation : Pb max p ( MQL min ) /(max min ), where, max is e maximum reshold and min is e minimum reshold.the final dropping probability, P a is computed by Pa ( Pa /(1 c * Pb )), where c is a counter used to avoid waiting too long before dropping a packet and maintains uniform packet dropping. The counter increases by one wi every arrival and reset to 0 when a drop occurs. The detailed parameterization is discussed in e original RED paper [Floyd and Jacobson, 1993]. Existing AQM algorims drop packets at increasing or decreasing intervals. Furermore, e modified ARED algorim drops packet only at e increasing intervals as shown by Figure 1. Modified ARED drops packets only when e MQL is increasing. The increasing degree of e MQL means an increasing degree of e correlation and congestion. The dropping probability increases as a reaction to e congestion status change. The dropping probability is 0 when e MQL is decreasing because max p is set to 0 at is instance as shown in e algorim. This way I. J. of SIMULATION Vol. 8 No 3 27 ISSN x online, print

3 congestion degree is decreased as will be shown in e next experiments section. 2.2 Experimental Settings A simulation package is developed using Java programming language to carry out ese experiments. All statistics are computed from e second half of e experiment time, which is ms. The experiment configuration is as follows. The number of TCP connections is 10, min is 15, max is 100, buffer size is 150 packets, exponential weighted moving average ( w q ) is.002 and e queue is initialized wi 10 packets. RED, ARED and modified ARED drop packet wi uniform dropping probability P a which increases linearly between e resholds, min and max from 0 up to max p. The router service rate is 4 packets/ms. The experiments are tested wi increasing sending rate of e TCP connections from.01 to 2 packets/ms. The next section uses ese configurations to invoke experiments on Drop Tail or DT, RED, ARED and e modified ARED. 3. RESULTS AND ANALYSIS 3.1 Performance results and analysis The modified ARED shows better performance gain in comparison to ARED and RED. RED is configured wi max p of 0.1 in order to show e comparison wi ARED and modified ARED. RED will show different performance wi different Figure 1: MQL Profile configurations because it is a static AQM mechanism. The main comparison is us between ARED and modified ARED, which are bo dynamic AQM mechanisms. Figure 2 shows e normalized delay and normalized roughput trade offs. The delay is 40 wi RED, 60 wi ARED, less an 40 wi e modified ARED and almost 120 wi DT. I. J. of SIMULATION Vol. 8 No 3 28 ISSN x online, print

4 Furermore, Figure 3 shows e difference in loss between e different algorims. Modified ARED shows lower delay for e same loss rate compared to ARED and RED. DT performance is e worst and e modified ARED is e best in all comparisons. Figure 2: Normalized Delay and Normalized Throughput Trade offs RED could show better performance wi different configurations, such as increasing e dropping probability. However, ARED and modified ARED maintain a dynamic dropping probability, which modifies e performance according to e traffic conditions. 3.2 Queue Oscillation The queue oscillation is investigated using e plotted profile of each AQM tested. Experiments for Figure 3: Normalized Delay and Loss Rate Trade offs is purpose are ms long and start wi 5 TCP connections. After 5000 ms anoer 5 TCP connections are initialized. Figure 4 show at ARED has succeeded in controlling he MQL at a specific target. Even after an increased load, e MQL goes back gradually to e target MQL. However, RED does not control e MQL wi low or high load, where e MQL max max is almost equal to e as Figure 5 illustrates. As shown by Figure 6, e MQL is stable and I. J. of SIMULATION Vol. 8 No 3 29 ISSN x online, print

5 changes level wi e load degree using e modified ARED. The modified ARED differs from ARED and RED in two points wi regard to MQL oscillation. The MQL is almost stable at a level relevant to e load condition and e queue does not oscillate much. Having e MQL stable is e reason of having a lower queue oscillation. Figure 4: ARED MQL Profile Figure 5: RED MQL profile I. J. of SIMULATION Vol. 8 No 3 30 ISSN x online, print

6 Figure 6: Modified ARED MQL Profile. 4. LIMITATIONS The modified ARED is very sensitive to e increment and decrement factors. Also, e modified ARED shows lower performance an ARED at high load and slower service time as shown by Figure 7. The following modification to e modified ARED algorim is used to help avoiding high MQL level when e load is high or e service time is relatively long. The increment factor is dynamically configured and e max p is decreased by a factor of 0.9 raer an being set to 0 when a decrease in e MQL occurs. In general e revised algorim uses two modifications which are e dynamic increment factor and slowly decreasing max p. The two algorims are compared using e same performance metrics. Wi e same conditions used earlier in Figure 7 to examine e limitations. The new algorim is successful in achieving a lower MQL at high load and long service time as shown in Figure 8. The modification of e algorim is as follows: I. J. of SIMULATION Vol. 8 No 3 31 ISSN x online, print

7 Figure 7: Comparing ARED and modified ARED wi Higher Load Figure 8: Comparing ARED and revised modified ARED wi Higher Load I. J. of SIMULATION Vol. 8 No 3 32 ISSN x online, print

8 The revised version shows at e algorim can retain its performance gain even when e traffic or e service time of e router changes. The worst case of e queueing delay is considered earlier to test e algorim at high load wi low service time. However, it is not a goal for is algorim to control e worst case queueing delay, as recommended by [Floyd et al,2001]. RED is max p configured wi of 0.2 raer an 0.1 as previous in order to give a different performance to compare wi ARED and revised modified ARED. The performance measures wi e revised algorim are shown wi Figure 9 and Figure 10. The performance conclusions are e same as earlier conclusions from Figure 1 and Figure 2. Figure 9: Normalized Delay and Normalized Throughput Trade off 5. CONCLUSIONS AND FUTURE WORK The modified ARED algorim gives superior performance in a trade off comparison of delay, loss rate and roughput an eier ARED or RED. The Figure 10: Normalized Delay and Loss Rate Trade off modified ARED shows better performance when having a dynamic increment factor and slowly decreasing max p. One important feature of ARED is to control e MQL to a specified target. Future I. J. of SIMULATION Vol. 8 No 3 33 ISSN x online, print

9 implementation will investigate e implementation of target MQL for modified ARED using an additional reshold or some oer appropriate meod. RED is well known for its early congestion notification. Modified ARED detect congestion even earlier regardless of traffic conditions. Such feature is important wi high speed networks. REFERENCES K. S. Alawfi and M. E. Woodward, "Implications of Self Similar Traffic on Congestion Control Mechanisms," presented at Proceedings of e Six Informatics Workshop for Research Students, University of Bradford, Bradford, UK, B. Braden, D. Clark, J. Crowcroft, B. Davie, S. Deering, D. Estrin, S. Floyd, V. Jacobson, G. Minshall, C. Partridge, L. Peterson, K. Amakrishnan, S. Shenker, J. Wroclawsk, and L. Z. a. UCLA, "Recommendations on Queue Management and Congestion Avoidance in e Internet," Request for Comments: 2309, April V. Firoiu and M. Borden, "A study of active queue management for congestion control," INFOCOM Nineteen Annual Joint Conference of e IEEE Computer and Communications Societies. Proceedings, vol. 3, pp , March W. Feng, D. Kandlur, D. Saha, and K. Shin, "Techniques for eliminating packet loss in congested TCP/IP networks," University of Michigan November March S. Floyd, R. Gummadi, and S. Shenker, "Adaptive RED: An Algorim for Increasing e Robustness of RED's Active Queue Management," unpublished, A. Mankin and K. Ramakrishnan, "Gateway Congestion Control Survey," Request for Comments: 1254, August Hiroyuki Ohsaki and M. Murata, "Steady State Analysis of e RED Gateway: Stability, Transient Behavior, and Parameter Setting," IEICE TRANS. COMM, vol. E85-B, January V. Paxson, "End-to-End Internet Packet Dynamics," IEEE\slash ACM Transactions on Networking, vol. 7(3), pp , S. A. M. Ostring and H. Sirisena, The influence of long-range dependence on traffic prediction, IEEE ICC01,Helsinki, Finland, vol. vol. 4,p. pp , June 2001 BIOGRAPHY Khalid S. R. Alawfi received his PhD degree (2006) in Active Queue Management based on e use of Self Similarity and Long Range Dependence found in Internet Traffic from e University of Bradford UK. He is a lecturer in e Department of Computing, University of Taibah, Madina, Saudi Arabia which he joined in W.C. Feng, D. Kandlur, D. Saha, and K. Shin, Professor Mike Woodward graduated wi a first "A Self Configuring RED Gateway," In class honours degree in Electronic and Electrical Infocom'99, March Engineering from e University of Nottingham in 1967 and received a PhD degree from e same S. Floyd and V. Jacobson, "Random early institution in 1971 for research into e decompostion detection gateways for congestion avoidance," of sequential logic systems. In 1970 he joined e Networking, IEEE/ACM Transactions, vol. 1, staff of e Department of Electronic and Electrical pp , Aug Engineering at Loughborough University as a lecturer, being promoted to Senior Lecturer in 1980 S. Floyd, "Recommendation on using e and Reader in Stochastic Modelling in He gentle variant of RED. remained at Loughborough unitil 1998 when he was appointed to e Chair in Telecommunications at e University of Bradford where he also became e I. J. of SIMULATION Vol. 8 No 3 34 ISSN x online, print

10 Director of e Telecommunications Research Centre. He is currently e Head of e Department of Computing at e University of Bradford. His current research interests include queueing networks, telecommunications traffic modelling, quality of service routing and mobile communications systems and he is e auor of two books and over one hundred research papers on e above and related topics. He currently holds two EPSRC grants and is supervisor to sixteen full time research students. Professor Woodward is a Fellow of e Institute of Maematics and its Applications (FIMA) and is a Chartered Maematician (Cma), Chartered Scientist (CSci) and a Chartered Engineer (Ceng). I. J. of SIMULATION Vol. 8 No 3 35 ISSN x online, print