On the Use of Rate Configuration in the Interoperation between DiffServ and 802.11e EDCA A. Bai, T. Skeie, P. E. Engelstad, Member, IEEE Abstract - This paper investigates rate configuration of the Expedited Forwarding (EF) class of Differentiated Services (DiffServ) when used with 802.11e EDCA. The rate configuration problem is presented, and several approaches are tested and evaluated in order to solve the problem. Results reveal that the contention window makes rate configuration very hard for the highest priority class (AC_VO) in 802.11e EDCA. Our evaluations show that 802.11e EDCA is not able to conform to DiffServ’s EF PHB specifications. 1 Keywords: IEEE 802.11e EDCA, DiffServ, PHB, Quality-ofService (QoS), Rate configuration, Drop. I. INTRODUCTION The demand for QoS provisioning is increasing in today’s best effort Internet, and the Differentiated Services (DiffServ) architecture [1] represents a promising candidate technology for end-to-end QoS. Among the approaches that exist, DiffServ is foreseen to become the de-facto standard for providing IP QoS. This architecture is simple and more scalable than per-flow QoS architectures, such as IntServ. As wireless networks are being more and more widely deployed, the demands for QoS in the wireless networks are also increasing. The wireless access network will often be the bottleneck on the path between the source and the receiver, because of the nature of wireless communication. It is therefore important to integrate the WLAN QoS architecture with the end-to-end QoS concept adopted, for example DiffServ, to support applications with QoS requirements on wireless terminals. The IEEE 802.11e standard was developed in order to meet the QoS requirements in 802.11 WLANs [2]. The focus in this paper is on the 802.11e EDCA, which gives differentiation between four different priority classes – called Access Categories (AC). This paper addresses the interoperation of 802.11e EDCA and the DiffServ architecture. It contributes with an Manuscript received Feb 10, 2006. This work has been supported by the OBAN project of the European Commissions 6th Framework Program. Other OBAN partners are not committed under any circumstances by its content. Alex Bai is a master student at University of Oslo and Telenor R&D, Norway (email: aleksab@ifi.uio.no). Tor Skeie is with Faculty of Informatics, University of Oslo, Norway (email: tskeie@simula.no). Paal E. Engelstad is with Telenor R&D, 1331 Fornebu, Norway (phone: +47 41633776; fax: +47 67891812; e-mail: engelstad@ieee.org). He is also associated with UniK / University of Oslo. evaluation of rate configuration for the Expedited Forwarding (EF) class. This evaluation is done both analytically and by simulations. Rate configuration (rate limiting) is a mechanism that restricts the throughput for a class or priority, so the throughput will never exceed a configured level. Rate configuration is needed by admission control algorithms in order to i.e. guarantee that one class will not harm transmission of other classes, and is therefore an important tool that must be well understood. The next section gives an introduction to the previous works in the field of interoperation between DiffServ and IEEE 802.11e. It also presents the rate configuration problem. Section 3 addresses this problem by an analytical approach. Section 4 presents results obtained from the simulations. An interesting property of the drop performance of the highest class in 802.11e EDCA is identified and discussed in Section 5. Finally the conclusions are drawn in Section 6, and suggestions for future work are highlighted in Section 7. II. BACKGROUND We have recently proposed a QoS architecture for DiffServ in 802.11e [3]. In this architecture the whole mapping between 802.11e's Access Categories and DiffServ's Per-Hop Behaviors (PHBs) is located in a separate and independent module. This module is called the “QoS Mapping Module” (QMM), and it makes sure that a given DiffServ PHB corresponds to a given 802.11e access category. Since all mappings are placed inside the QMM, neither signaling nor communication is needed between the DiffServ node and the QoS-enabled Access Point (QAP). In this way transparency is achieved. The QMM is very simple in its functioning, because it only performs table lookup, as well as reading and writing of the MAC and IP headers in a packet. The QMM may be placed inside the QAP, as depicted in Figure 1 (the figure just illustrates the QMM concept and therefore the QAP is not given in detail). When a packet arrives from a DiffServ edge node, the module intercepts the packet and gives it an 802.11e priority that corresponds to the DiffServ's DSCP value. When a packet comes from a QoS-enabled Station (QSTA), it is given a DiffServ DSCP value. A typical example is the mapping of the DiffServ Code Point (DSCP) for Best Effort to 802.11e AC_BE. http://folk.uio.no/paalee/ specification [4] will go to infinity, and the system will not conform to the specification. Setting the rate configuration of the EF traffic to a reasonable level is a prerequisite for conformance with the standard. Since this rate configuration setting is obviously strongly dependent on the number of actively transmitting stations, finding an appropriate rate control value can be a difficult problem. This is what we refer to as the rate configuration problem. In the following, we try to solve this problem by an analytical approach. III. ANALYTICAL APPROACH Fig. 1. Proposed architecture. The proposed mappings in [3] between DiffServ's PHBs and 802.11e's traffic classes use only three out of the four available ACs. The EF PHB is mapped to the highest priority AC, AC_VO. The AF PHB is mapped to AC_VI and the BE PHB is mapped to AC_BE. To be able to control the rate of the EF traffic transmitted on the wireless channel, admission control is needed for AC_VO. This rate control value is here assumed statically set in the QAP. For further information about the architecture, the QMM module and how 802.11e's access categories conform to the EF, AF and BE PHB requirements, refer to [3]. A. Rate configuration Throughput - 16 stations scenario 2000 1800 parameters τ Throughput [Kbps] 1600 1400 i , Pi , Pb , P s ,i and Ps . τi is the probability for a station in priority class i to transmit during a generic slot time, Pi is the collision probability for the 1200 1000 800 priority class i, P 600 400 b is the probability that the channel is busy, P s , i is the probability of a successful transmission in 200 0 480 To conform to the EF PHB specification [4], queue overflow for the EF class must be avoided. It would therefore be beneficial if it was possible to make an algorithm or a method for deciding the correct rate configuration value [3]. Then it would be possible to calculate the maximum rate configuration value for a given set of stations that would satisfy the latency demands for the EF class. An obvious approach would be to analyze an analytical model of 802.11e and extract an algorithm. Xiao has presented an analytic model of 802.11e, which has proven to be reliable when the system is in saturation [5]. Since rate configuration only makes sense when the system is in saturation, it is not necessary to use more advanced nonsaturation models, e.g. such as the model proposed by Paal Engelstad and Olav Østerbø in [6]. Only the relevant equations from Xiao’s 802.11e analytical model are given in this paper. A full overview of Xiao’s analytic model can be found in [5]. Xiao’s model encapsulates everything into five basic equations for the 2480 4480 6480 8480 10480 12480 Total offered load [Kbps] EF AF BE Fig. 2. Throughput for 16 stations. For the scenarios with 12 and 16 stations the throughput results of the simulations were not as expected [3]: The EF class was rate configured (limited) to 5% of the channel rate, but instead of staying at the configured throughput level, the EF class converged to a value below this (Figure 2). The configured rate for EF traffic was obviously set too high, and because of the high number of stations, this lead to many collisions on the channel. Therefore the actual throughput for EF traffic became lower than the rate configuration level of 5% of channel rate (550Kbps). This is clearly not a desired result since it will make the queue length for EF traffic grow towards maximum, and latency demands for EF will not be satisfied. Thus, the E_p parameter presented in the EF PHB http://www.unik.no/personer/paalee a slot time for priority i and P s is the probability for a successful transmission in a slot time. By combining these five equations, Xiao finds the maximum throughput for each class as: Si = p s , iT E ( L ) (1 − p b ) ∂ + p s T s + [ p b − p s ]T c where T E ( L ) , ∂ ,T s and T c (1) are constants. A maximum rate configuration level can be found by letting the number of stations in each priority class go to infinity, and see if S i has an asymptotic value. If an asymptote can be found, a maximum throughput level is also found. It would then be possible to find a generic algorithm for a given set of nodes. In order to find the limits of S i when the number of stations goes to infinitive, a closer look at the behavior of the parameters τ i , P i , Pb , P s , i and P s In search of a rate configuration method, numerous simulations where conducted. The ns-2 discrete event simulator (version 2.26) was used for the evaluations, together with an 802.11e EDCA extension model implemented by the TKN group at the Technical University of Berlin [7][8]. The configurations that were used for the simulations are listed in Table 1 is needed. A. The limits of P s , i P s , i is given by Ps ,i = niτ i (1 − τ i ) N −1 ∏ (1 − τ ni −1 h ) nh (2) h =0,h ≠ i where n is the number of stations. The valid domains for τi and τh are 0 ≤ τ i ≤ 1, 0 ≤ τ h ≤ 1 . For simplicity, the analysis can be broken into two cases. The first case is for τ i ∈ 0 ,1 ,τ h ∈ 0 ,1 , and the other case [ ] [ ] is for 0 < τ i < 1,0 < τ h < 1 . It is easy to see from Eq. (2) that the term τ i ( 1 − τ i ) will be zero when τ i ∈ [0 ,1 ] . The N −1 ∏ (1 − τ h ) n h term will be zero when τ h = 1 , and h = 0 ,h ≠ i when τ h = 0. Hence, when τ i ∈ [0 ,1 ], τ h ∈ [0 ,1 ] . Ps,i will be zero For the remaining case ( 0 < τ i < 1,0 < τ h < 1 ) , we observe that the terms (1 − τ i ) n i − 1 and also go to zero when n → ∞ , since: lim ln(1 − τ )n = lim n ln(1 − τ ) = −∞ n−>∞ n−>∞ IV. SIMULATION APPROACH (1 − τ h ) nh will (3) Parameter Channel Rate Packet size Traffic generator Link delay Stabilize time Simulation duration Queue length Physical layer settings EDCA parameters [AIFS, CWmin, Cwmax, TXOP limit] Table 1. Detailed configuration for the scenarios The goal of the simulations was to find the correct rate configuration value for the EF class in the 16 nodes scenario. The motivation behind this was to compare the results found by simulation with numerical analysis. If numerical and simulation results correlate, further numerical results can be extracted and extended to a rate configuration formula. While performing the simulations, an interesting property was revealed. It was not possible to set a rate configuration level that would guarantee that the delay bounds would be fulfilled. In other words, it was not possible to find a rate configuration level that the EF class could maintain. Drop percentage - 16 stations scenario 25 It can be shown that the divider of S i will go towards 1 n → ∞ by looking at the limits for when τ i , P i , P b , P s . These proofs have been left out because of limited space, but the reader is encouraged to confirm the results. Therefore the limits of S i is lim S i = 0 . n −>∞ This is not a desired result because the motivation behind this analysis was to find a non-zero asymptote for S i , and thus be able to find a rate configuration method for the EF class. Another approach is needed if a rate configuration method is to be found. Drop percentage [%] n→∞. Si AF (AC_1) = [2, 7, 15, 3] BE (AC_3) = [7, 15, 1023, 0] (3, 5, 8, 12, 16) Number of stations Hence, in any of the possible cases, Ps ,i → 0 when B. The limits of Value 11 Mbps 500 bytes (fixed) All nodes sending CBR traffic 1us 50s 180s 5 packets 802.11b EF (AC_0) = [2, 3, 7, 1.5] 20 15 10 5 0 30 120 210 300 390 480 570 660 750 840 930 1020 1110 1200 1290 1380 Total offered throughput [Kbps] EF AF Fig. 3. Drop percentage for the EF class (in a 16 node scenario). Figure 3 shows the drop percentage for the EF class for the scenario given in section 2.1. The EF class was rate configured to 5% of the channel rate (i.e. to 550 kb/s) while the AF class was not rate configured at all. As shown with both throughput and drop percentage results (Figure 2 and 3), the EF class was not able stay at the configured level. We also ran other simulations where the EF class was rate configured to 3%, 1%, 0,5% and even 0,03%. Those p=e 1 ln 0, 24 7 = 0,82 (4) The division by seven in Eq. (4) reflects the recommended number of retries of 802.11e EDCA (which our simulations also used). A collision probability of 82% is not unthinkable when there are 16 stations and the channel is in saturation. Thus it really should not come as a surprise that it is not possible to rate configure the EF class when the number of stations is high. Several other simulations were performed with different EDCA parameters. These results are omitted because of limited space. The simulations were still not able to perform Another interesting feature also emerged during the simulations. If rate configuration was applied to all the three classes (EF, AF and BE) during the simulation, the drop percentage would stabilize. Both the AF class and the BE class had to be rate configured along with the EF class in order to achieve this result. If just one of the three classes was rate configured too high or not at all, the drop percentage would rise and the queue delay would also continue to rise. Drop percentage - all classes rate configured 5 4,5 4 3,5 3 2,5 2 1,5 1 0,5 99 0 11 25 12 60 13 95 15 30 16 65 18 00 19 35 20 70 85 5 72 0 58 5 45 0 31 5 0 45 The reason for this unexpected behavior of the rate configuration is the EDCA parameters. All the simulations were performed with recommended EDCA parameters. However, the contention window for the highest priority class in 802.11e is initially very small [2]. An understanding of what kind of data the highest priority class is supposed to carry is needed in order to understand why the EDCA parameters are set the way they are. When IEEE standardized 802.11e, the purpose of the highest priority class (priority 0) was different from DiffServ’s highest priority class (EF). The EF class is not intended for any specific data type, but is just meant to deliver the data as fast and reliable as possible (and better than all lower priority classes). The voice class (priority 0) in 802.11e is, however, intended for voice traffic. This is the fundamental difference between the two standards and the reason why rate configuration is so difficult. The recommended EDCA parameters for the highest class in 802.11e is set to achieve fast access to the medium at all costs, even on behalf of some packet loss. This is why the contention window is so small, since the intended data for the class is voice traffic. It is less crucial if a voice packet is discarded than a video packet (which has lower priority and a higher contention window). The impacts of a small contention window are obvious in our simulations results. Because of the high number of stations and many packet collisions, the drop percentage for the highest 802.11e class rises very high. A drop percentage of 24% (maximum drop percentage in Figure 3) for the highest priority in 802.11e would indicate a collision probability of 82%: A. Drop percentage stabilizes 18 0 V. DROP PERCENTAGE at the rate configuration level, although in some scenarios it was possible to find a better match. Whether it is possible to find an EDCA setting that makes it possible to rate configure the EF class is left open. A more detailed study is needed to confirm or discard this option. Drop percentage [%] simulations showed the same result as in Figure 3. The EF class could not maintain the throughput level and dropped below the configured rate level. This is a very interesting property of 802.11e EDCA that has not been documented before to our best knowledge. This result has a large impact on admission control algorithms, since it will be difficult to rate limit the highest priority class in 802.11e as expected. This means that admission control will be very hard to perform. Total offered load [Kbps] EF AF BE Fig. 4. Drop percentage stabilizes. As shown in Figure 4, the drop percentages for the EF and AF classes stabilized and the throughput therefore also stabilized. In this simulation, the EF class was rate configured to 5%, AF to 1% and BE to 1% of the channel rate. This is exactly the same scenario that is listed in section 2.1, except that the AF and BE class were also rate configured. The drop percentages stabilize in this scenario because the channel is not in saturation. This might be an approach for an admission control that needs rate configuration. By rate configuring all the 802.11e classes and taking the drop percentage into account, a maximum throughput limit can be obtained. Figure 5 shows a scenario where this approach could have been used. It is the same scenario as shown in Figure 4, where the highest priority class was rate configured to 5%, the AF class to 1% and BE class to 1%. In Figure 5 it is observed that the throughput stabilizes for the EF class. Throughput - All classes rate configured 600 Throughput [Kbps] 500 400 300 200 100 5 0 5 0 0 20 7 19 3 18 0 16 6 15 3 13 9 5 0 5 12 6 99 0 11 2 85 5 58 5 72 0 45 0 31 5 45 18 0 0 Total offered load [Kpbs] EF AF A simple and effective solution to the rate configuration problem is to rate configure all three classes simultaneously. This is because the number of collisions is brought under control, and the drop percentages are stabilized. However, since the drop probability for the EF class is still higher than the AF and BE class, 802.11e is still not able to fulfill the EF PHB specification [4]. In summary, our evaluations show that the 802.11e EDCA is not able to conform to DiffServ’s EF PHB specifications, because of the rate configuration problem regarding the drop mechanism for the highest priority class in 802.11e EDCA. BE VII. FUTURE WORK Fig. 5. Throughput for a scenario where all the classes are rate configured. As both Figure 4 and 5 shows, better results are achieved when rate configuring is applied to all classes. Comparing the EF class throughput in Figure 5 with that in Figure 2 shows radical improvements; the EF class achieves a stabile throughput and it gets 100 Kb/s more throughput when the AF and BE class is also rate configured. The EF class is, however, still not able to stay at the configured rate level because of the drop percentage. As explained in section 5, this is because of how the contention window works. VI. CONCLUSION This paper has gone deeper into the evaluation of how the IEEE 802.11e standard for WLAN QoS is able to interoperate with DiffServ. A brief overview of the architecture along with its mapping was given, before the rate configuration problem was explained in more detail. An analytical study of the throughput for IEEE 802.11e was given. The analysis concluded that the throughput goes towards zero when the number of stations goes toward infinity. This was perhaps not an unexpected result, but even so it is important to investigate the properties of the throughput and establish the result. The goal behind finding an asymptote for the throughput was to find a rate configuration level for the EF class. Since this turned out negative, more simulations were performed in order to find a rate configuration level by simulations. It turned out to be impossible to rate configure the EF class in 802.11e, regardless of how low or how high the rate configuration was. Because of the EDCA parameters, many collisions will occur because of a small contention window and therefore the drop percentage could not be eliminated no matter how the EF class was configured. This problem has never been documented before, and is a very serious problem when considering admission control for 802.11e EDCA. Being able to rate limit a class is usually an absolute demand in admission control algorithms, and solutions to work around the rate configuration problem should therefore be investigated. More study on how to set the EDCA parameters when rate configuring the EF class is needed. It may be possible to achieve close to zero drop probability for the EF class if the EDCA parameters are set differently. This must, however, not destroy the transmission of the other classes. A more comprehensive study of the drop problem for the highest class of 802.11e EDCA is needed. There might exist solutions to the rate configuration problem, or solutions that will efficiently eliminate the problem. VIII. REFERENCES [1] S. Blake, D. Black, M. Carlson, E. Davies , Z. Wang, W. Weiss, 'An Architecture for Differentiated Services', IETF RFC 2475, December 1998. [2] Wireless Medium Access Control (MAC) and Physical Layer (PHY) specifications: Medium Access Control (MAC) Quality of Service enhancements, IEEE draft standard P802.11e/D8.0, February 2004. [3] Aleksander Bai, Paal Engelstad, Bjørn Selvig, Tor Skeie, ’Interoperation between DiffServ and 802.11e EDCA’, Norwegian Network Research Seminar, 2005. (See also: http://www.unik.no/personer/paalee .) [4] B. Davie, A. Charny, J.C.R. Bennett , K. Benson, J.Y. Le Boudec, W. Courtney, S. Davari, V. Firoiu, D. Stiliadis, 'An Expedited Forwarding PHB (Per-Hop Behavior)', IETF RFC 3246, March 2002. [5] Yang Xiao, ‘Performance Analysis of IEEE 802.11e EDCF under Saturation Condition’, IEEE Communications Society, 2004. [6] Paal Engelstad, Olav Østerbø, ‘Non-Saturation and Saturation Analysis of IEEE 802.11e EDCA with Starvation Prediction’, ACM, Analysis and Simulation of Wireless and Mobile Systems, 2005. (See also: http://folk.uio.no/paalee .) [7] Sven Wietholter, Christian Hoene, 'Design and Verification of an IEEE 802.11e EDCF Simulation Model in ns-2.26', Technical University Berlin, Telecommunications Networks Group, Berlin November 2003. [8] Kevin Fall, Kannan Varadhan, http://www.isi.edu/nsnam/ns/doc/index.html. 'The ns Manual',