You are on page 1of 5

1

Analysis of Unfairness between TCP Uplink and Downlink Flows in Wi-Fi Hot Spots

Samsung Electronics Co., LTD., Suwon, Korea, E-mail: {eunchan.park,dy470.kim}@samsung.com Seoul National University, Seoul, Korea, E-mail: chchoi@csl.snu.ac.kr

Eun-Chan Park , Dong-Young Kim , and Chong-Ho Choi

Abstract This paper focuses on the unfairness problem between TCP uplink and downlink ows in the 802.11 Wi-Fi hot spots and shows that the service is prone to be unfair. The cause of unfairness is analyzed from two aspects: TCPinduced asymmetry and MAC-induced asymmetry. Due to the asymmetric behavior of TCP congestion control with a cumulative acknowledgment mechanism between uplink and downlink ows, the service is biased toward the uplink ow and the downlink ow tends to starve. The contention-based channel access mechanism of 802.11 MAC exacerbates this unfairness problem because it intends to provide fair access opportunity only to the sending stations. Next, the analysis of the interaction between congestion control of TCP and contention control of MAC reveals interesting and counter-intuitive results: (i) Even when a station has a sufciently large amount of trafc to send, it does not always participate in the MAC-layer contention, its opportunity for MAC-layer contention is controlled by the TCP congestion control. (ii) The aggregate throughput remains almost constant with respect to the number of stations sending/receiving TCP trafc. (iii) Both TCP-induced unfairness and MAC-induced unfairness can be resolved if packet loss due to buffer overow in an access point does not occur. Index Terms fairness, TCP congestion control, contention control, IEEE 802.11 WLAN

I. I NTRODUCTION Nowadays, wireless communication has become very popular along with the explosive growth of the Internet. In order to realize users demand to access the Internet at any time and at any place, wireless Internet service providers (WISPs) widely deploy wireless local area networks (WLANs) based on the IEEE 802.11 standard [1] in public areas such as campuses, cafes, and convention centers. These WLAN access networks are also called Wi-Fi hot spots. The number of Wi-Fi hot spot users and/or devices is increasing explosively. Besides laptops and personal digital assistants, many new mobile devices such as cellular phones, portable media players, and portable game devices tend to support Internet connectivity in Wi-Fi hot spots. Therefore, fair service is one of the most crucial concerns to be considered. This paper focuses on the unfairness problem between TCP uplink (from a station to access point (AP)) and downlink (from AP to a station) ows in Wi-Fi hot spots. Note that most of the trafc in the current Internet utilizes TCP as transport-layer protocol and that most Wi-Fi hot spots operate in the IEEE 802.11 infrastructure mode. Also, we consider that wireless stations deploy a contention-based medium access control (MAC) protocol (e.g., IEEE 802.11 distributed coordination function (DCF)). In this paper, it is shown that the service in the current Wi-Fi hot spot is prone to be unfair; the service is biased toward uplink ows and downlink ows tend

to starve. Unlike ad-hoc networks with UDP trafc, hot spot networks with TCP trafc have two asymmetric properties; TCP-induced asymmetry and MAC-induced asymmetry, which are the causes of the unfairness problem. This paper shows that the TCP congestion control mechanism with a cumulative ACK mechanism affects the behavior of a ow quite differently, depending on whether the ow is uplink or downlink. This TCP-induced asymmetry makes the service biased toward uplink ows. On the other hand, we show that MAC-induced asymmetry exacerbates the unfairness problem. A contentionbased channel access mechanism attempts to provide equal access opportunity to all competing stations. However, in a hot spot network where most stations receive service from an AP, only sending stations and the AP participate in the contention. This MAC-induced asymmetry leads to unfair service. Recently, the issue of fairness between uplink and downlink ows in infrastructure WLANs has been studied in [2][6]. In order to mitigate the unfairness problem, the approach in [2] gives more transmission opportunities to the AP by increasing the minimum contention window size of stations. Similarly, the downlink compensation access (DCA) algorithm in [3] gives higher channel access priority to the AP by shortening its interframe space. The schemes proposed in [4] modies channel access mechanism with mean of backoff distribution, in order to provide fairness between uplink and downlink ows. These mechanisms [2][4] are MAC-layer solutions for fairness and do not consider the characteristics of upperlayer trafc with an implicit assumption that every station has innite amount of trafc to send and is always involved in the MAC layer contention. However, this assumption is no longer valid for TCP trafc in Wi-Fi hot spot networks due to the asymmetric properties addressed above. On the other hand, the studies in [5], [6] focus on the unfairness problem among TCP ows, and their approaches do not modify the channel access mechanism for fairness. In [5], it is shown that the fairness is affected by the buffer availability of the AP. The solution proposed in [5] is based on the manipulation of the advertised TCP window size at the AP. The approach proposed in [6] deploys per-ow queue at the AP and is based on per-ow scheduling, without modifying both MAC and TCP protocols. Also, the study in [7] has derived an analytical model of TCP throughput in Wi-Fi hop spot. However, this model has a limitation in that only downlink TCP ows are considered and uplink TCP ows are not taken into account. This paper analyzes the unfairness problem by considering the interaction between congestion control of TCP and contention control of MAC. It reveals the following interesting and counter-intuitive results, which are the contributions of

corresponding wired nodes


throughput (Mb/s)
11 00 11 00 11 00 11 00

WiFi hot spot router router


100Mb/s 10ms 100Mb/s 10ms

throughput (Mb/s)

4 3 2 1 0

UP_STA DN_STA

4 3 2 1 0

UP_STA DN_STA

100Mb/s [2040]ms

uplink flow
AP
11 00 11 00 11 00 UP_STA 11 00 11 00 11 00

downlink flow

20

40

60

80

100

20

40

60

80

100

time (sec)

time (sec)

(a) UDP

(b) TCP

wireless stations

DN_STA

Fig. 2. Comparison of UDP throughput and TCP throughput achieved by DN STA and UP STA.
TCP DATA for DN_STA TCP ACK for UP_STA

Fig. 1.

Network conguration of a Wi-Fi hot spot.

this study; (i) Even when a station has a sufciently large amount of TCP trafc to send, it does not always participate in the MAC-layer contention to occupy channel. (ii) The aggregate throughput is little affected by the number of stations as long as they send/receive TCP trafc. (iii) As well as TCP-induced unfairness, MAC-induced unfairness is resolved if packet loss due to buffer overow in the AP does not occur. The reason is that the sending rate of TCP trafc is intrinsically controlled by the TCP congestion control. Thus, in contrast to the case of ad-hoc networks with UDP trafc, the opportunity for the MAC-layer contention in Wi-Fi hot spots is affected by the TCP congestion control. The rest of the paper is organized as follows. In Section II and III, the cause of unfairness is investigated from two aspects: TCP-induced asymmetry and MAC-induced asymmetry, respectively. The effect of packet loss due to buffer overow in the AP is discussed in Section IV. In Section V, we analyze the interaction between congestion control of TCP and contention control of MAC. The conclusion follows in Section VI. II. TCP- INDUCED UNFAIRNESS In this section, we perform ns-2 [8] simulation studies to investigate the cause of unfairness between TCP uplink and downlink ows focusing on the asymmetry of TCP congestion control responding to packet loss. For the simplicity of analysis, this study ignores wireless channel error and focuses on a simple scenario where the wireless stations in the Wi-Fi hot spot communicate with the corresponding wired nodes via the AP as shown in Fig. 1, which is a typical conguration of the current Wi-Fi hot spot. Here, Ndn stations receive data trafc from the AP while Nup stations send data to the AP. We refer to these stations as DN STAs and UP STAs, respectively.1 The simulations use the standard values for IEEE 802.11b MAC/PHY parameters [1] and the transmission rate is set to 11 Mb/s. The version of TCP used in the simulations is TCP/Reno. The simulation topology, link capacity and propagation delay are given in Fig. 1. The packet size for data trafc is set to 1 Kbyte. Here, both the advertised window size of TCP ow and the buffer size of AP are set to 50 packets.
1 For the coherence of notation between a station and a ow, DN STA/UP STA indicate stations having a downlink/uplink ow, respectively.

1111 0000 1111 0000 1111 0000 1111 0000 1111 0000
AP
TCP DATA TCP ACK

packet (DATA/ACK) drop due to bufferoverflow

TCP DATA

DN STA

UP STA

Fig. 3. Schematic diagram of access point downlink buffer shared by TCP uplink and downlink ows.

First, we consider a case of Nup = Ndn = 1, which is simple nevertheless sufcient to observe the effect of TCPinduced asymmetry on the channel sharing problem. In this case, it is expected that UP STA and AP (that is responsible for the service for DN STA) compete for the wireless channel following the contention-based channel access mechanism (i.e., DCF) so that UP STA and DN STA have similar amount of throughput. However, this is not true when TCP is used as a transport-layer protocol. By comparing Fig. 2(a) and Fig. 2(b), we observe a bias toward UP STA in the case of TCP trafc. For UDP trafc (Fig. 2(a)), there is little difference between the throughputs achieved by DN STA and UP STA. On the other hand, for TCP trafc (Fig. 2(b)), we observe that the average throughputs of UP STA ( 2.26 Mb/s) is higher than that of DN STAs ( 1.28 Mb/s) by more than 75 %. Also, Fig. 2 shows that the throughput of UDP trafc is almost constant; however, that of TCP trafc uctuates severely, which is due to the TCP congestion control. This bias toward UP STA results from the asymmetric behavior of TCP ows responding to packet loss. Note that a TCP connection is bidirectional, so a sender transmits a data packet to a receiver and the receiver sends the corresponding acknowledgement (ACK) packet to the sender. When DN STA and UP STA share a wireless channel, there are two different kinds of packets buffered in the interface queue of the AP: data packets to DN STA and ACK packets to UP STA, as illustrated in Fig. 3. In general, the capacity of a wireless link is smaller than that of a wired link. Moreover, in Wi-Fi hot spots, AP is always involved in the communication and so it can be easily congested. Thus, the wireless link easily becomes a bottleneck link and packets are occasionally dropped due to buffer overow. The TCP ows for DN STA and UP STA react to this packet loss in different ways. When a data packet for DN STA is

congestion window (packets)

throughput ratio (UP/DN)

throughput ratio (UP/DN)

UP_STA DN_STA

congestion window (packets)

60

60

UP_STA DN_STA

12 9 6 3 0 TCP UDP estimate

25 20 15 10 5 0 0 2 4 6 8 number of DN STAs (Ndn) 10 0 2 4 6 8 10 number of UP/DN STAs (Nup = Ndn) TCP UDP estimate

40

40

20

20

0 0 20 40 60 80 100 time (sec)

0 0 20 40 60 80 100 time (sec)

(a) Ndn = Nup = 1

(b) Ndn = Nup = 2

Fig. 4. Comparison of TCP congestion window size for DN STA and UP STA (B = Wmax = 50 packets).

(a) Case1 (Nup = 1, Ndn = 1 10)

(b) Case2 (Nup = Ndn = 1 10)

Fig. 5. Throughput ratios (tcp , udp , est ) for Case1 and Case2 (B = Wmax = 50 packets).

lost, the receiver (DN STA) transmits duplicate ACKs or a time-out occurs at the sender (wired node). Then, the sender Case1 : Nup = 1, and Ndn = 1 10. reduces its congestion window size, decreasing the throughput Case2 : Nup = Ndn = 1 10. of DN STA. However, the loss of an ACK packet for UP STA We quantify the degree of unfairness by introducing which does not affect its throughput much due to the cumulative is dened as the ratio of average throughput of UP STA to that ACK mechanism of TCP. Even though an ACK packet is of DN STA. Here, we observe how the number of contending lost due to buffer overow, the loss does not necessarily stations affects . We dene est as the estimated value of invoke the TCP congestion control mechanism of UP STA. under two assumptions: The cumulative ACK mechanism lets UP STA tolerate the loss (A1) All data sending stations, i.e., UP STAs and AP, are of ACK packet as long as the next ACK packet with a higher always backlogged so that they always participate in the sequence number is delivered timely to UP STA. Then, the contention. sender (UP STA) does not reduce its congestion window. Due (A2) The channel access mechanism gives approximately equal to this asymmetric behavior in response to packet loss in the opportunities to the competing stations on average. AP buffer, the throughput of UP STA becomes higher than For Case1, one UP STA and the AP that contends on behalf that of DN STA. We further investigate the asymmetric behavior of TCP by of all DN STAs compete to access the channel. The single comparing the TCP congestion window size for DN STA and UP STA occupies the channel with probability of one half, UP STA, which are denoted as wdn and wup , respectively. In while each DN STA has opportunity to be served by the AP the case of Ndn = Nup = 1 (Fig. 4(a)), wup increases up with probability of 0.5/Ndn . In a similar way, we can estimate Case2. to its maximum size (advertised window size), Wmax , and est for 1 1 then does not change from Wmax . However, wdn decreases 1 : Case1, 2 2 Ndn = Ndn abruptly whenever a data packet is dropped in the AP buffer, est = Nup 1 1 1 : Case2. resulting in the sudden decrease of throughput for DN STA Nup +1 Nup Nup +1 Ndn = Ndn (1) (see Fig. 2(b)), and is always smaller than Wmax . In the case of Ndn = Nup = 2 (Fig. 4(b)), wdn shrinks to only a few packets Note that the values of est for both cases are the same. Let us denote tcp and udp as the values of obtained from and cannot grow any more. This phenomenon is caused by the burstiness of TCP trafc combined with its asymmetric the simulations with TCP trafc and UDP trafc, respectively. behavior regarding to packet loss. Since TCP sender transmits In the simulation, Wmax = B = 50 packets. For UDP trafc, packets burstily up to the size of the congestion window, constant bit rate (CBR) trafc is used with a sufciently high several packets belonging to the same congestion window rate to fully utilize the channel capacity. From Fig. 52 , we may be dropped with a higher probability as the queue size observe the following: approaches its maximum size. These multiple losses of data The values of est calculated by (1) agree well with packets may trigger the TCP time-out mechanism, shrinking those of udp . However, there exists a notable difference the congestion window size to the initial value (i.e., one). If between est and tcp . this event of multiple packet losses occurs successively, the For both Case1 and Case2, the fairness problem becomes time-out interval is doubled (see Fig. 4(b) during time t 40 severe as the number of stations increases. Furthermore, s 80 s), and the sender is frozen for the increased timethe degree of unfairness is magnied in the case of TCP out interval, resulting in the drastic decrease of throughput. trafc; tcp is always higher than udp . Consequently, the burstiness of TCP trafc with its asymmetric While the difference between tcp and est is nearly behavior exacerbates the unfairness between DN STAs and constant in Case1, it increases almost linearly as Nup UP STAs. increases in Case2. These results that est is nearly equal to udp conrm that III. MAC- INDUCED UNFAIRNESS two assumptions made in computing est , (A1) and (A2), In this section, we study the cause of unfair channel sharing 2 The simulation results in Fig. 2 and Fig. 4 are obtained from a single focusing on the asymmetric property of MAC in Wi-Fi hot instance of simulation to show the results with respect to time. The other spots. To illustrate the unfairness problem, we consider two results shown with respect to parameters (i.e., Fig. 5 Fig. 7) are averaged over ten instances of simulation with random seed. cases:

are reasonable. In the case of TCP trafc, however, these assumptions do not hold any more; tcp does not agree with est . The difference between tcp and est results from the asymmetric behavior of TCP discussed in Section II. While this difference is between 1 and 1.5 for Case1, it ranges from 1 to 14 for Case2. For Case1 where Nup = 1, the increase of Ndn does not have any signicant effect on packet loss rate for DN STAs because wdn shrinks to a very small value (See Fig. 4(b)). Therefore, the difference between tcp and est is almost constant, regardless of Ndn . On the other hand, for Case2 where Nup changes, the increase of Nup tends to increase the queue length of the AP buffer because UP STA has large congestion window size. Accordingly, packet loss rate for DN STAs increases, which reduces their sending rates, and the degree of unfairness also increases. As Nup increases, the fairness degrades signicantly, e.g, tcp 25 when Nup ( = Ndn ) = 10. IV. E FFECT OF PACKET LOSS DUE TO BUFFER OVERFLOW Here, we study how the packet loss due to buffer overow in the AP buffer affects TCP-induced unfairness and MACinduced unfairness. It is worthwhile to note that packet loss due to wireless channel error or contention can be recovered by the 802.11 link-layer retransmission mechanism and this kind of packet loss is not asymmetric between UP STA and DN STA. Thus, we mainly focus on the effect of packet loss due to buffer overow in the AP buffer. A. On TCP-induced unfairness Based on the analysis in Section II, we can infer that the TCP-induced unfairness can be alleviated if packet loss rate decreases. We can avoid packet loss due to buffer overow by either making the buffer size, B, sufciently large or restricting the maximum congestion window size, Wmax . In this subsection, we study the effect of packet loss on TCP-induced unfairness. For this purpose, we consider the simple case of Ndn = Nup = 1, and set B = 50 packets Wmax = 10 80 packets. We observe the following from Fig. 6: By restricting Wmax properly, i.e., Wmax < 40, it can be guaranteed that packets are not lost due to buffer overow in the AP, as shown in Fig. 6(a). As long as no packet is dropped, there is no notable discrepancy between throughputs of DN STA and UP STA (see Fig. 6(b)). However, the unfairness problem becomes severe as packet loss rate increases. There exists a trade-off between aggregate throughput and fairness when Wmax < 30. In this case, the channel is not fully utilized due to the limit on Wmax , and the fairness is achieved at the cost of reduced aggregate throughput. The aggregated throughput is nearly insensitive to the change of Wmax if it exceeds a certain threshold value, i.e., Wmax > 30. From these observations, we conclude that there exists an optimal value for Wmax to achieve fairness and high channel utilization simultaneously, e.g., Wmax = 30 40 packets in this simulation. It is obvious that, however, this optimal value depends on network condition and trafc load, e.g., the buffer size, the number of ows, and round trip time (RTT) of TCP ows.

2 packet loss rate (%) 1.6 1.2 0.8 0.4 0

UP_STA DN_STA throughput (Mb/s) 0 20 40 60 80

UP_STA 4 DN_STA total 3 2 1 0 0 20 40 60 80

maximum congestion window (packets)

maximum congestion window (packets)

(a) packet loss rate

(b) throughput

Fig. 6. Effect of packet loss on TCP-induced unfairness and utilization (B = 50 packets and Wmax = 10 80 packets).

B. On MAC-induced unfairness Here, we investigate the effect of packet loss on MACinduced unfairness. By setting the APs buffer size, B, to a sufciently large value, we can reduce the packet loss rate and mitigate the adverse effect of the TCP-induced asymmetry. We consider Case1 and Case2 in Section III and set B = 500 packets and Wmax = 50 packets. Figure 7(a) shows the loss rate of data packets for DN STA, ldn . In Case1, no packet is dropped for the entire range of Ndn . In Case2, however, some portion of packets are dropped if Ndn (= Nup ) exceeds 4. From Figures 7(b) and 7(c), we make the following observations: As long as ldn = 0 in Case1 and Case2, all stations share the wireless channel equally, i.e., 1. Once packets are dropped in Case2; however, UP STA occupies the channel more than DN STA, and increases up to 20. The aggregate throughput for both Case1 and Case2 is almost independent of the number of stations, even when ldn > 0. Also, the throughput difference between Case1 and Case2 is negligible. Intuitively, we expect that every contending station (AP and UP STAs) under the CSMA-based contention shares the wireless channel equally on average. This expectation has already been conrmed by the simulation results for UDP trafc in Fig. 5, i.e., est udp . However, the simulation results described in Fig. 7(b) are counter to this expectation for the case of TCP trafc. For example, in the cases of Nup = 1, Ndn = 10 (Case1) and Nup = Ndn = 4 (Case2), all stations have nearly the same throughput, i.e., 1. This expectation holds true if all contending stations always take part in the contention to occupy channel, which is not true for TCP trafc. TCP sender adjusts its sending rate depending on the network condition (e.g., packet loss and RTT). If packet loss rate or RTT increases, the TCP congestion control mechanism limits the opportunity for a station to participate in the MAC-layer contention. We can nd another important observation about aggregate throughput from Fig. 7(c). In Case1 where Nup = 1, the increase of Ndn has little effect on the degree of contention, because DN STAs do not participate in the contention directly. Therefore, the aggregate throughput in Case1 is almost constant regardless of the value of Ndn . Even in Case2 where Nup = 1 10, the aggregate throughput is little affected by the increase of Nup . As opposed to the intuition that the aggregate throughput decreases as the number of contending stations increases, the actual aggregate throughput in the

aggregate throughput (Mb/s)

10 packet loss rate (%) 8 6 4 2 0 0

throughput ratio (UP/DN)

Case1 Case2

10 8 6 4 2 1 0

Case1 Case2

5 4 3 2 1 0 0

Case1 Case2

2 4 6 8 number of UP/DN STAs

10

2 4 6 8 number of UP/DN STAs

10

2 4 6 8 number of UP/DN STAs

10

(a) packet loss rate Fig. 7.

(b) throughput ratio

(c) aggregate throughput

Effect of packet loss on MAC-induced unfairness and utilization (B = 500 packets and Wmax = 50 packets).

simulation maintains nearly a constant value, regardless of Nup . This observation implies that the degree of contention is not in proportion to the number of contending stations if the stations send/receive TCP trafc whose rate is controlled by the congestion control mechanism. These counter-intuitive results will be explained in the next section by the analysis of interaction between congestion control of TCP and contention control of MAC. V. I NTERACTION BETWEEN TCP AND MAC We consider that the channel sharing problem in Wi-Fi hot spots is analogous to a card game. While UP STAs participate in the game by themselves, DN STAs do not participate directly. Instead, an AP, which is analogous to a dealer, participates in the game on behalf of DN STAs. Also, TCP window size is analogous to a credit for betting games. In the game of occupying shared channel, there are two kinds of rules; contention control of MAC and congestion control of TCP. We consider that the rst rule of MAC is fair so that it assures almost equal probability of winning the game (occupying channel) among players taking part in the current game (contending stations, i.e., UP STAs and AP). It is important to note that a player cannot join all the games because its credit is limited. According to the second rule of TCP, the player is allowed to join the game. If a player exhausts its credit, it cannot join the game any more and should wait until the credit is regenerated. After some time (RTT) has passed, the credit is regenerated. Note that the interval of credit generation is typically much greater than the interval of games (RTT is in the order of tens or hundreds of milliseconds, while the contention interval of 802.11 DCF is less than a few milliseconds). If the credit generation rate is excessive so that the dealer cannot accommodate all the credits for the players, less credit is generated in the next interval (the congestion window size reduces due to packet loss). But, some players may not consider this as a loss and continues to get the maximum credit (an UP STA has the maximum congestion window due to TCP cumulative ACK mechanism). Consequently, some of the players (UP STAs) tend to get more chances to win the game, while other players (DN STAs) tend to get less chances, which results in the unfairness problem. However, consider that a credit loss does not occur (e.g., by a large buffer or a small maximum congestion window), all the players (UP STAs and DN STAs) get the same amount of credit

for every credit generation interval. Compared to an individual player (UP STA), the dealer has more credits in proportion to the number of trusters (DN STAs) and participates in more games. Therefore, the probability that the dealer wins the games increases accordingly, and fairness among all the players can be assured. VI. C ONCLUSION In this paper, we have studied the issue of fairness between TCP uplink and downlink ows in Wi-Fi hot spots. We identied the unfairness problem - namely that the service is biased toward uplink ows while downlink ows tend to starve. The cause of this unfairness was analyzed from the viewpoint of TCP-induced and MAC-induced asymmetries. Also, we analyzed the interaction between congestion control of TCP and contention control of MAC. Based on this analysis, we can conclude that the TCP congestion control adjusts the opportunity for a station to participate in the MAC-layer contention and MAC-induced unfairness may be resolved by the TCP congestion control mechanism. As our future work, we will study a cross-layer approach between TCP and MAC that can provide fair service and high utilization. ACKNOWLEDGMENT This work was supported in part by POSCO, Korea. R EFERENCES
[1] IEEE 802.11 WG, Part 11: Wireless LAN medium access control (MAC) and physical layer (PHY) specication, IEEE 802.11 Standard, 1999. [2] M. Bottigliengo, C. Casetti, C.-F. Chiasserini, and M. Meo, Smart trafc scheduling in 802.11 WLANs with access point, in Proceedings of IEEE VTC-Fall, 2003. [3] S. Kim, B.-S. Kim, and Y. Fang, Downlink and uplink resource allocation in IEEE 802.11 wireless LANs, IEEE Trans. on Vehicular Technology, vol. 54, pp. 320327, January 2005. [4] J. Jeong, S. Choi, and C. Kim, Achieving weighted fairness between uplink and downlink in IEEE 802.11 DCF-based WLANs, in Proceedings of IEEE QShine, 2005. [5] S. Pilosof, R. Ramjee, D. Raz, Y. Shavitt, and P. Sinha, Understanding TCP fairness over wireless LAN, in Proceedings of IEEE INFOCOM, 2003. [6] Y. Wu, Z. Niu, and J. Zheng, Study of the TCP upstream/downstream unfairness issue with per-ow queuing over infrastructure-mode WLANs, Wireless Communications and Mobile Computing, vol. 5, no. 4, pp. 459 471, 2005. [7] R. Bruno, M. Conti, and E. Gregori, Analytical modeling of TCP clients in Wi-Fi hot spot networks, in Proceedings of IFIP Networking, 2004. [8] The network simulator: ns-2. available at http://www.isi.edu/nsnam/ns.

You might also like