An Asymmetric Access Point for Solving the Unfairness Problem in ...

IEEE TRANSACTIONS ON MOBILE COMPUTING,

VOL. 7,

NO. 10,

OCTOBER 2008

1213

An Asymmetric Access Point for Solving the Unfairness Problem in WLANs Elena Lopez-Aguilera, Martin Heusse, Yan Grunenberger, Franck Rousseau, Andrzej Duda, and Jordi Casademont Abstract—In a typical deployment of IEEE 802.11 wireless LANs in the infrastructure mode, an access point acts as a bridge between the wireless and the wired parts of the network. Under the current IEEE 802.11 Distributed Coordination Function (DCF) access method, which provides equal channel access probability to all devices in a cell, the access point cannot relay all the frames that it receives on the downlink. This causes significant unfairness between upload and download connections, long delays, and frame losses. This unfairness problem comes from the not-so-complex interaction of transport-layer protocols with the MAC-layer access method. The main problem is that the access point requires more transmission attempt probability than wireless stations for correct operation at the transport layer. In this paper, we propose to solve the unfairness problem in a simple elegant way at the MAC layer. We define the operation of an Asymmetric Access Point that benefits from a sufficient transmission capacity with respect to wireless stations so that the overall performance improves. The proposed method of operation is intrinsically adaptive so that when the access point does not need the increased capacity, it is used by wireless stations. We validate the proposed access method by simulation to compare it with other solutions based on IEEE 802.11e. Unlike many papers in this domain, which only validate MAC-layer modifications through simulation or analytical modeling, we provide measurement data gathered on an experimental prototype that uses wireless cards implementing the proposed method. Index Terms—Local area networks access schemes, wireless local area networks, IEEE 802.11, prioritization, TCP performance over IEEE 802.11.

Ç 1

INTRODUCTION

W

E consider a performance problem that arises in IEEE 802.11 [1] wireless local area networks (WLANs) in the infrastructure mode. In a typical setup, an access point (AP) acts as a bridge for wireless stations: either it forwards frames between stations or it interconnects the wired and the wireless parts of the network by forwarding data flows in two directions (upload or download) on behalf of wireless stations. As most of the traffic goes through the AP, it requires a greater transmission attempt probability than wireless stations. However, this requirement is not currently supported by the standard IEEE 802.11 Distributed Coordination Function (DCF) access method that provides approximately equal or symmetric channel access probability to all devices in a wireless cell. Thus, if there are N wireless stations in a cell, the AP only benefits from 1=ðN þ 1Þ the channel access probability, which leads to an important performance degradation of connections transmitted over the wireless

. E. Lopez-Aguilera is with the Wireless Networks Group, Department of Telematics Engineering, Technical University of Catalonia (UPC), C4, Campus de Castelldefels, Av. Del Canal Olimpic, sn. 08860 Castelldefels, Spain. E-mail: [email protected]. . M. Heusse, Y. Grunenberger, F. Rousseau, and A. Duda are with the Grenoble Computer Science Laboratory (LIG), 38041 Grenoble, France. E-mail: {martin.heusse, yan.grunenberger, frank.rousseau, andrzej.duda}@imag.fr. . J. Casademont is with the Wireless Networks Group, Department of Telematics Engineering, Technical University of Catalonia (UPC), ModulC3 Campus Nord c/ Jordi Girona 1-3, 08034 Barcelona, Spain. E-mail: [email protected]. Manuscript received 19 Apr. 2007; revised 29 Oct. 2007; accepted 13 Feb. 2008; published online 7 Mar. 2008. For information on obtaining reprints of this article, please send e-mail to: [email protected], and reference IEEECS Log Number TMC-2007-04-0111. Digital Object Identifier no. 10.1109/TMC.2008.44. 1536-1233/08/$25.00 ß 2008 IEEE

link, known as the unfairness problem largely studied in the literature [2], [3], [4], [5], [6]. When real-time flows such as VoIP coexist with TCP connections, their delay may also become unacceptably long. The main cause of the unfairness problem is that the AP requires much more capacity than wireless stations to convey transport-level connections. Note that the unfairness problem addressed in this paper differs from other performance issues of the IEEE 802.11 DCF such as the performance anomaly due to the coexistence of multirate stations [7]. Many solutions to the unfairness problem have already been proposed at different layers: transport, network, and MAC. Proposals at upper layers try to fix the unfairness problem of the MAC layer and only provide partial solutions. MAC-layer solutions mainly make use of priority differentiation schemes such as the IEEE 802.11e, the recently published standard for Quality-of-Service (QoS) enhancement [8]. However, priority assignment is not straightforward, because it needs to adapt to changing load conditions in a wireless cell. In this paper, we propose to solve the unfairness problem in a simple elegant way at the MAC layer. We define the operation of an Asymmetric AP (AAP) that obtains a sufficient transmission capacity with respect to wireless stations so that the overall performance at the transport layer improves. The asymmetric operation consists of giving to the AAP transmission capacity that is k times greater than the capacity of all the wireless stations in the cell, i.e., the AAP benefits from kN times greater channel access probability compared to one wireless station, with N being the number of active stations in the cell. Recall that DCF results in equal or symmetric channel access probabilities of 1=ðN þ 1Þ for all Published by the IEEE CS, CASS, ComSoc, IES, & SPS

1214

IEEE TRANSACTIONS ON MOBILE COMPUTING,

contending entities, including the AP. Factor k reflects a common acknowledgment (ACK) scheme used by transport-layer protocols in which a receiver acknowledges every k data segments, with k being 2 in most of the current implementations. We show that this allocation is required in the worst case of the pure download scenario, so it leads to an increased overall throughput and good fairness. The AAP achieves the allocation by applying the following surprisingly simple principles: The AAP sets its contention window (CW) to a constant value, independent of the number of active wireless stations. 2. Wireless stations use the Idle Sense access method [9] and its CW adaptation mechanism. Similar to the IEEE 802.11 DCF access method, our solution is fully distributed and independent of N, the number of active wireless stations in the cell; i.e., the AAP does not need to know N to achieve the required allocation. The main advantage of our solution consists of faster draining of the frame queue at the AAP so that the throughput of download TCP connections is significantly improved—it becomes only limited by contention among stations and by their flow control. Our method privileges download connections (download means that data segments are sent to wireless stations) with respect to upload ones, because the current usage of wireless networks shows the dominance of download traffic [10]. This view is similar to the asymmetric capacity allocation in ADSL access networks. For upload connections, our scheme also results in an improved throughput compared to the DCF. Moreover, shorter queues at the AP mean that real-time UDP flows may benefit from shorter delays. In this paper, we derive the constant values of the CW that should be used by the AAP to obtain transmission capacity k times greater than all active stations for different variants of IEEE 802.11 PHY and MAC layers. We validate the analytical results through simulation and compare the performance of the AAP with a solution based on IEEE 802.11e priorities. Unlike many proposals for a modified wireless MAC layer, which are only evaluated via simulation or analytical modeling, we report measurement data of TCP and UDP performance gathered on an experimental prototype that uses wireless cards implementing the proposed method. This way, we provide a complete analysis of our solution. This paper is organized as follows: Section 2 states the problem and overviews the related work. Section 3 introduces the principles of the IEEE 802.11 DCF and Idle Sense access methods. Section 4 presents the operation of the AAP, our solution to the unfairness problem. In Section 5, we evaluate the throughput that can be achieved by the AAP. Section 6 presents performance validation of our proposal through simulation and comparison with a solution based on the IEEE 802.11e. Section 7 shows measurement data gathered on an experimental prototype that uses wireless cards implementing the proposed method. Section 8 concludes this paper. 1.

2

PROBLEM STATEMENT

AND

RELATED WORK

Let us go into the details of the unfairness problem. In the current Internet architecture, IEEE 802.11 WLANs are

VOL. 7,

NO. 10,

OCTOBER 2008

Fig. 1. Basic configuration studied in this paper.

mostly access networks connecting wireless stations to the high-capacity wired backbone (cf. Fig. 1). In most configurations, they are bottleneck links, because their bit rates are smaller than largely deployed Ethernet local area networks (LANs) and long haul fiber links between routers. The second characteristic of IEEE 802.11 WLANs is their halfduplex operation: all wireless stations and the AP need to share the capacity of the radio channel. We believe that these characteristics should be taken into account in the design of the access method for the wireless link and in deciding how the capacity of the wireless link can be distributed among the AP and wireless stations. Recall that the standard IEEE 802.11 DCF access method works on a per-frame basis: it gives to each contending entity (a station or an AP) approximately equal transmission opportunity, regardless of the length of a data frame. We consider two basic cases to illustrate the unfairness problem: 1) local communication between wireless stations, with the AP acting as a bridge, and 2) wireless stations communicate with hosts on the Internet via the AP. Consider the first case: the AP is used as a bridge for exchanging local two-way traffic between wireless stations—it forwards frames received from source stations to respective destination stations. Each response frame goes on the same path back from the destination to the source. We can easily see that during one frame exchange, the AP sends two frames, whereas each station sends one frame. It is thus clear that for efficient forwarding, the AP requires more transmission opportunity over the channel than wireless stations: in this scenario, the AP needs to have twice the transmission capacity of a wireless station. Otherwise, it becomes a bottleneck that limits the throughput and leads to lost frames due to buffer overflow. The second case is more complex. Let us consider a scenario in which wireless stations communicate with hosts on the Internet through download or upload transport connections that go across the AP. Download means that the AP forwards data segments to a station that sends transport-layer ACKs in the opposite direction. In an upload connection, the directions of data and ACKs flows are inversed. Without loss of generality, we can assume that connections are TCP, the most used transport protocol in the Internet (it accounts for more than 85 percent of the Internet traffic). Other standardized transport protocols such as Datagram Congestion Control Protocol (DCCP) [11] and Stream Control Transmission Protocol (SCTP) [12] adopt a similar error control mechanism. Recall that TCP provides several communication functions such as reliable data transfer through the use of the Automatic Repeat Request (ARQ) error control based on ACK segments and flow and congestion control.

LOPEZ-AGUILERA ET AL.: AN ASYMMETRIC ACCESS POINT FOR SOLVING THE UNFAIRNESS PROBLEM IN WLANS

1215

Fig. 2. Upload scenario.

Fig. 4. Download scenario.

The error control uses one ACK per k data segments, with k being 2 for most implementations of transport protocols. We consider a cell with N stations: M stations have active TCP upload connections, and N  M stations have TCP download connections (one connection per station). Pure uploads ðM ¼ NÞ or downloads ðM ¼ 0Þ are extreme cases that set up performance limits for such mixed scenarios. Figs. 2 and 4 illustrate them and show how many segments are backlogged on the average. In the pure upload case ðM ¼ NÞ, stations have backlogged segments to be sent, and they participate in the contention for the channel access, while the AP needs to send ACKs for each connection. As the AP working under IEEE 802.11 DCF has the same transmission opportunity as the other stations, some frames to be transmitted over the wireless link may be lost due to the limited size of buffers at the AP. For the upload connections, lost ACKs sent by the AP toward wireless stations do not raise much problems, because TCP ACKs are cumulative (an ACK acknowledges all previous data segments), so the upload throughput will only be marginally affected by the TCP congestion control mechanisms: to keep uploading sources active, it is sufficient to occasionally receive an ACK. Fig. 3 presents the measured performance of an upload oriented scenario: one download competing with four uploads (M ¼ 4, and N ¼ 5). We can observe the starvation of the download traffic under the standard IEEE 802.11 DCF that allocates to the AP the proportion of transmission attempts of 1=ðN þ 1Þ ¼ 1=6. If we assume that the destination generates one ACK per k ¼ 2 data segments, the performance of uploads may be good if the AP benefits from the proportion of transmission attempts equal to

N=ðN þ 2NÞ ¼ 1=3, whereas all other stations need to benefit from 2N=ðN þ 2NÞ ¼ 2=3 (cf. Fig. 2). For the pure download case ðM ¼ 0Þ, the situation is different, with data segments backlogged at the AP (cf. Fig. 4). Because of an insufficient capacity share under the IEEE 802.11 DCF, the segments of the download connection fill up the AP buffer and are dropped, unless it is large enough for the TCP sources to choke on the window size [3]. This leads to important performance degradation: long delays, lost data segments, and retransmissions. Fig. 5 illustrates this problem in a download-oriented scenario: one upload competing with four downloads (M ¼ 1, and N ¼ 5). We can observe that the uploading station obtains a far better throughput than the downloading stations, because TCP segments directed to wireless stations queue at the AP, while the uploading station has the same channel access share as the AP ð1=ðN þ 1Þ ¼ 1=6Þ and thus sends its segments at a greater rate. However, the performance of downloads may be good if the AP benefits from the proportion of transmission attempts equal to 2N=ðN þ 2NÞ ¼ 2=3, whereas all other stations require only N=ðN þ 2NÞ ¼ 1=3 (cf. Fig. 4). With mixed upload-download scenarios, the proportion of transmission attempts of the AP needs to vary between 1/3 and 2/3. In general, for a protocol with one ACK per k data segments, the AP requires at most the proportion of transmission attempts equal to k=ðk þ 1Þ. For mixed upload-download scenarios, this proportion may vary between 1=ðk þ 1Þ and k=ðk þ 1Þ, whereas with the IEEE 802.11 DCF, the share of the AP is 1=ðN þ 1Þ.

Fig. 3. Data transferred versus elapsed time, with one TCP download and four uploads for the IEEE 802.11 DCF. Experimental measurements.

Fig. 5. Data transferred versus elapsed time, with four TCP downloads and one upload for the IEEE 802.11 DCF. Experimental measurements.

1216

IEEE TRANSACTIONS ON MOBILE COMPUTING,

As we can see, the unfairness problem comes from an insufficient capacity of the AP to convey transport-level connections. In a IEEE 802.11 cell with N active stations, there are N þ 1 contending entities. However, if the transport layer ACK scheme makes use of one ACK per k data segments and we consider the worst case, i.e., the pure download (the AP sends data segments to all stations that need to send one ACK per k data segments), there should be ðk þ 1ÞN virtual contending entities, with kN entities associated with the AP and N entities associated with stations. Thus, the AP requires transmission capacity k times greater than the capacity of all stations (or kN times the capacity of one station) to empty its buffer of data segments. The objective of our proposal is to provide a solution at the MAC layer, which potentially gives the AP sufficient capacity to correctly operate in the worst case: kN times the capacity of a single station. Moreover, we want to allocate the capacity in a way that is independent of N, the number of active stations, to avoid its estimation, which usually requires some complex algorithms [13]. In other mixed upload-download scenarios or when the traffic is a mixture between other types of flows, the AP will simply not use the available capacity, leaving it to active stations. In the case of multiple uploads, a greater potential capacity of the AP will not affect the overall performance, because the limiting factor is contention among stations. Such a principle represents an analogy with ADSL access networks in which it is common to allocate capacity at layer 2 according to the actual use of the link. ADSL links provide asymmetric capacity, with the downlink exceeding the uplink at about one order of magnitude, e.g., 24 megabits per second (Mbps) versus 1 Mbps in ADSL2+. Similarly, the AP in IEEE 802.11 WLANs needs to have capacity proportional to the number of segments that active wireless stations may receive.

2.1 Related Work The performance problem raised by an IEEE 802.11 AP that does not benefit from enough radio-channel share has been largely studied in the literature [3], [4], [5], [6], [14], [15], [16], [17] and has led to many analysis of the TCP behavior in wireless LANs [2], [3], [15], [16], [17], [18]. It results in severe throughput unfairness observed at the level of TCP connections; for instance, the throughput of a download connection may be 10 times lower than that of an upload one [3]. The problem is important, because the mixed uploaddownload scenario is common in a typical deployment of wireless LANs in the infrastructure mode [19]. A similar unfairness has also been observed for UDP flows [15]. We can categorize different solutions proposed to solve the problem according to the layer at which they are placed: 1.

Transport. Several solutions have been proposed to address the problems of TCP over wireless networks. The early work has focused on coping with higher frame error rates over wireless links that reduce TCP performance: TCP Snoop [20], [21] and Indirect TCP [22]. Balakrishnan and Padmanabhan discuss many TCP performance issues related to network asymmetry [2]. One of the problems is the ACK compression caused by queuing in the reverse path of a TCP connection [23], [24]. It results in performance

VOL. 7,

NO. 10,

OCTOBER 2008

degradation of TCP over wireless links. The TCP New Jersey with Time-stamp-based Available Bandwidth Estimation (TABE) [25] is immune to ACK compression. To deal with the unfairness problem of TCP connections, Pilosof et al. proposed to modify the receiver window in TCP ACKs to pace sources on wireless stations and provide, this way, more bandwidth for the download traffic [3]. 2. Network. Some authors proposed to solve the unfairness problem by providing a suitable scheduling mechanism at the IP layer. Eckhardt and Steenkiste [26] defined an Effort-Limited Fair scheduling for wireless networks. Other authors proposed to use QoS support or service differentiation to cope with performance problems over WLANs [27], [28]. Ha and Choi address the unfairness problem with two distinct queues for data segments and ACKs at the AP [29]. They tune their relative priorities according to the number of flows in both directions and the corresponding offered window field in ACK segments. 3. MAC. Many authors propose to solve the unfairness problem by defining an adequate access method. Vaidya et al. defined a fully distributed mechanism for fair scheduling in a WLAN that consists of allocating bandwidth in proportion to the weights of flows sharing the radio channel [16]. Their mechanism trades weighted fairness for an additional wait time before each frame transmission. Banchs and Perez proposed to modify the IEEE 802.11 DCF so that CWs are scaled according to a distributed weighted fair queuing algorithm [17]. Siris and Stamatakis defined a centralized mechanism for assigning different CW values to stations in proportion to their weights [30]. Shin and Shulzrinne proposed to increase the AP priority through the transmission of a number of frames in a ContentionFree Period (CFP) [31]. The AP determines the required number of frames based on its queue size and the number of wireless stations. Bruno et al. defined a method in which the AP can periodically transmit frame bursts [32]. The AP adapts the burst length to the estimated collision probability of uplink transmissions. Another approach was to use MAC-layer priorities such as those offered by the IEEE 802.11e [8]. Leith et al. choose suitable parameters of the IEEE 802.11e to provide fairness between competing TCP uploads and downloads [5], [6]. Other authors propose algorithms for enhancing performance in asymmetric traffic load conditions by giving more priority to the AP. However, they rely on exchanging information between the AP and wireless stations [4], [15], [33]. To conclude the description of the related work, we can observe that solutions above the MAC layer are not really suitable, because they do not eliminate the cause of the problem—the AP obtains less capacity than what is required for forwarding flows. Moreover, they are usually much complex, because they require processing at the network or transport layers. So, it seems more promising to try to solve it by an appropriate solution at the MAC layer.

LOPEZ-AGUILERA ET AL.: AN ASYMMETRIC ACCESS POINT FOR SOLVING THE UNFAIRNESS PROBLEM IN WLANS

1217

The advantage of such an approach is that it will work for any kind of upper protocols (TCP, UDP, DCCP, SCTP, or even multicast traffic). Concerning the MAC layer, many authors have proposed to solve the unfairness problem by defining an adequate access method that supports differentiation between stations [16], [17], [31], [32]. However, setting up differentiation parameters is not a simple task, because the priority given to the AP needs to adapt to the current load of a cell and the number of active stations. Actually, what we need is an access method that dynamically allocates the required capacity of the AP with respect to other stations and their current traffic.

The original Idle Sense algorithm [9] makes ni converge to ntarget by applying Additive Increase Multiplicative Decrease i (AIMD) [34] to transmission attempt probability Pe . When stations do not perform the exponential backoff mechanism after collision, the transmission attempt probability is given as follows [35]:

3

, Pe Pe þ . . If ni  ntarget i , P Pe . . If ni < ntarget e i Here,  and  are some adaptation parameters. In practice, stations update their CWs by combining these updating rules with (1). However, our experience with the implementation and tuning of Idle Sense [36] resulted in some modifications of the basic algorithm. We have observed that applying AIMD to Pe is not the right thing to do, in part because it requires a ceiling function to prevent Pe from becoming greater than 1. Instead, we propose to directly apply AIMD to the CW, which yields the following algorithm:

IEEE 802.11 DCF METHODS

AND IDLE

SENSE ACCESS

We have shown in the previous section that to solve the unfairness problem in WLANs, we need to allocate sufficient capacity to the AP compared to wireless stations. How can we achieve the required capacity allocation in a simple way, adaptive to the load in a cell and independent of the actual number of wireless stations? We propose to implement the proposed allocation principles by using the Idle Sense method [9], which we describe briefly in the following. For completeness, we also shortly recall the IEEE 802.11 DCF access method. The IEEE 802.11 DCF uses the Carrier Sense Multiple Access with Collision Avoidance (CSMA/CA) access method: before initiating a transmission, a station senses the state of the channel. If the channel is free, the station transmits its frame after the DCF Interframe Space (DIFS) interval. When the medium is sensed busy, the station defers until the channel is free and waits for the DIFS interval and a random backoff: an additional contention time composed of a number of slots uniformly distributed in [0, CW  1]. If the station senses a transmission during the contention time, it uses the residual backoff for the next contention period. The value of CW is set to CWmin for the first transmission attempt, and it is doubled at each failed transmission (collision or frame loss) up to CWmax (the exponential backoff mechanism). Idle Sense optimizes the IEEE 802.11 DCF for high throughput and fairness: contending stations do not use the exponential backoff mechanism after collisions and failed transmissions; they rather make their CWs dynamically converge in a fully distributed way to similar values solely by tracking the number of idle slots between consecutive transmissions. The method works as follows: Each station estimates the number of consecutive idle slots ni between two transmission attempts and uses this estimate n^i to adjust its CW to approach the predefined target value ntarget . ntarget is computed numerically for a i i given variant of the IEEE 802.11 from the parameters of the PHY and MAC layers (its value is 5.68 for the IEEE 802.11b and 3.91 for the IEEE 802.11a/g [9]). When stations adjust their CW so that ni converges to ntarget , their throughput is i close to the optimal. After the initial convergence, their values of CW are proportional to the number of contending entities.

Pe ¼

2 : CW þ 1

ð1Þ

If a station observes too many idle slots compared to the target, it needs to increase Pe additively, which, in turn, will decrease ni , whereas if it observes too few idle slots, it needs to decrease Pe in a multiplicative way, which, in turn, will increase ni . This leads to the following algorithm:

.

If ni  ntarget , CW i

CW .

ntarget , i

CW CW þ . . If ni < We have also observed that the accuracy of the algorithm improves if we adjust the parameter maxtrans [9] so that it becomes proportional to the number of stations (recall that a station refreshes CW every maxtrans transmissions on the channel). We use the fact that Idle Sense results in a CW value proportional to the number of contending stations. Then, to speed up convergence when n^i is clearly off target, we use a small value of maxtrans when ni is significantly different from the target value (a value of 5 [9] seems to be the right value). The refined adaptation mechanism is thus the following [36]: <  ! maxtrans ¼ CW . . If ni  ntarget i  . If ni  ntarget   ! maxtrans ¼ 5. i

Here,  and  are some constants. The first condition makes maxtrans a multiple of the number of contending stations when ni is close to the target value ntarget . Based on the i properties of Idle Sense [9], this leads to the following values: .

For IEEE 802.11b, maxtrans ¼ CW   3N.

. For IEEE 802.11a/g, maxtrans ¼ CW   2N. Algorithm 1 [36] specifies Idle Sense more formally. Algorithm 1: Idle Sense. maxtrans 5; sum 0; ntrans 0 After each transmission { /* Station observes ni idle slots before a transmission */ sum sum þ ni ntrans ntrans þ 1 if ðntrans >¼ maxtransÞ then

1218

IEEE TRANSACTIONS ON MOBILE COMPUTING,

/* Compute the estimator */ n^i sum=ntrans /* Reset variables */ sum 0 ntrans 0 Þ then if ðn^i < ntarget i /* Increase CW */ CW CW þ  else /* Decrease CW */ CW CW end if /* Adapt update interval */ ^  n if ð ntarget i < Þ then i maxtrans ¼ CW  else maxtrans ¼ 5 end if end if } After several tests and validation through simulation, we have chosen the following values of the adaptation parameters that lead to the best balance between accuracy and convergence speed for any number of contending stations: 1  ¼ 1:0666.  ¼ 6:0.  ¼ 0:75.  ¼ 4.

. . . .

4

ASYMMETRIC ACCESS POINT

Our objective is to define a solution at the MAC layer, which potentially gives the AP sufficient capacity to correctly operate in all traffic conditions while wireless stations dynamically share the rest of the capacity in a fair manner. We can achieve this by the following operation of the WLAN cell: The AAP sets its CW to a constant value. Wireless stations use the Idle Sense access method and its CW adaptation mechanism. We assume that there are N active wireless stations in the cell (active means that they have frames to be sent over the channel). Note that neither the AAP nor wireless stations need to know the value of N. As explained in Section 2, we also assume that there are N upload or download connections at different stations (one connection per station) and the transport layer ACK scheme makes use of one ACK per k data segments. Our approach for solving the TCP unfairness problem is to provide the AP with a sufficient probability of transmission to drain its frame queue. The smallest probability for achieving this for any mix of download/upload connections is the transmission probability required in the pure download scenario: the AAP needs to obtain kN times the capacity of a single station. This way, we provide a solution for the case in which the AP suffers the most from insufficient capacity. This also means that our scheme will privilege download connections with respect to upload ones, which is 1. 2.

VOL. 7,

NO. 10,

OCTOBER 2008

reasonable in view of dominating download traffic [10]. In the case of pure uploads, this greater potential capacity will not be effectively used by the AP: it will empty the queue of ACK segments, and the limiting factor will be the contention among stations that try to obtain as much capacity as possible. As the queue at stations will not grow up because of the self-pacing behavior of TCP, stations will adapt themselves to the available capacity. The Idle Sense method used by stations leads to more efficient channel usage than under DCF, especially for a larger number of active stations. Thus, we need to have the following relation between the probabilities of a successful transmission: PtAAP ¼ kNPtSTA :

ð2Þ

Consequently, if all contending entities send frames of the same size and at the same transmission rate, the corresponding throughputs satisfy the following relation: XAAP ¼ kNXSTA :

ð3Þ

We show in the following that this relation, combined with the Idle Sense mechanism used at stations, leads to a fixed value of CW at the AAP, and we derive its value. We start from (2) to find the corresponding transmission attempt probabilities PeSTA and PeAAP that are directly controlled by the CW (cf. (1)). We follow the model in [35] to express the probabilities of a successful transmission: PtAAP ¼ PeAAP 1  PeSTA


N

;

ð4Þ

and PtSTA ¼ PeSTA 1  PeSTA


N1


1  PeAAP :

ð5Þ

Combining them with (2) yields PeAAP PeSTA ¼ kN  kNPeSTA : 1  PeAAP 1  PeSTA

ð6Þ

Then, we can express the probability of an idle slot, which is the same for the AP and all stations, as
AAP 1  Pe 
N 1  1  PeSTA : kNPeSTA þ 1

Pi ¼ 1  PeSTA


N

ð7Þ

Let us consider now how the principles of Idle Sense can be applied to the AAP. Idle Sense makes the CW of stations converge to the values close to the optimal for any number of contending entities. When all contending entities use Idle Sense, the number of idle slots between two transmissions is kept constant so that the probability of an idle slot Pi also remains constant. Thus, Idle Sense maintains NPe ¼ , with  being some constant. When N ! 1, the probability of an idle slot becomes Pitarget  e [9]. In the case of the AAP, only wireless stations use Idle Sense. The probability of an idle slot remains Pitarget  e , and the access probability of wireless stations PeSTA is such that NPeSTA ¼   , where   is another constant. For N ! 1, we thus obtain 1  PeSTA


N



! e :

ð8Þ

LOPEZ-AGUILERA ET AL.: AN ASYMMETRIC ACCESS POINT FOR SOLVING THE UNFAIRNESS PROBLEM IN WLANS

Using this in (7) leads to Pi  e







1 : k  þ 1

Thus,  ¼   þ lnð1 þ k  Þ:

ð9Þ

We can numerically solve this equation to find the value of constant   corresponding to a given value of Pitarget (the target value of Pi and, consequently,  are derived elsewhere [9]). Then, from (6), we have PeAAP ¼

k  : 1 þ k 

CWAAP ¼

PeAAP

5

 1:

This result is important, because it allows us to set the CW of the AAP to a constant value, independent of the number of active stations in a cell, so that the AAP obtains kN times the throughput of each wireless station. We obtain the following values of CWAAP for different variants of IEEE 802.11 PHY and MAC layers: .

IEEE 802.11b

.

– CWAAP ¼ 25:18 for k ¼ 1. – CWAAP ¼ 18:86 for k ¼ 2. IEEE 802.11a/g

– CWAAP ¼ 18:09 for k ¼ 1. – CWAAP ¼ 13:54 for k ¼ 2. Obviously, we need to use integer values in the access method. We propose to round the derived values to the smallest integer to give slightly more weight to the AAP so that its queue does not build up. Thus, we fix 25 and 18 for the IEEE 802.11b and fix 18 and 13 for the IEEE 802.11a/g. Finally, we want to derive the target value of the CW for stations that use the Idle Sense access method. If, as stated above, NPeSTA ¼   , then PeSTA

 ¼ ; N

ð11Þ

and from (1), we obtain CWSTA ¼

use the same frame size, the throughput allocation also follows the same distribution (cf. (3)). The value of CWAAP is a local parameter of a given AP and can be set up by an administrator based on the information of the local TCP variants (value of k) and a given traffic mix (expected proportion of uploads and downloads). We think that equalizing the throughputs for any mixture of uploads and downloads for any value of k cannot be achieved at the MAC layer. It is still an open problem that requires a more complex mechanism, probably based on a cross-layer approach (need for TCP-layer information) and belongs to our future work.

ð10Þ

Finally, we can derive the CW of the AAP from (1): 2

1219

ANALYSIS

OF

THROUGHPUT

In this section, we want to evaluate the throughput at the MAC layer in a cell, with the AAP operating as described above but with stations that use the fixed value of the CW derived in the previous section; i.e., they do not rely on the adaptation mechanism of Idle Sense. So, we assume that all stations use the fixed CWSTA given in (12) and we know the number of stations. These assumptions correspond to an access method with a perfect adaptation mechanism at stations, so we will be able to compare its throughput, which we call target, with the throughput achieved when stations dynamically adapt their CWs according to Idle Sense. We start again with Bianchi’s model [35], which analyses the saturation throughput of the IEEE 802.11 DCF. Thus, we consider that all N stations and the AP are greedy—we assume that the frame buffers of all contending entities are always backlogged. Consider the event consisting of a station attempting transmission in a given time slot. The transmission attempt probabilities of AAP and STA are given in (10) and (11), respectively. The probability of an idle slot is thus

N Pi ¼ 1  PeAAP 1  PeSTA :

ð13Þ

The probability of a successful transmission in a given slot Pt can be expressed as Pt ¼ PeAAP 1  PeSTA


N



N1 þNPeSTA 1  PeAAP 1  PeSTA : ð14Þ

The collision probability in a given slot Pc is thus 2N  1: 

ð12Þ

To summarize, the AP in our solution operates with a fixed value of the CW, which depends on a given variant of IEEE 802.11 PHY and MAC layers and the value of k. Wireless stations use Idle Sense to dynamically adjust their CW to the target value (cf. (12)). Note that stations do not need to know this value nor the number of contending stations due to the adaptation algorithm of Idle Sense that takes care of this convergence. Moreover, we remark that stations operate, independent of factor k, because only CWAAP depends on k. Under these conditions, AAP benefits from the successful transmission probability increased by the factor of kN compared to a single station. If all entities

Pc ¼ 1  ðPi þ Pt Þ:

ð15Þ

We can express the system throughput X as a function of the probabilities defined above: X¼

Pt sd ; Pi  þ Pc Tc þ Pt Tt

ð16Þ

where sd is the average frame size,  is the duration of an idle slot time (defined by a given variant of IEEE 802.11 PHY and MAC layers), Tc is the average collision duration, and Tt is the time of a successful transmission. Tc and Tt depend on the parameters of the PHY and MAC layers, on TDATA , the transmission time of data frames, on TACK , the transmission time of ACK frames, and on the propagation

1220

IEEE TRANSACTIONS ON MOBILE COMPUTING,

TABLE 1 Target Values of Throughput (in Megabits per Second) and CW for the IEEE 802.11b, with k ¼ 1

VOL. 7,

NO. 10,

OCTOBER 2008

TABLE 3 Target Values of Throughput (in Megabits per Second) and CW for IEEE 802.11b, with k ¼ 2

Analytical results.

Analytical results.

TABLE 4 Target Values of Throughput (in Megabits per Second) and CW for IEEE 802.11a/g, with k ¼ 2

TABLE 2 Target Values of Throughput (in Megabits per Second) and CW for the IEEE 802.11a/g, with k ¼ 1

Analytical results.

Analytical results.

TABLE 5 PHY and MAC Parameters

time . If we consider data frames of a fixed size, we have the following:1 Tt ¼ TDATA þ SIF S þ  þ TACK þ DIF S þ ; Tc ¼ TDATA þ DIF S þ :

ð17Þ ð18Þ

Finally, we obtain the following expressions for XAAP , the target throughput of the AP, and XSTA , the target throughput of a station:

PtAAP

XAAP ¼

PtAAP sd ; Pi  þ Pc Tc þ Pt Tt

ð19Þ

XSTA ¼

PtSTA sd ; Pi  þ Pc Tc þ Pt Tt

ð20Þ

PtSTA

and are given in (4) and (5), respectively. where Tables 1 and 2 present the target throughput2 of our proposed access method for data frames of 1,500 bytes and k ¼ 1 (the PHY and MAC parameters of the IEEE 802.11b and IEEE 802.11a/g are given in Table 5). Thus, when the AAP and stations use the fixed values of CW derived in Section 4, they benefit from the throughput presented in the tables for a given number of active stations N. We can see that the AAP benefits from throughput values almost equal to the throughput obtained by all the wireless stations together, which means that the AAP obtains a throughput share around 0.50. Tables 3 and 4 present the target throughput for k ¼ 2. In this case, the AAP benefits from a double throughput of all the wireless stations together; i.e., the AAP obtains a throughput share around 0.66. Thus, the AAP obtains kN times the throughput of each wireless station and a throughput share of k=ðk þ 1Þ (cf. Section 2). We can observe 1. We do not consider EIFS in this analysis. 2. We use the exact and not rounded values of the contention window, because the target throughput is computed analytically.

that the throughput share becomes closer to k=ðk þ 1Þ for greater N.

6

PERFORMANCE COMPARISONS BASED SIMULATION

ON

In this section, we compare the performance of our solution with the standard IEEE 802.11 DCF and with another proposed approach based on the IEEE 802.11e. The results come from simulation: we use a self-made discrete-event simulator that accurately models the PHY and MAC layers of the original IEEE 802.11a/b/g and Idle Sense access methods. In the simulation of the AAP, wireless stations behave according to the Idle Sense adaptation algorithm so that they try to keep their CWs around the target values derived in Section 4. The simulator has been extensively validated and used in published papers [37], [38]. We also favorably compared its results with the MAC simulator used to evaluate the Idle Sense access method [9]. To obtain our simulation results, we have run a large number of independent simulations and obtained very small confidence intervals so that they are not shown in the following figures. We present the results for IEEE 802.11g and IEEE 802.11b PHY and MAC layers (cf. parameters in Table 5) and for data frames of 1,500 bytes. In this section, we only evaluate the performance at the MAC layer and the ability of the proposed solution to

LOPEZ-AGUILERA ET AL.: AN ASYMMETRIC ACCESS POINT FOR SOLVING THE UNFAIRNESS PROBLEM IN WLANS

1221

Fig. 6. AAP throughput versus the number of stations for the IEEE 802.11g. Simulation results.

Fig. 8. Throughput per station versus the number of stations for the IEEE 802.11g. Simulation results.

Fig. 7. AAP throughput versus the number of stations for the IEEE 802.11b. Simulation results.

Fig. 9. Throughput per station versus the number of stations for the IEEE 802.11b. Simulation results.

correctly distribute required throughput shares between the AP and wireless stations. So, we consider that all N stations and the AP are greedy. Such a situation arises, for example, when each station has a UDP upload flow and there are N downloads passing through the AP and going to wireless stations so that the frame buffers of all contending entities are always backlogged. In this case, the throughput of each flow exceeds the share of radio channel capacity that a station or the AP can obtain. We consider the performance of TCP connections over a WLAN in the next section, which presents data measured on an experimental platform.

To better understand the behavior of the proposed solution, we have gathered statistics during simulations on the number of idle slots and on the CW of wireless stations (cf. Figs. 10 and 11, respectively).3 As wireless stations operate according to the Idle Sense access method, they do not perform the exponential backoff algorithm unlike the standard IEEE 802.11 DCF, which makes ni to be close to the target values and consequently results in CW of stations close to the target values (cf. Section 3). Finally, we have evaluated the convergence speed in the following scenario. At the beginning, there are four stations and AAP contending for the channel. After 2,000 transmissions, when the CW of stations has already converged for a long time to the target value, six more stations become active, so now, 10 stations are contending for the channel. After 3,000 more transmissions, six stations become inactive again. Fig. 12 presents the CW of stations as a function of the time represented as the number of transmissions. We can observe based on this figure that the AIMD adaptation mechanism applied to CW (cf. Section 3) quickly reacts to changing conditions: only few periods of idle slot estimation are required to reach a new stable state.

6.1 Comparisons with the Target Throughput First, we evaluate the influence of the adaptation mechanism of Idle Sense used in our solution based on the AAP by comparing the simulated results with the analytical values of the target throughput derived in Section 5. Figs. 6 and 7 present the throughput of the AAP for IEEE 802.11g and IEEE 802.11b PHY and MAC layers, respectively, whereas Figs. 8 and 9 give the corresponding results for wireless stations. The results show that the AAP obtains the throughput close to the target values. We can also see that the solution based on the AAP attains its performance objective: the AP obtains the throughput share close to k=ðk þ 1Þ.

3. We only include the results for the IEEE 802.11g PHY and MAC layers, because the corresponding results for the IEEE 802.11b are quite similar.

1222

IEEE TRANSACTIONS ON MOBILE COMPUTING,

VOL. 7,

NO. 10,

OCTOBER 2008

Fig. 10. Number of idle slots versus the number of stations for the IEEE 802.11g. Simulation results.

Fig. 12. Convergence speed with AAP for the IEEE 802.11g. Simulation results.

6.2

categories (ACs), with AC 4 having the highest priority. Each AC uses parameters that depend on AC: interframe intervals AIF S½AC instead of standard DIFS and CWs CWmin ½AC and CWmax ½AC instead of the standard CWmin and CWmax . Table 7 gives the default values of these parameters for aCWmin and aCWmax corresponding to the values specified for the IEEE 802.11g: aCWmin ¼ 16, and aCWmax ¼ 1;024. Some authors have proposed to use the IEEE 802.11e to solve the unfairness problem [5], [6], so we wanted to compare our approach with the solution in which the AP benefits from a higher priority than wireless stations. As in

Comparisons with the IEEE 802.11 DCF and Idle Sense Access Methods Next, we have compared the performance of our proposed solution based on the AAP with the standard IEEE 802.11 DCF and Idle Sense. Recall that both access methods consider the AP as any other contending entity, so it does not obtain more throughput than a wireless station. Table 6 shows the throughput comparison with the IEEE 802.11 DCF and Idle Sense. We can observe that our solution achieves the desired throughput distribution between the AP and stations of k=ðk þ 1Þ fairly well. Nevertheless, the throughput share of AAP changes slightly with the number of stations: it is greater than k=ðk þ 1Þ for few stations and less than k=ðk þ 1Þ for N around 10 or greater. This effect is due to the existence of the adaptation mechanism in Idle Sense, which makes the throughput of stations oscillating around the target value. We can also see that our solution leads to higher bandwidth efficiency by obtaining better aggregate throughput X.

TABLE 6 Throughput Comparison (in Megabits per Second) for the IEEE 802.11g

6.3 Comparison with the IEEE 802.11e Recently, the standardized IEEE 802.11e [8] has defined mechanisms for channel access differentiation based on the Enhanced Distributed Channel Access (EDCA) method. EDCA offers eight traffic priorities mapped into four default access

Fig. 11. CW versus the number of stations for the IEEE 802.11g. Simulation results.

Simulation results.

LOPEZ-AGUILERA ET AL.: AN ASYMMETRIC ACCESS POINT FOR SOLVING THE UNFAIRNESS PROBLEM IN WLANS

1223

TABLE 7 Default EDCA Parameters

TABLE 8 Assignment of IEEE 802.11e Parameters for Each Case

Fig. 14. AP throughput versus the number of stations for IEEE 802.11e priorities versus the AAP, with k ¼ 2. Simulation results.

Fig. 15. Throughput per station versus the number of stations for IEEE 802.11e priorities versus the AAP, with k ¼ 1. Simulation results.

Fig. 13. AP throughput versus the number of stations for IEEE 802.11e priorities versus the AAP, with k ¼ 1. Simulation results.

previous simulations, we consider a cell with N greedy stations and the AP with N download flows. All stations transmit data frames of 1,500 bytes at the nominal bit rate of 54 Mbps. We have chosen seven different assignments of AIF S, CWmin , and CWmax to the AP and wireless stations (cf. Table 8). In the first four cases, the AP has a higher priority than the default assignment (cf. Table 7), whereas in the last three cases, we give it a default priority but is still higher than the priority of wireless stations. Figs. 13, 14, 15, 16, 17, and 18 present the throughput of the AP, the throughput of wireless stations, and the aggregate throughput, respectively. The results show that it is not easy to find the right values for priority differentiation parameters so that the AP obtains the required throughput share of k=ðk þ 1Þ, with wireless stations obtaining 1=ðk þ 1Þ. In Cases 1, 2, and 6, the throughput of the AP and the aggregate throughput are high, but it is obtained at the expense of fairness: the

Fig. 16. Throughput per station versus the number of stations for IEEE 802.11e priorities versus the AAP, with k ¼ 2. Simulation results.

throughput of wireless stations is zero. Such an effect has already been shown: under heavy load of high-priority traffic, EDCA leads to the starvation of low-priority traffic [39]. As the different AIFSs are separated by values that are multiples of the slot time and the backoff interval is slotted, queues of different priorities work in a synchronized way under heavy load: they attempt transmissions simultaneously, which results in collisions between high-priority frames transmitted after a backoff interval and lower

1224

IEEE TRANSACTIONS ON MOBILE COMPUTING,

Fig. 17. Aggregate throughput versus the number of stations for IEEE 802.11e priorities versus the AAP, with k ¼ 1. Simulation results.

priority frames transmitted without any backoff time. Such synchronization leads to the starvation of low-priority traffic. Cases 3, 4, 5, and 7 result in the AP throughput closer to the target. Wireless stations also achieve the throughput closer to the target for Cases 5 and 7. For Cases 3 and 4, the differences in station throughput with respect to the target value are significant. The comparisons show that all the cases of the IEEE 802.11e present lower aggregate throughput compared to the target values (only Cases 1, 2, and 6 show better aggregate throughput but at the expense of fairness) and our solution based on the AAP offers the best overall performances. Besides, even if we know the number of stations and traffic sources, it is not easy to adequately assign the required IEEE 802.11e priorities to the AP and wireless stations. It may even become more difficult with a number of stations and traffic load changing fast, probably requiring a coordinator and a signaling protocol for dynamically adapting priority assignments. The IEEE 802.11e EDCA can be used at the AP to fix its CW by specifying the same value for both CWmin and CWmax (we did it in an early implementation of the AAP to fix the CW). However, this approach is limited, because EDCA only allows us to set the value of the CW to a power of 2. Besides EDCA, the IEEE 802.11e standard also defines the HCF Controlled Channel Access (HCCA) polling access method based on the Hybrid Coordination Function (HCF), which extends the IEEE 802.11 PCF. It can be used to give more channel access share to the AP; however, the scheme for channel allocation still needs to be designed.

7

IMPLEMENTATION MEASUREMENTS

AND

EXPERIMENTAL

We have implemented Idle Sense on Intel PRO/Wireless 2200BG 802.11 a/b/g cards and set up an experimental testbed on FreeBSD boxes. The AAP uses the same card, with the CW fixed to the value derived in Section 4. In addition to the AAP, the testbed includes five wireless stations, with one TCP connection each (either download or upload). All traffic on the wired part involves a host two

VOL. 7,

NO. 10,

OCTOBER 2008

Fig. 18. Aggregate throughput versus the number of stations for IEEE 802.11e priorities versus the AAP, with k ¼ 2. Simulation results.

TABLE 9 Throughput Comparison (in Megabits per Second) for the IEEE 802.11a, with N ¼ 4

Experimental measurements.

hops away from the AAP. We have run measurement sessions with the IEEE 802.11a variant to avoid interference with the deployed wireless infrastructure. The wireless cards use the maximum available bit rate of 54 Mbps, and the AAP has a buffer of 20 frames. First, we have compared the measured performance of the standard IEEE 802.11 DCF (IEEE 802.11a PHY and MAC layers) and our solution with the analytical and simulation results given in the previous section. We have set up the same greedy scenario used in the simulation, with N ¼ 4 stations. We have set the CW of the AAP to 13, which corresponds to the required value for k ¼ 2. Table 9 presents the average throughput over 10 s and its distribution between stations and the AP. We can see that the AAP provides the throughput share close to the expected value of 0.66. Next, we have analyzed the behavior of TCP connections in mixed upload-download scenarios common to IEEE 802.11 cells that act as gateways to the Internet: some users download data to their wireless devices, whereas some others generate data uploaded via the AP. In the first experiment, we have measured the performance of the standard IEEE 802.11 DCF and our solution in a download oriented scenario: one upload ðM ¼ 1Þ competing with four downloads (N ¼ 5; cf. Figs. 5 and 19). We can observe that for DCF, the uploading station obtains a far better throughput than the downloading stations, because TCP segments destined to wireless stations queue at the AP, while the uploading station has the same channel access share as the AP and thus sends its segments at a greater rate. The AAP scheme alleviates this effect: the uploading station benefits from less throughput and the performance of downloads is improved. We use the FreeBSD default TCP congestion control, which is similar to Vegas.

LOPEZ-AGUILERA ET AL.: AN ASYMMETRIC ACCESS POINT FOR SOLVING THE UNFAIRNESS PROBLEM IN WLANS

Fig. 19. Data transferred versus the elapsed time, with four TCP downloads and one upload for the AAP. Experimental measurements.

In the case of an upload-oriented scenario: one download competing with four uploads (M ¼ 4, and N ¼ 5), the starvation of the download traffic under the IEEE 802.11 DCF is even more striking (cf. Figs. 3 and 20), whereas the AAP presents an opposite behavior by giving more share to the downloading station. We can also see that the unfairness among TCP uploads is not as important as previously reported in a study based on simulation results [6]. Nevertheless, the AAP improves fairness further on. Finally, we wanted to study the influence of the proposed solution on the performance of real-time flows such as VoIP. In a realistic situation, an AP also needs to support UDP flows that require short delays besides optimizing the performance of upload/download TCP connections. If we consider such a mixed UDP-TCP scenario, the downlink queue of the AP under the IEEE 802.11 DCF is permanently full so that TCP connections choke on the receiver window size or lower their congestion window due to frame losses when the AP downlink queue overflows. Fig. 21 presents histograms of the one-way delay experienced by VoIP RTP/UDP packets that originated in a wireless station in the presence of four TCP downloads initiated on other stations. VoIP traffic is bidirectional between the wireless station and a host connected to

Fig. 20. Data transferred versus the elapsed time, with one TCP download and four uploads for the AAP. Experimental measurements.

1225

Fig. 21. Histograms of one-way delays experienced by VoIP RTP/ UDP packets in the presence of four TCP downloads. (a) DCF: downlink. (b) DCF: uplink. (c) AAP: downlink. (d) AAP: uplink.

the wired part of the network. All stations involved in this experiment maintain precise clock synchronization with respect to a third computer to guarantee the correctness of one-way delay-measurements-based packet time stamps. We generate histograms from 30-s measurement sessions, which represent approximately 3,000 UDP packets observed during an emulated typical bidirectional VoIP call making use of a G.711 codec and RTP (92 bytes of UDP payload per frame). We can observe that the downlink delay under DCF exhibits really poor performance (recall that in our experimental setup, we have set up the buffer size to a small value of 20 frames, which mitigates the effect of queuing at the AP; with the default value of 50 frames used by FreeBSD, the delays would be even longer). The difference with AAP is striking: under DCF, the queue is always full, leading to long delays, while it is near empty with the AAP so that packets experience fairly short delays.

8

CONCLUSIONS

In this paper, we have proposed a simple elegant solution at the MAC layer to the unfairness problem. Our approach results in an adequate distribution of transmission opportunity between the AAP and wireless stations that supports efficient operation of TCP connections over IEEE 802.11 wireless LANs. By giving the AP a sufficient channel access share, we force the downlink queue at the AP to be almost empty all the time so that TCP connections benefit from the desired operational state, in which the destination controls the rate of the packet flow. Thus, instead of a low throughput due to congestion control and long delays caused by the saturated AP queue, we obtain a significant performance improvement in terms of throughput and delay, along with fair sharing of the wireless part of the network. Our analytical and simulation results show very good performance of the AAP in terms of throughput, access probability, and convergence speed. It compares favorably with the standard IEEE 802.11 DCF and with recent approaches based on IEEE 802.11e. Moreover, we present measurement data of TCP and UDP performance gathered

1226

IEEE TRANSACTIONS ON MOBILE COMPUTING,

on an experimental prototype that uses wireless cards implementing the proposed method. The measurements show that in common mixed download-upload TCP scenarios, our method increases the overall throughput and alleviates unfairness. Our solution limits itself to the MAC layer and our choice of allocating to the AAP the share corresponding to the worst case scenario (2/3 in case of k ¼ 2) results in a slight preference given to download connections. We think that this is the right choice, because the current usage of the network shows the dominance of download traffic [10]. Providing the AP with the optimal share of transmission opportunity in a given mixed upload-download scenario requires more information, which may be available at upper layers, and we plan to investigate this problem in the future. We also leave for future research the performance analysis of the AAP scheme for different frame sizes and multiple bit rates due to channel adaptation mechanisms. Moreover, we also plan to study the performance of the AAP scheme in the presence of hidden and exposed nodes due to overlapping cells. The major contribution of our work is the discovery of a surprisingly simple principle for allocating a required channel access share to the AP: it is sufficient to set its CW to a constant value if wireless stations operate according to the Idle Sense method. The main advantage of this principle is its extreme simplicity of implementation. Although the AAP requires to upgrade the cards firmware, it is compatible with the IEEE 802.11 PHY layer. So, we think that its integration with IEEE 802.11 products is possible either through proprietary extensions or by enhancing the current standards. Finally, we want to emphasize our measurement results gathered on the experimental testbed. This is something new in the domain concerned with the evaluation of new MAC layers: most of the proposed access methods or improvements to the IEEE 802.11 MAC are only validated by means of simulation or analytical modeling.

[4]

[5]

[6]

[7]

[8] [9]

[10]

[11] [12] [13]

[14]

[15]

[16]

[17]

[18]

[19]

ACKNOWLEDGMENTS The authors would like to thank the Intel Research Laboratory, Cambridge, in particular Dina Papagiannaki, for making the implementation of Idle Sense possible and the anonymous referees for many helpful comments. This work was partially supported by the European Commission Project WIP under Contract 27402, by the French Ministry of Research Project AIRNET under Contract ANR-05RNRT-012-01, by the Spanish Ministry of Education and Science through the CICYT Project TEC2006-04504, and by the DURSI of the Catalan Government.

REFERENCES [1] [2] [3]

IEEE 802.11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications, IEEE, 1999. H. Balakrishnan and V.N. Padmanabhan, “How Network Asymmetry Affects TCP,” IEEE Comm. Magazine, vol. 39, no. 4, pp. 60-67, Apr. 2001. S. Pilosof, R. Ramjee, D. Raz, Y. Shavitt, and P. Sinha, “Understanding TCP Fairness over Wireless LAN,” Proc. IEEE INFOCOM ’03, vol. 2, pp. 863-872, Mar.-Apr. 2003.

[20]

[21]

[22]

[23]

[24]

[25]

[26]

VOL. 7,

NO. 10,

OCTOBER 2008

X. Wang and S.A. Mujtaba, “Performance Enhancement of 802.11 Wireless LAN for Asymmetric Traffic Using an Adaptive MAC Layer Protocol,” Proc. 56th IEEE Vehicular Technology Conf. (VTC-Fall ’02), vol. 2, pp. 753-757, Sept. 2002. D.J. Leith and P. Clifford, “Using the 802.11e EDCF to Achieve TCP Upload Fairness over WLAN Links,” Proc. Third IEEE Int’l Symp. Modeling and Optimization in Mobile, Ad-Hoc and Wireless Networks (WiOpt ’05), pp. 109-118, Apr. 2005. D.J. Leith, P. Clifford, D. Malone, and A. Ng, “TCP Fairness in 802.11e WLANs,” IEEE Comm. Letters, vol. 9, no. 11, pp. 964-966, Nov. 2005. M. Heusse, F. Rousseau, G. Berger-Sabbatel, and A. Duda, “Performance Anomaly of 802.11b,” Proc. IEEE INFOCOM ’03, vol. 2, pp. 836-843, Mar.-Apr. 2003. IEEE 802.11e: Wireless LAN MAC PHY: Medium Access Control (MAC) Quality-of-Service Enhancements, IEEE, 2005. M. Heusse, F. Rousseau, R. Guillier, and A. Duda, “Idle Sense: An Optimal Access Method for High Throughput and Fairness in Rate Diverse Wireless LANs,” Proc. ACM SIGCOMM ’05, vol. 35, pp. 121-132, Aug. 2005. C. Na, J.K. Chen, and T.S. Rappaport, “Measured Traffic Statistics and Throughput of IEEE 802.11b Public WLAN Hotspots with Three Different Applications,” IEEE Trans. Wireless Comm., vol. 5, no. 11, pp. 3296-3305, Nov. 2006. E. Kohler, M. Handley, and S. Floyd, Datagram Congestion Control Protocol (DCCP), IETF RFC 4340, Mar. 2006. L. Ong and J. Yoakum, An Introduction to the Stream Control Transmission Protocol (SCTP), IETF RFC 3286, May 2002. G. Bianchi and I. Tinnirello, “Kalman Filter Estimation of the Number of Competing Terminals in an IEEE 802.11 Network,” Proc. IEEE INFOCOM ’03, vol. 2, pp. 844-852, Mar.-Apr. 2003. S. Lu, V. Bhargavan, and R. Srikant, “Fair Scheduling in Wireless Packet Networks,” Proc. ACM SIGCOMM ’97, vol. 27, pp. 63-74, Oct. 1997. M. Bottigliengo, C. Casetti, C.F. Chiasserini, and M. Meo, “Smart Traffic Scheduling in 802.11 WLANs with Access Point,” Proc. 58th IEEE Vehicular Technology Conf. (VTC-Fall ’03), vol. 4, pp. 2227-2231, Oct. 2003. N.H. Vaidya, P. Bahl, and S. Gupta, “Distributed Fair Scheduling in a Wireless LAN,” Proc. ACM MobiCom ’00, pp. 167-178, Aug. 2000. A. Banchs and X. Perez, “Distributed Weighted Fair Queueing in 802.11 Wireless LAN,” Proc. IEEE Int’l Conf. Comm. (ICC ’02), vol. 5, pp. 3121-3127, Apr.-May 2002. H. Wu, Y. Peng, K. Long, S. Cheng, and J. Ma, “Performance of a Reliable Transport Protocol over IEEE 802.11 Wireless LAN: Analysis and Enhancement,” Proc. IEEE INFOCOM ’02, vol. 2, pp. 599-607, June 2002. A. Balachandran, G.M. Voelker, P. Bahl, and V. Rangan, “Characterizing User Behavior and Network Performance in a Public Wireless LAN,” Proc. ACM SIGMETRICS ’02, vol. 30, pp. 195-205, June 2002. E. Amir, H. Balakrishnan, S. Seshan, and R. Katz, “Efficient TCP over Networks and with Wireless Links,” Proc. Fifth IEEE Workshop Hot Topics in Operating Systems (HotOS ’95), pp. 35-40, May 1995. M. Garcia, J. Choque, L. Sa´nchez, and L. Mun˜oz, “An Experimental Study of Snoop TCP Performance over the IEEE 802.11b WLAN,” Proc. Fifth IEEE Int’l Symp. Wireless Personal Multimedia Comm. (WPMC ’02), vol. 3, pp. 1068-1072, Oct. 2002. A.V. Bakre and B.R. Badrinath, “Implementation and Performance Evaluation of Indirect TCP,” IEEE Trans. Computers, vol. 46, no. 3, pp. 260-278, Mar. 1997. L. Zhang, S. Shenker, and D.D. Clark, “Observations and Dynamics of a Congestion Control Algorithm: The Effects of Two-Way Traffic,” Proc. ACM SIGCOMM ’91, vol. 21, pp. 133-147, Aug. 1991. L. Kalampoukas, A. Varma, and K.K. Ramakrishnan, “Performance of Two-Way TCP Traffic over Asymmetric Access Links,” Proc. Interop Eng. Conf., May 1997. K. Xu, Y. Tian, and N. Ansari, “Improving TCP Performance in Integrated Wireless Communications Networks,” Elsevier Computer Networks, vol. 47, no. 2, pp. 219-237, Feb. 2005. D. Eckhardt and P. Steenkiste, “Effort-Limited Fair (ELF) Scheduling for Wireless Networks,” Proc. IEEE INFOCOM ’00, vol. 3, pp. 1097-1106, Mar. 2000.

LOPEZ-AGUILERA ET AL.: AN ASYMMETRIC ACCESS POINT FOR SOLVING THE UNFAIRNESS PROBLEM IN WLANS

[27] M. Heusse, P. Starzetz, F. Rousseau, G. Berger-Sabbatel, and A. Duda, “Bandwidth Allocation for DiffServ-Based Quality of Service over 802.11b,” Proc. IEEE Global Telecomm. Conf. (GLOBECOM ’03), vol. 2, pp. 992-997, Dec. 2003. [28] D. Skyrianoglou, N. Passas, and A. Salkintzis, “Support of IP QoS over Wireless LANs,” Proc. 59th IEEE Vehicular Technology Conf. (VTC-Spring ’04), vol. 5, pp. 2993-2997, May 2004. [29] J. Ha and C.-H. Choi, “TCP Fairness for Uplink and Downlink Flows in WLANs,” Proc. IEEE Global Telecomm. Conf. (GLOBECOM ’06), Nov. 2006. [30] V.A. Siris and G. Stamatakis, “Optimal CWmin Selection for Achieving Proportional Fairness in Multi-Rate 802.11e WLANs: Testbed Implementation and Evaluation,” Proc. First ACM Int’l Workshop Wireless Network Testbeds, Experimental Evaluation and Characterization (WiNTECH ’06), pp. 41-48, Sept. 2006. [31] S. Shin and H. Shulzrinne, “Balancing Uplink and Downlink Delay of VoIP Traffic in WLANs Using Adaptive Priority Control (APC),” Proc. Third ACM Int’l Conf. Quality of Service in Heterogeneous Wired/Wireless Networks (QShine ’06), Aug. 2006. [32] R. Bruno, M. Conti, and E. Gregori, “Design of an Enhanced Access Point to Optimize TCP Performance in Wi-Fi Hotspot Networks,” ACM/Springer Wireless Networks J., vol. 13, no. 2, pp. 259-274, Apr. 2007. [33] A. Banchs, A. Azcorra, C. Garcia, and R. Cuevas, “Applications and Challenges of the 802.11e EDCA Mechanism: An Experimental Study,” IEEE Network, vol. 19, no. 4, pp. 54-58, July-Aug. 2005. [34] D. Chiu and R. Jain, “Analysis of the Increase and Decrease Algorithms for Congestion Avoidance in Computer Networks,” J. Computer Networks and ISDN, vol. 17, no. 1, June 1989. [35] G. Bianchi, “Performance Analysis of the IEEE 802.11 Distributed Coordination Function,” IEEE J. Selected Areas in Comm.: Wireless Series, vol. 18, no. 3, pp. 535-547, Mar. 2000. [36] Y. Grunenberger, M. Heusse, F. Rousseau, and A. Duda, “Experience with an Implementation of the Idle Sense Wireless Access Method,” to appear in, Proc. Third Conf. Future Networking Technologies (CoNEXT), 2007. [37] E. Lopez-Aguilera, J. Casademont, and J. Cotrina, “Outdoor IEEE 802.11g Cellular Network Performance,” Proc. IEEE Global Telecomm. Conf. (GLOBECOM ’04), vol. 5, pp. 2992-2996, Nov.-Dec. 2004. [38] E. Lopez-Aguilera, M. Heusse, F. Rousseau, A. Duda, and J. Casademont, “Performance of Wireless LAN Access Methods in Multicell Environments,” Proc. IEEE Global Telecomm. Conf. (GLOBECOM ’06), Nov.-Dec. 2006. [39] S. Kuppa and R. Prakash, “Service Differentiation Mechanisms for IEEE 802.11-Based Wireless Networks,” Proc. IEEE Wireless Comm. and Networking Conf. (WCNC ’04), vol. 2, pp. 796-801, Mar. 2004. Elena Lopez-Aguilera received the MS degree in telecommunications engineering from the Technical University of Catalonia (UPC), Barcelona, in 2001. In 2002, she joined the Wireless Networks Group (WNG), Department of Telematics Engineering, UPC, where she is currently an associate professor. From 2005 to 2006, she was an invited researcher at the Grenoble Computer Science Laboratory (LIG). Her research interests include the study of medium access protocols in WLANs and mesh networks. She has published papers in the area of wireless communications and MAC mechanisms. Martin Heusse received the diploma in telecommunications engineering from TELECOM Bretagne Engineering School in 1996 and the PhD degree in 2001. He is an assistant professor at the Joseph Fourier University, Grenoble. He is currently with the Grenoble Computer Science Laboratory (LIG). His research interests include routing and the design and evaluation of access methods in wireless LANs and sensor networks.

1227

Yan Grunenberger received the diploma in telecommunications engineering in 2004 from Grenoble Institute of Technology (INPG), where he is currently working toward the PhD degree. His PhD research is focused on cross-layer techniques in next-generation wireless networks. In 2006, he was an intern at Intel Research, Cambridge, implementing new access methods on the Intel hardware. He is currently with the Grenoble Computer Science Laboratory (LIG). Franck Rousseau received the PhD degree from the Grenoble Institute of Technology (INPG) in 1999. He was a member of technical staff at the OSF Research Institute, Grenoble.  He is an assistant professor at the INPG-Ecole Nationale Supe´rieure d’Informatique et de Mathe´matiques Applique´es de Grenoble (ENSIMAG). He is currently with the Grenoble Computer Science Laboratory (LIG). He is actively involved in national and European projects and has served as program committee member for several networking conferences. His research interests include wireless networking, in particular mobility, ad hoc, mesh, and medium access networks. He has published papers in the fields of multimedia and networking. Andrzej Duda received the PhD degree from the Universite´ de Paris-Sud in 1984 and the Habilitation diploma from Grenoble University in 1994. He was an assistant professor at the Universite´ de Paris-Sud, a charge´ de recherche at CNRS, and a visiting scientist at the MIT Laboratory for Computer Science. From 2002 to 2003, he was an invited professor at the Swiss Federal Institute of Technology, Lausanne (EPFL). He is a professor at the Grenoble Institute of Technology-Ecole Nationale Supe´rieure d’Informatique et de Mathe´matiques Applique´es de Grenoble (ENSIMAG). He is currently with the Grenoble Computer Science Laboratory (LIG). He has been on the program committee of several international conferences and acts as an expert for the European Commission. He published more than 90 papers in the areas of performance evaluation, distributed systems, multimedia, and networks. Jordi Casademont received the MS degree in telecommunications and the PhD degree from the Technical University of Catalonia (UPC), Barcelona, in 1992 and 1998, respectively. He is currently an assistant professor at UPC. He is an active member of the Wireless Networks Group (WNG), Department of Telematics Engineering, UPC. His research interests include the design and evaluation of MAC mechanisms and mesh networks. He is actively involved in national and European projects. He has published papers in the fields of networking and medium access control. . For more information on this or any other computing topic, please visit our Digital Library at www.computer.org/publications/dlib.