RBA-RIO Rate Based Adaptive Red With In and Out Algorithm for DiffServ AF PHB Zhang Mgjie Zhu Peidong Su Jshu Lu Xicheng School of Computer, National University of Defense Technology, Changsha 410073, Cha canicula@263.net Abstract-RIO is the active queue management algorithm for supportg DiffServ AF PHB. When the subscription level or the load varies, the performance of RIO will change accordgly. The average queue delay and lk utilization oscillate and do not converge to the ideal values. The paper proposes a new algorithm for DiffServ AF PHB, which is called RBA-RIO (Rate based Adaptive RED with In and Out). RBA-RIO consists of two sub-algorithms: adaptive RED and LPD (Loss Probability Divider). LPD calculates the drop probability of and out packets dynamically based on their arrival rate. Compared with RIO, RBA-RIO only needs to configure one set of parameters. RBA-RIO can achieve smaller average queue delay and higher lk utilization, and its performance advantage is verified usg ns simulations by comparison with RIO. Keywods: DiffServ, AF PHB, RIO, subscription level, utilization, average queue delay, adaptive I. INTRODUCTION The Internet, based on the TCP protocol, has succeeded providg worldwide data communication service for the past few decades. However, Internet does not provide any Quality of Service (QoS) guarantee to applications. With creasg emergence of new service types, such as real-time audio/video applications, there is an creasg demand for providg QoS support the Internet. The differentiated services (Diffserv) architecture [1] proposed by IETF has recently become the preferred method to address QoS issues IP networks. Customer s traffic is classified to different service classes and marked with different drop priorities (such as /out packets) at edge routers. Core routers only need to implement simple packet schedulg and droppg algorithm. This packet markg based approach to IP QoS is attractive due to its simplicity and scalability. In DiffServ networks, the externally observable forwardg behavior applied at a DiffServ-compliant node to a behavior aggregate is called Per-Hop-Behavior (PHB). IETF have standardized two basic PHBs: Expedited Supported by the National Natural Science Foundation of Cha under Grant No. 90204005 and 90104001; the National Hi-Tech Research and Development Program of Cha under Grant No. 2003AA121510. Forwardg (EF) [2] PHB and the Assured Forwardg (AF) [3] PHB. The EF PHB is used to build services that require low delay, low jitter and low loss like the Virtual Leased Le (VLL) services, while the AF PHB is used to build more elastic services that impose requirements only on throughput without any delay or jitter restrictions. This paper is focus on AF PHB. To build an end-to-end service with AF, subscribed traffic profiles for customers are mataed at the traffic conditiong nodes at the edge of the network. The aggregated traffic is monitored and packets are marked at the traffic conditioner. When the measured traffic exceeds the committed target rate, the packets are marked with high drop precedence (out); otherwise, packets are marked with low drop precedence (). Core routers implement active queue management schemes, such as RED with In and Out (RIO) [4], and provide service differentiation to the traffic accordg to pre-assigned service classes and drop priorities carried the packet header. As illustrated Fig.1, RIO uses the same mechanism as RED [5] but is configured with two sets of parameters, one for packets and the other for out packets. Upon each packet arrival at the router, the router checks whether the packet is tagged as or out. If it is an packet, the router calculates avgq_, the average queue for the packets; if it is an out packet, the router calculates avgq_out, the average total queue size for all (both and out) arrivg packets. The probability of droppg an packet depends on avgq_, and the probability of droppg an out packet depends on avgq_out. In Fig.1, m_ is bigger or equal to max_out. P max_out P out 1 1 avgq_out m_out max_out m_ max_ Fig. 1 RIO Algorithm Weighted RED (WRED) [6] is another AQM for supportg AF PHB. WRED calculates a sgle average P P max_ avgq_
queue that cludes arrivg packets of all priorities. For an arrival or departure of or out packets, WRED updates a sgle average queue based on total number of packets of or out. However, multiple RED threshold parameters are mataed - one for each priority. In [7] May et al perform some analytical modelg of DiffServ architecture schemes. Based on analytic evaluation of the loss probability, they conclude: choice of different [RIO] parameter values can have a clear impact on performance. Reference [8], through experiments, evaluates the performance of WRED and RIO. Performance dicators used the study cluded drop count of low drop precedence packets, transaction rate, throughput and number of retransmissions. They fd that the performance of RIO is better than WRED. This paper, through simulations, shows that when the subscription level or the load (connection number) varies, the performance of RIO will change accordgly. The average queue delay and lk utilization oscillate and do not converge to the ideal values. To improve the performance of RIO, the paper proposes a new algorithm for AF PHB, which is called RBA-RIO (Rate based Adaptive RED with In and Out). The rest of the paper is organized as follows. In Section II we demonstrate the weakness of RIO through simulations. We then propose RBA-RIO algorithm Section III. In Section IV, the validity of RBA-RIO is verified through ns simulations. Fally, we conclude our research Section V. II. WEAKNESS of RIO This section demonstrates the weakness of RIO through ns-2 [13] simulations. The simulation topology is illustrated as Fig.2. In Fig.2, FTP/TCPReno connections are established between S i and R i. E 1 E 2 are edge routers and mark packets accordg to the traffic profile. Time Slidg Wdow (TSW) [4] marker is adopted by E 1. The bottleneck lk is between core routers C 1 and C 2. RIO is run on C 1 and its parameters are listed Table I. Each S i R i connection pair has a target rate of 10 SL N Mb/s, where SL is subscription level, and N is the connection number. TABLE I RIO PAREMETER SETTINGS IN-profile OUT-profile M th 5 15 Max th 15 30 Max p 0.02 0.2 W q 0.002 0.002 Total simulation time lasted 400s and lk utilization from C 1 to C 2 is calculated by dividg total sent packet count durg terval [100s-300s] by the maximal packet count that can be sent on the lk. Fig. 3 shows the lk utilization. Fig. 3. Lk Utilization of RIO From Fig.3, we can observe that when resource is under subscribed, lk utilization creases along with the connection number crease. When subscription level approaches saturation, lk utilization changes little as the connection number creases. We also fd if connection number is fixed, lk utilization creases along with the subscription level crease. This phenomenon is more significant when connection number is small. Average queue length on router C 1, which is calculated by dividg the summation of sampled average total queue length by the samplg count, is depicted Fig.4. S 1 R 1 S 2 10M E 1 C 1 C 2 E 2 R 2 50ms S n Fig. 2 Simulation Topology Different scenarios have been simulated on this network to evaluate the performance of RIO. The subscription level is changed from under-subscription (20%) to over-subscription (140%), and connection number is changed from 20 to 100. R n Fig. 4. Average Queue Length of RIO From Fig.4, we can observe that when subscription level is fixed, average queue length creases along with the
connection number crease, and when connection number is fixed, average queue length creases along with the subscription level crease. Average queue length (delay) of RIO will reach different values at different network scenarios and therefore cannot be predicted advance. Network operators wish the lk utilization is as high as possible and dependent of subscription level or connection number. At the same time, real-time audio and video customers wish average queue delay is stable. But from the above simulations, we can see that RIO cannot achieve high lk utilization and stable average queue delay simultaneously. III. THE RBA-RIO ALGORITHM In this section, we discuss the design of the RBA-RIO algorithm detail. A. RBA-RIO Design To improve the performance of RIO, the paper design RBA-RIO for AF PHB. RBA-RIO is depicted Fig. 5. Upon a packet arrival: if(packet is ) update rate ; else update rate out ; update avgqlen; if( avgqlen max TH ) drop the packet; else if( mth avgqlen max TH ) calculate packet drop probability p; calculate p, p out accordg to p, rate, rate out if(packet is ) with probability p drop the arrivg packet; else with probability p out drop the arrivg packet; else if( avgqlen mth ) queue the arrivg packet. Variable: rate /rate out avgqlen m TH /max TH p p /p out (out) packet arrival rate; average queue length; low/high threshold of avgqlen; packet (/out) drop probability; /out packet drop probability; Fig. 5. RBA-RIO Algorithm In RBA-RIO, packet arrivg rate of and out is estimated usg TSW [4] method. Packet drop probability of and out packet is calculated accordg to the followg equations. rate rateout p pout p (1) rate rateout rate rateout 0 p (2) p out p rate rate rateout 1 (3) pout p p rate rate rateout 1 When calculatg, first let p 0, and then calculate p out accordg to (1). If p out 1, then let p out 1, calculate p accordg to (1). Above method can be extended simply to three-drop priorities (GREEN/YELLOW/RED). In RBA-RIO, the maximal packet drop probability is constantly tuned to adjust to current traffic conditions. The detailed adjustg procedure is described [9]. p 0 B. Characteristics of RBA-RIO Compared with RIO, RBA-RIO can achieve high lk utilization and stable average queue delay simultaneously. Furthermore, RBA-RIO only needs to configure one set of parameters rather than two or three sets of parameters. IV. SIMULATIONS In this section we evaluate the performance of RBA-RIO various traffic conditions usg simulations and compare it with RIO. In simulations, parameters of RBA-RIO are listed Table II. Parameter max p is the maximal packet drop probability and is adjusted to current traffic conditions. Parameter WLen is used for estimatg packet arrival rate. TABLE II RBA-RIO PAREMETER SETTINGS m TH 5 max TH 15 max p 0.02 W q 0.0008 WLen 5s A. Static Traffic Scenario Here we repeat the simulations section II, but core router C 1 runs RBA-RIO this time. Lk utilization and average queue length are depicted as Fig. 6 and Fig. 7, respectively.
Fig. 6. Lk Utilization of RBA-RIO B.1 Under subscription In this experiment, n (20) FTP sources start to send bulk data durg [0s-5s] and aggregate subscription is 4Mbps, and m (30) FTP sources beg to send data durg [100s-105s] and aggregate subscription is also 4Mbps. The total subscription level is 80%. Fig. 9-1 and Fig. 9-2 show the time evolution of lk utilization and average queue length, respectively. From Fig.9-1 and Fig. 9-2, we can observe that RBA-RIO has higher lk utilization and smaller average queue length oscillations than RIO. Fig. 7. Average Queue Length of RBA-RIO From Fig.6 we can see that lk utilization is high (>99%) for various subscription level and connection number. Fig.7 shows that average queue length, which is around target queue length (10~11), is much more stable than RIO,. Fig. 9-1. Utilization Comparison B. Dynamical Traffic Scenario To evaluate the performance of RBA-RIO dynamical conditions, the simulation topology depicted as Fig.8 is somewhat different from Fig.1. We add senders s 1 ~s m, and receivers r 1 ~r m. The latency from senders to edge routers is uniformly distributed between 10ms to 50ms. S 1 S 2 S n s 1 s2 s m E 1 E 3 10M C 1 C 2 E 2 E 4 Fig. 8. Simulation Topology R 1 R 2 Rn r 1 r 2 r m Fig. 9-2. Average Queue Length Comparison Packet drop rate of and out, which is calculated by dividg discarded packet count 1.5s by received packet count, are depicted Fig.9-3 and Fig.9-4, respectively. In Packet drop rate of RBA-RIO and RIO are both very small when the system is stable. But at time around 0s and 100s, packet drop rate of RIO appears a peak. From Fig.9-4, we can observe that out packet drop rate is quite different durg [100s-200s]. We measure the number of received packet and dropped packet between 140s and 150s. Duration [140s-150s], RIO received 2151 out packets and dropped 1057 out packets; the drop rate of out is 49%. RBA-RIO received 3324 out packets and dropped 723 out packets; the drop rate is 22%. Low out packets drop rate is consistent with high lk utilization.
Fig. 9-3. Loss Rate (IN) Comparison Fig. 10-2. Average Queue Length Comparison Packet drop rate of and out are depicted Fig.10-3 and Fig.10-4, respectively. It is seen (Fig.10-3 and Fig.10-4) that when resource is over subscription, packet drop rate of and out has little difference. Fig. 9-4. Loss Rate (OUT) Comparison B.2 Over Subscription Here, n (20) FTP sources start to send bulk data durg [0s-5s] and aggregate subscription is 5Mbps, and then durg [100s-105s], m (30) FTP sources beg to send data and aggregate subscription is 6Mbps. The subscription level is 110%. Fig.10-1 and Fig.10-2 show the lk utilization and average queue length evolution of RBA-RIO and RIO, respectively. From Fig.10-1, we can observe that lk utilization of RIO and RBA-RIO has little difference between 100s and 200s. But from Fig.10-2, the average queue length and oscillation of RIO is much bigger than RBA-RIO. Fig.10-3. Loss Rate (IN) Comparison Fig.10-4. Loss Rate (OUT) Comparison We also conducted extensive simulations usg web traffic, and the simulation results show that the performance of RBA-RIO under web traffic is similar to that of FTP traffic. Fig. 10-1. Utilization Comparison
V. CONCLUSIONS The paper proposes a new algorithm for AF PHB to elimate the weakness of RIO. The performance advantage of RBA-RIO over RIO is verified usg ns-2 simulations. It is demonstrated that RBA-RIO can achieve higher utilization and more stable average queue delay than RIO. The stablibility of average queue delay is crucial for real-time audio and video applications. The idea of RBA-RIO is not only applicable to RED but to some other active queue management algorithms. RBA-RIO can be abstracted as Fig. 11. CBA LPD p Green p Yellow Technical report, Nortel Networks, May 2000. [9] S. Floyd, R. Gummadi, and S. Shenker. Adaptive RED: an algorithm for creasg the robustness of RED s Active Queue Management. http://www.icir.org/ floyd. Aug 2001. [10] Hollot, V. Misra, D. Towsley, and W. Gong. On Designg Improved Controllers for AQM Routers Supportg TCP Flows. Proceedgs of IEEE INFOCOM 2001. [11] Zhang Heyg, Liu Baohong and Dou Wenhua, Design of a Robust Active Queue Management Algorithm Based on Feedback Compensation. ACM SIGCOMM 2003. [12] Yuan Gao, Jennifer C. Hou. A State Feedback Control Approach to Stabilizg Queues for ECN-Enabled TCP Connections. IEEE INFOCOM 2003. [13] Ns-2 Network simulator, http://www.isi.edu/nsnam/ns. Fig. 11. RBA-RIO Abstraction p Red CBA (Color-bld AQM) block is the basic AQM for calculatg total packet drop probability p, and LPD (Loss Probability Divider) block is responsible for dividg p to p GREEN, p YELLOW and p RED. Besides RED and its variants, PI [10], PIP [11] and SFC [12] are also AQM algorithms used best-effort networks. If CBA is substituted by these AQMs, we can design different buffer management algorithms for DiffServ AF PHB. Performance evaluation of these algorithms will be part of our future work.. REFERENCES [1] S. Blake, D. Black, M. Carlson, E. Davies, Z. Wang, W. Weiss. An Architecture for Differentiated Services, RFC 2475, December 1998. [2] V. Jacobson, K. Nichols, K. Poduri, An Expedited Forwardg PHB, RFC2598, June 1999. [3] J. Heanen, F. Baker, W. Weiss, and J. Wroclawski, Assured forwardg PHB group, RFC 2597, June 1999. [4] Clark D. and Fang W., Explicit Allocation of Best Effort Packet Delivery Service, IEEE/ACM Transactions on Networkg, V.6 N. 4, August 1998. [5] S. Floyd and V. Jacobson. Random Early Detection Gateways for Congestion Avoidance. IEEE/ACM Transactions on Networkg, 1(4): 397 413, Aug.1993. [6] http://www.cisco.com/univercd/cc/td/doc/product/softwa re/ios112/ios112p/gsr/wred_gs.htm [7] May M, Bolot JC, Jean-Marie A, and Diot C, Simple performance Models of differentiated services schemes for the Internet, Proceedgs of INFOCOM'99. [8] R. Makkar et al., Empirical study of buffer management schemes for diffserv assured forwardg PHB,