IMPROVING FAIRNESS AND THROUGHPUT FOR VOICE TRAFFIC IN 802.11E EDCA C. Casetti and C.-F. Chiasserini CERCOM - Dipartimento di Elettronica, Politecnico di Torino - Italy E-mail: casetti,chiasserini
@polito.it

Abstract— Wireless Local Area Networks (WLANs) using 802.11 technology are expected to become a widespread networking solution to provide both real-time services and data applications. In this paper, we consider a WLAN with infrastructure using the 802.11e distributed, contention-based access scheme, the so-called EDCA. We investigate the EDCA performance in presence of voice and data integrated traffic, and observe the inefficiencies that arise when access differentiation is performed on the basis of traffic type only. We then present a solution that improves both fairness and throughput of real-time traffic in WLANs employing EDCA, and we show the benefits of our proposal through some simulation results obtained with ns-2.

I. INTRODUCTION It can be argued that one of the most promising applications of Wireless LANs (WLANs) based on the IEEE 802.11 technology is the support of Voice-over-IP (VoIP) in corporate networks. Indeed, beside allowing an almost instantaneous deployment of network services without affecting existing infrastructures, WLANs could well be used to carry voice traffic over an IP architecture. However, this raises concerns due to the limited degree of Quality of Service (QoS) support that current 802.11b WLANs can provide. These demands have lead to the development of a new IEEE Draft, the so-called IEEE 802.11e [1], defining the MAC procedures to support LAN applications with QoS requirements, including the transport of voice, audio and video over WLANs. The IEEE 802.11e Draft Standard introduces the Hybrid Coordination Function (HCF) and specifies two access schemes: the Enhanced Distributed Channel Access (EDCA) and the HCF Controlled Channel Access (HCCA) [1]. An overview of the 802.11e access mechanisms can be found in [2]. In this paper, we focus on the performance of EDCA, i.e., the distributed, contention-based access scheme, which involves the use of various Access Categories (ACs). Each AC corresponds to a different access priority, as well as to a different set of parameters to be used for channel contention. The performance of 802.11e-based WLANs are currently under investigation; in particular an analytical study is presented in [3] while some simulation results can be found in [4], [5], [6]. We highlight that in [4], [5], [6] each AC is

associated with a different type of traffic (e.g., data, video or voice), as indicated in the IEEE 802.11e Draft Standard. Also, of particular relevance to our work is the observation made in [4], where the authors note an unfairness between uplink and downlink flows when EDCA is employed in a WLAN with Access Point (AP). The reason is that in a WLAN with stations there are uplink CSMA instances contending with only one downlink CSMA instance (the one at the AP). We notice that this phenomenon mainly affects UDP traffic flows, while it does not have an impact on TCP traffic, since TCP is a closed-loop protocol. Taking this observation as a starting point, in our work we address the problem of unfairness between uplink and downlink traffic under EDCA and in presence of voice and data integrated traffic. We propose a solution that is based on an innovative way to set the EDCA parameters, i.e., according not only to the traffic type but also to the traffic direction. Simulation results, obtained by using ns-2, show that our solution significantly improves fairness and throughput for voice traffic. II. QUALITY OF SERVICE PROVISIONING IN IEEE 802.11 E : THE EDCA FUNCTION In this section, we briefly describe the main features of EDCA in order to make the paper self-contained. The interested reader can refer to [1], [2] for more details on the EDCA. Like the 802.11 Distributed Coordination Function (DCF), EDCA is based on the CSMA/CA scheme and employs the concept of Inter Frame Space (IFS) as well as the backoff mechanism; furthermore it introduces the following innovations. ¯ When an 802.11e station seizes the channel, it is entitled to transmit one or more frames for a time interval named Transmission Opportunity (TXOP); a TXOP is characterized by a maximum duration, called TXOP Limit. ¯ Various Access Categories (ACs) are defined, each of which corresponds to a different priority level and to a different set of parameters to be used for contending the channel. In particular, an 802.11e station

operating under the EDCA function includes up to four MAC queues; each queue corresponds to a different AC and represents a separate instance of the CSMA/CA protocol. A queue employs the following parameters to access the channel: (i) the Arbitration Inter Frame Spacing (AIFS AC
), similar to the DIFS used in DCF, (ii) the Minimum and the Maximum Contention Window (CWmin AC
, CWmax AC
), (iii) and the TXOP Limit AC
. The higher the AC priority is, the smaller the AIFS AC
, CWmin AC
and CWmax AC
are. The larger the TXOP Limit AC
, the greater the share of capacity of the AC. However, the values of CWmin AC
and CWmax AC
have to be carefully chosen so as to avoid high collision probability among traffic flows belonging to the same AC, and the value of AIFS must be at least as long as the DIFS interval (the only exception is for the AP that can use an AIFS 30 s long). ¯ Within every 802.11e station, a scheduler solves virtual collisions among the AC queues, i.e., among the various CSMA/CA instances, by always enabling the queue associated with the highest priority to transmit.




III. SIMULATION SCENARIO We consider a wireless-cum-wired network scenario; the wireless portion is an 802.11e Basic Service Set (BSS) with wireless stations. A fixed an Access Point (AP) and node S is connected to the AP through a wired link of 55 Mb/s speed and propagation delay of 2 ms. This link is overprovisioned so that no packets are dropped at its ends. Within the BSS, each wireless station can directly communicate only with the AP, and we assume that there are no hidden stations. Also, we assume that wireless stations and the AP operate at a data rate of 11 Mb/s. The wireless channel is modeled using an independent error model for each communicating pair of nodes. Thus, given wireless stations, we have independent error models representing the channel between each station and the AP,   independent error models representing the and  channel between each pair of stations. An error model is represented by a three-state discrete-time Markov chain. The Markov chain time slot is equal to the 802.11 aSlotTime parameter. Errors over the channel occur in the states long bad and short bad, while the good state is error-free. Thus, a frame transmission is successful only if the error model is in state good for all slots it takes to the frame to be transmitted, while it fails otherwise. The difference between the long bad and short bad states is the time correlation of errors: the former corresponds to long bursts of errors, the latter to short ones.



TABLE I B YTES TO SEND FOR TCP TRAFFIC SOURCES Flow Length (bytes) 119

179

251

334

428

529

658

948

1650

2861

4706

8015

13681

26641

284454

A. Higher-layer Settings The traffic offered to the network is generated by TCP connections and UDP flows. TCP traffic sources exhibit an on-off behavior. On periods do not have a preset duration but, rather, are determined by the amount of traffic that the source has chosen to send. In particular, in the results shown in the following, TCP traffic sources are modeled so as to represent a client-server interaction, such as Web browsing. A specific TCP source is thus either a client or a server, and the activities of client-server pairs are intertwined, meaning that when one half of the pair is active, the other is silent. If a source is a server, its on period spans the transmission of a file whose length is determined by a uniformly random selection of one of the values in Table I. The values were obtained empirically from real traffic measurements. If a source is a client, the length of the file is selected in a similar fashion, although the random selection is truncated at 2861 bytes (yielding three TCP segments at most). All wireless stations on the simulated WLAN are TCP clients, while node S is a TCP server. The TCP version used in our simulation is NewReno, which is supposed to improve TCP performance during Fast Recovery phases (i.e., it recovers from multiple losses within the same window without waiting for a timeout expiration). The Maximum Segment Size (MSS) of TCP segments is equal to 1000 bytes. The initial congestion window and its maximum value are set, respectively, at 1 and 100 segments. UDP traffic simulates Voice-over-IP (VoIP). The model for VoIP calls mimics a two-speaker interaction: when one speaker, at one end, talks (i.e., sends a talkspurt), the other is silent. The duration of talkspurts is determined by a Paretodistributed random variable with average 4 s and shape factor   ; this corresponds to a set of duration values ranging between 4 and 30 s. It follows that silences have the same distribution. During a talkspurt, the active source sends data following a real-life trace collected at the output of a g.729A vocoder, with an encoded bit rate of 8 Kb/s and an average rate of 2.8 Kb/s resulting from the introduction of VAD (Voice Activity Detection). The framing time is 20 ms. The quality of the recorded performance can be related to voice applications by cross-checking Table II, where delay and loss requirements are listed [7]. Notice that the loss percentage refers to both network losses as well as delay losses, i.e.,





TABLE II V OICE - OVER -IP Q O S REQUIREMENTS Delay (ms)

Losses

good

0-150

medium

150-400








400

poor

TABLE III CW MIN AND CW MAX SETTINGS FOR DIFFERENTIATING ON THE BASIS

total

1000 Throughput [Kbps]

Quality

10000

100 TCP on-off VoIP

10

OF TRAFFIC TYPE ONLY

1

AC

AIFS

CW

AC0

50 s

31:1023

TCP

AC1

50 s

7:511

VoIP

 

uplink downlink

Traffic

frames discarded – either preemptively or at playback time – because they were past their playback time. B. Data Link-layer Settings Only the EDCA scheme is used at the MAC layer. MAC queues at wireless stations have a capacity of 25 Kbytes each, while the capacity of each queue at the AP is equal to 400 Kbytes. UDP-carrying frames are timestamped when they enter the MAC queue. These frames are preemptively discarded before transmission if they exceed a timeout of 400 ms so as to avoid to uselessly forward time-critical data when they are past their playout time. We set the RTS threshold at 400 bytes, i.e., only TCP data packets are transmitted using the RTS/CTS (Request to Send/Clear to Send) handshaking. The Short Retry Limit (SRL) and the Long Retry Limit (LRL) are set to 7 and 4, respectively. The TXOP Limit is set to 2.8 ms for all ACs. IV. PERFORMANCE STUDY AND PROPOSED SOLUTION TO IMPROVE FAIRNESS AND THROUGHPUT Here, we first consider the case where an exact mapping between ACs and traffic type exists. Hence, under data and VoIP traffic, two ACs are defined and associated to the parameter settings reported in Table III. Simulation results suggest that in this case an unfairness between uplink and downlink traffic arises, which prevents us from efficiently exploiting the network resources. To cope with this problem, we propose to differentiate the channel access on the basis of both traffic type and traffic direction. Simulation results show the improvement in performance that is obtained through our simple and innovative solution. A. Differentiation Based on Traffic Type Only Our simulation scenario includes a number of wireless stations, , varying between 4 and 40. All stations are involved

0.1 5

10

15 20 25 30 Number of Wireless Stations

35

40

Fig. 1. Average throughput for TCP and VoIP flows and total system throughput versus the number of wireless stations - Differentiation based on traffic type only

in a TCP client-server connection and in a bidirectional VoIP traffic flow. Results are derived through ns-2, assuming that the average error probability over the wireless channel is equal to 0.01 and the average error burst is 88 ms long. Figures 1–3 present the results obtained by mapping the 802.11e ACs onto the existing traffic types according to the parameter settings reported in Table III. Further results have been obtained by using other CW min and CWmax values, or by differentiating on the basis of the AIFS parameter only, as well as of both the AIFS and the CW parameters. However, the parameters settings in Table III gave the best performance. Figure 1 shows the throughput achieved by TCP clientserver connections and VoIP flows, both on the uplink and on the downlink direction, as well as the overall network throughput. Note that, since TCP sources have an onoff behavior, the per-connection throughput refers to the throughput during on periods only, while the overall throughput is computed as the amount of transport-layer traffic successfully transmitted during the simulation, divided by the simulation time. With the settings described above, it can be clearly seen that the network manages to sustain an acceptable throughput for both types of traffic only as long as the number of wireless stations does not exceed 30. Beyond this number, the network population increases to a point where VoIP traffic on the downlink direction is severely penalized. TCP connections on the uplink direction manage to obtain larger throughputs than on the downlink direction. Indeed, downlink TCP is affected by a larger number of “virtual” collisions within the AP MAC, losing the contention in favor of UDP traffic enjoying higher priority, hence smaller CWs. Figure 2 shows the average delay experienced by frames

TABLE IV CW MIN AND CW MAX SETTINGS WITH UPLINK / DOWNLINK

10

DIFFERENTIATION

Average Delay [s]

1

AC

AIFS

CW

Traffic

AC0

50 s

63:4095

Uplink TCP

31:1023

Downlink TCP

AC1 TCP on-off VoIP

0.1

uplink downlink

AC2 AC3

 50s 50s 50s

7:511

Uplink VoIP

3:255

Downlink VoIP

0.01 10000 total 0.001 10

15 20 25 30 Number of Wireless Stations

35

Fig. 2. Average packet delay for TCP and VoIP flows versus the number of wireless stations - Differentiation based on traffic type only

100

100

TCP on-off VoIP

10 TCP on-off VoIP

10 Loss Probability [%]

1000

40 Throughput [Kbps]

5

uplink downlink

uplink downlink 1 5

1

10

15 20 25 30 Number of Wireless Stations

35

40

Fig. 4. Average throughput for TCP and VoIP flows and total system throughput versus the number of wireless stations - Differentiation based on traffic type and direction

0.1

0.01

0.001 5

10

15 20 25 30 Number of Wireless Stations

35

40

Fig. 3. Packet loss probability for TCP and VoIP flows, versus the number of wireless stations - Differentiation based on traffic type only

that were correctly received; remember that VoIP frames that linger in the send buffer for more than 400 ms are preemptively discarded to avoid unnecessary transmission of late voice samples: thus, their contribution does not appear in the average delay computation. ¿From Figure 2, we observe that the latency of downlink VoIP flows quickly converges toward the 400 ms buffer limit. The higher delays experienced by downlink VoIP (compared to uplink VoIP) are explained by a “contention overhead”: when one of few collisions occur, there is only one downlink CSMA/CA instance versus several uplink CSMA/CA instances. This means that downlink UDP has to draw a new backoff interval for every transmission attempt, while all instances of uplink UDP draw one backoff interval each, and they decrement it in parallel. As a result, uplink UDP traffic has a lower latency

than downlink UDP. The packet loss probability for TCP and VoIP flows, presented in Figure 3, clearly shows that no more than 28 VoIP users can be supported without incurring in unacceptably high losses on the downlink channel. B. Differentiation Based on Traffic Type and Traffic Direction One key observation in the results shown in the previous subsection was that downlink VoIP traffic was enjoying a weaker protection than uplink VoIP traffic. Indeed, all metrics indicated that VoIP users transmitting in the uplink direction could be sustained without severe QoS degradation. Thus, our goal was to restructure the AC settings so that downlink traffic could experience a milder contention overhead, i.e., by imposing smaller CW min and CWmax values for downlink traffic versus uplink traffic. We have set new AC parameters to preserve the CW-based differentiation for TCP and VoIP traffic, while introducing further differentiation depending on the direction of traffic flows. Table IV summarizes the new choices. Simulations using the new settings have highlighted two important results: i) more downlink VoIP users can be supported without compromising the service offered to uplink

10

1.1

1

uplink downlink

0.1

Fairness Index

Average Delay [s]

1

TCP on-off VoIP

0.9

0.8

Traffic type Traffic type and direction

0.01 0.7

0.001

0.6 5

10

15 20 25 30 Number of Wireless Stations

35

40

Fig. 5. Average packet delay for TCP and VoIP flows versus the number of wireless stations - Differentiation based on traffic type and direction

100

10

15 20 25 30 Number of Wireless Stations

35

40

Fig. 7. Fairness index for VoIP flows, versus the number of wireless stations - Comparison between differentiation based on traffic type only and differentiation based on traffic type and direction TABLE V AIFS- ONLY AND AIFS + CW SETTINGS WITH UPLINK / DOWNLINK

TCP on-off VoIP

10 Loss Probability [%]

5

DIFFERENTIATION

uplink downlink

AC

AIFS Only AIFS CW

AIFS & CW AIFS CW

AC0

90 s

31:1023

70 s

63:4095

Uplink TCP

AC1

70 s

31:1023

50 s

31:1023

Downlink TCP

1

0.1

AC2 AC3

  50s 30s

31:1023 31:1023

  50s 30s

Traffic

7:511

Uplink VoIP

3:255

Downlink VoIP

0.01

0.001 5

10

15 20 25 30 Number of Wireless Stations

35

40

Fig. 6. Packet loss probability for TCP and VoIP flows, versus the number of wireless stations - Differentiation based on traffic type and direction

VoIP users; ii) the fairness of per-station throughput is dramatically increased, regardless of the direction of the traffic. Figures 4–6 support our claims regarding the increased downlink support. It can be seen that up to 32 stations now experience delays and losses that qualify their quality between good and medium. We also note that TCP traffic is only marginally affected by the new choice of parameters. The reason is that downlink TCP traffic faces the competition of higher-priority VoIP traffic, as shown in Figure 1, and differentiating on the basis of traffic direction does not solve this issue. The improvement in throughput fairness for voice traffic is depicted in Figure 7. We plotted Jain’s fairness index, which was computed from the throughputs achieved by all

VoIP flows, regardless of whether they were in the uplink or in the downlink direction. The fairness curve for traffic differentiated only by traffic type (labeled as “Traffic type” in Figure 7) starts its downward slope when the number of stations is relatively small, a trend which is caused by the growing performance gap between uplink and downlink stations. Differentiating also by direction (see the curve labeled as “Traffic type and direction” in Figure 7) yields an almost constant fairness index even for large number of stations, which indicates that the new choice of parameters equally supports downlink and uplink traffic. For the sake of completeness, we also present loss percentage results obtained with other parameter choices, all yielding traffic differentiation on the uplink and downlink channels. In particular, we used the AIFS-only as well as the AIFS+CW parameters, shown in Table V. Our aim is to confirm that CW-only differentiation represents the best choice. Indeed, Figure 8 shows a very poor performance of downlink VoIP traffic when only AIFSs are different, while Figure 9 highlights fairness problems that appear to favor downlink VoIP traffic with respect to uplink VoIP traffic, when both AIFS and CW are different.

R EFERENCES

100

TCP on-off VoIP

Loss Probability [%]

10

uplink downlink

1

0.1

0.01

0.001 5

10

15 20 25 30 Number of Wireless Stations

35

40

Fig. 8. Packet loss probability for TCP and VoIP flows, versus the number of wireless stations - AIFS differentiation based on traffic type and direction 100

TCP on-off VoIP

Loss Probability [%]

10

uplink downlink

1

0.1

0.01

0.001 5

10

15 20 25 30 Number of Wireless Stations

35

40

Fig. 9. Packet loss probability for TCP and VoIP flows, versus the number of wireless stations - AIFS + CW differentiation based on traffic type and direction

V. CONCLUSIONS We studied QoS provisioning in WLANs using the 802.11 technology. We considered a WLAN with infrastructure where stations employ the distributed, contention-based access scheme (EDCA) under a combination of voice and data traffic. Simulation results obtained through ns-2 showed that mapping access priorities onto traffic types leads to some inefficiencies for voice traffic, in terms of both fairness and throughput. We then proposed a solution based on an innovative way to set the EDCA parameters, which significantly improves the performance of voice traffic. VI. ACKNOWLEDGMENTS This work was supported by a contract between Politecnico di Torino and Alcatel France.

[1] IEEE 802.11 WG, Draft Supplement to Standard for Telecommunication and Information Exchange between Systems – LAN/MAN Specific Requirements – Part II: Wireless Medium Access Control (MAC) and Physical (PHY) Layer Specifications: MAC Enhancements for Quality of Service, IEEE 802.11e Draft 5.0, July 2003. [2] S. Mangold, S. Choi, G. R. Hiertz, O. Klein, and B. Walke, “Analysis of IEEE 802.11e for QoS Support in Wireless LANs,” IEEE Wireless Communications Mag., Dec. 2003, pp. 40–50. [3] G. Bianchi and I. Tinnirello, “Analysis of Priority Mechanisms Based on Differentiated Inter Frame Spacing in CSMA-CA,” IEEE VTC-Fall, Orlando, FL, Oct. 2003. [4] A. Grilo and M. Nunes, “Performance evaluation of IEEE 802.11e,” IEEE PIMRC’02, Lisboa, Portugal, Sep. 2002. [5] S. Choi, J. del Prado, S. Shankar N, and S. Mangold, ”IEEE 802.11e Contention-Based Channel Access (EDCF) Performance Evaluation,” IEEE ICC 2003, Anchorage, Alaska, May 2003. [6] A. Lindgren, A. Almquist, and O. Schel´en, “Quality of Service Schemes for IEEE 802.11 Wireless LANs - An Evaluation,” MONET, Special Issue on Performance Evaluation of Qos Architectures in Mobile Networks, Vol. 8, No. 3, June 2003, pp. 223–35. [7] R. Caputo, CISCO Packetized Voice & Data Integration, McGraw-Hill, 2000.