Game-theoretic EDCA in IEEE 802.11e WLANs

Liqiang Zhao, Li Cong and Hailin Zhang State Key Laboratory of Integrated Services Networks

Xidian University Xi’an, Shaanxi, China

lqzhao@mail.xidian.edu.cn

Wei Ding and Jie Zhang Centre for Wireless Network Design

University of Bedfordshire Luton, Bedfordshire, UK

Abstract—In IEEE 802.11 WLANs, the fundamental MAC mechanism – DCF, only supports best-effort service, and is unaware of QoS. IEEE 802.11e EDCA supports service differentiation by differentiating contention parameters. This may introduce the problem of selfish traffics with rational nodes. Hence, Game-theoretic EDCA (G-EDCA) is proposed in this paper to solve the problem. Moreover, a simple and run-time estimation mechanism is presented to implement G-EDCA in current WLANs. Extensive simulations are also carried out to compare the performances of DCF, EDCA and G-EDCA. The simulation results show that G-EDCA performs better than DCF and EDCA in terms of system throughput and QoS.

Keywords- WLAN; MAC; QoS; Game Theory

I. INTRODUCTION In recent years, wireless LANs (WLANs) have been widely

used as one of the essential technologies to provide broadband wireless access, and the performance analysis and improvement of WLANs have attracted much research interests [1]. IEEE 802.11 is a widely adopted WLAN standard. Its fundamental medium access control (MAC) mechanism – distributed coordination function (DCF), is based on carrier sense multiple access with collision avoidance (CSMA/CA), a typical contention-based MAC protocol. DCF only supports best-effort service, and does not ensure any quality-of-service (QoS) requirement. However, there is an increasing demand that multimedia services with different QoS requirements be supported by WLANs. Different multimedia services have different QoS requirements with respect to bandwidth, delay, jitter and packet-loss-rate. In order for WLANs to support multimedia applications, the MAC protocol must support a degree of service differentiation. In DCF, all the traffic classes have the same contention parameters. Hence, all the services have the same priority when accessing the channel. One possible solution is to provide a good priority scheme by differentiating the contention parameters, thereby achieving service differentiation [2-7]. The IEEE 802.11e standard introduces enhanced distributed channel access (EDCA) to support QoS applications. In EDCA, the traffic which is assigned a smaller contention window has a higher priority to contend for the channel. However, all the above protocols have not considered the selfish behaviors of nodes.

Due to the selfish behaviors of nodes, the system performance would be significantly degraded [8-11]. The current Internet is vulnerable to the greedy TCP end-point behavior, having either low network throughput or high drop-rates, or both [8]. In IEEE 802.11e selfish nodes can gain

enhanced performance by selectively classifying low priority traffic as high priority, potentially destroying the QoS capability of the system [9]. [8-11] assumes that the selfish nodes only prepare to harm other honest nodes if and when they can derive a benefit from this misbehavior. However, in the real scenarios, there is no incentive for a node to “cheat” the system, as it will be disciplined more severely. On the other hand, the selfish behavior of nodes in CSMA/CA networks always results in a network which provides an incentive for cheaters to cooperate with each other [10-11]. So in this paper all the users are assumed to be rational ones [12-13], i.e., they play according to the predetermined MAC protocols. However, in EDCA, the high priority traffic (e.g., voice and video applications) with a lower contention window will enjoy greater throughput and less delay, compared to the low priority traffic (e.g., file transfer). That is to say, the traffics are selfish. In such a selfish manner, a smaller contention window may cause more collisions, and significantly affect the system performance. However, all the rational nodes would try their best to support the selfish traffics and get help from the system, e.g., the service differentiation-based MAC protocol. To support service differentiation, EDCA integrates selfish traffics with rational nodes, and causes the problem of selfish traffics with rational nodes. However, the existing research does not consider this problem. A game-theoretic EDCA (G-EDCA) protocol was very briefly presented as a letter in [14]. In [14], the incompletely cooperative game is introduced into G-EDCA, and each node tunes its equilibrium strategy based on the estimated game state. However, [14] did not present any algorithm to estimate the game state. In this paper, the preliminary results presented in the letter will be revised and substantially extended, e.g., a simple and run-time estimate mechanism is presented, and the performance of G-EDCA is evaluated in much greater details.

The rest of this paper is organized as follows. Game theory, DCF, and EDCA are introduced in section II. In section III, all the aspects of G-EDCA are discussed in detail. In section IV, simulation studies are carried out to evaluate the performance of G-EDCA. The concluding remarks are given in Section V.

II. PRELIMINARY

A. Description of DCF and EDCA In DCF if a node has a new packet to transmit, it will

monitor the channel activities first. If the channel has been idle for a period of time, i.e. a distributed IFS (DIFS), the node will transmit. Otherwise, if the channel is sensed busy, the node will continuously monitor the channel until it is found to be idle for

This work is supported by the 111 Project (B08038), National Natural Science Foundation of China (No. 60772137), Natural Science Basic Research Plan in Shaanxi Province of China (Program No. 2006F30), and the EU FP6 “GAWIND” project on broadband wireless access network design under the grant number MTKD-CT-2006-042783.

978-1-4244-1722-3/08/$25.00 ©2008 IEEE. 1

a DIFS. At this point, the node generates a random backoff interval before transmitting. The backoff time is slotted and uniformly chosen in the range [0, CW-1]. At the first transmission attempt, the contention window, CW, is set equal to a value CWmin called the minimum contention window. After each unsuccessful transmission, CW is doubled up to the maximum value CWmax = 2mCWmin. The value m is called the maximum backoff stage, and CWmax is called the maximum contention window. Once CW reaches CWmax, it will remain at the value until the packet is transmitted successfully or the retransmission time reaches the retry limit (r). While the limit is being reached, retransmission attempts will cease and the packet will be discarded.

EDCA is one of the main parts of the IEEE 802.11e standard for service differentiation. In EDCA traffic flows are categorized into four access categories (ACs). Different ACs adopt different values of arbitrary interframe space (AIFS), CWmin, m and transmission opportunity (TXOP) for acquiring channel access. For example, before transmitting, each node needs to wait for a period of time called AIFS for the channel to be idle. The value of AIFS is associated with the corresponding AC. TXOP is the time interval permitted for a particular node to transmit packets. During the TXOP, there can be a series of frames transmitted by a node. Typically, a shorter AIFS and a longer TXOP is associated with an AC with a higher priority.

B. Description of Game Theory Game theory is a powerful tool to study situations of

conflict and cooperation. This theory is concerned with finding the best actions for individual decision makers (i.e., players) in these situations and recognizing stable outcomes [15]. Games may generally be categorized as non-cooperative and cooperative games. Non-cooperative game theory is concerned with the analysis of strategic choices and explicitly models the decision making process of a player in its own interests. Unlike in non-cooperative games, in cooperative games, the players can make binding commitments. Recently, game theory has become a powerful tool to analyze and improve the performance of WLANs [16]. However, to use the game effectively in WLANs, much attention should be paid to the relevant context of DCF or EDCA. For example, explicit cooperation is practically impossible in WLANs. So a novel concept of incompletely cooperative game theory is presented for DCF [17].

III. DESCRIPTION OF G-EDCA In G-EDCA, first, each player (i.e., node) estimates the

current game state according to its history; second, the player adjusts its own equilibrium strategy by tuning its local contention parameters, and then transmits its packets [17]. As all the nodes take actions simultaneously, i.e., transmitting their packets or not, the player does not know which action the other nodes (i.e., its opponents) are actually taking. However, the player can predict its opponents’ actions according to its history, and make the optimal decisions.

The game state includes not only the player’s local states, such as signal-to-noise ratio [16], but also its opponents’ states, such as the number of competing nodes of each AC; the former

can be retrieved from DCF or EDCA operation and the latter cannot. Several performance evaluation studies [17] show that the performance of DCF is very sensitive to the number of nodes competing on the channel, i.e., the number of nodes that are simultaneously trying to send a packet on the shared medium. However, to our best knowledge, no one estimates the number of nodes of each AC. For simplicity, the following will only discuss how the player estimates the number of competing nodes of each AC.

A. Estimation Mechanism In this paper, four subtypes of data-frames are proposed in

order to support the corresponding ACs. The four data-frames have the same MAC frame format except the subtype field in the MAC layer header, which are 0011, 0100, 0101 and 0110 respectively, as shown in Fig. 1. These four values have been reserved in IEEE 802.11x.

Figure 1. Format of data-frames in IEEE 802.11x

During the operation of G-EDCA, as all the frames are broadcasted to the destination node, all the other neighboring nodes can obtain the source address (SA) and the subtype field (SF) from the received data-frames. By counting the number of the corresponding data-frames with different SAs, the player can estimate the number of competing nodes of ACi (ni), where i denotes a given AC (0 ≤ i ≤ 3). The estimation process includes the following four main steps:

1. To define a variable ni to count the number of competing nodes of ACi.

2. To define an array state[J][K] to record the SAs and SFs of the received frames, and its relevant time values, e.g., state[1][1] records a SA, state[1][2] records a SF, state[1][3] records the latest frame arrival time from relative SA and SF.

3. When the player receives a new frame, it will compare the SA and SF value of the frame with the recorded SAs and SFs in state[j][1] and state[j][2] (1 ≤ j ≤ J) respectively. If this is a new SA or SF, the player will record the SA and SF value and its relevant time in the array state[J+1][1], state[J+1][2] and state[J+1][3] respectively. In addition, if this is a new SA, the player will increase ni by one. Otherwise, it will update the relevant time of the SA and SF in the array.

4. To scan the whole array. If a SA is found to be out of date, the player will delete this item and decrease ni by one. If a SF is found to be out of date, the player will delete this item.

Now the question is how to determine if an SA or an SF in the array is out of date, that is, how to set the value of the SA or SF survival time LifeTime. With a smaller LifeTime, the item in the array will be deleted earlier; but with a larger LifeTime, the player will be less sensitive to the changing size of the network.

2

Suppose a node has a frame to send, if the transmission fails, it should get another chance after a time called ∆t to transmit the same frame; if the transmission fails again, then the node should transmit again after another ∆t time until it reaches the maximum retry limit r. So for simplicity, the survival time for the SA of state[j][1] and the SF of state[j][2] is set as:

[ ][ ] ( )23

−×=∆×= ∑=

KkjstatertrLifeTimeK

k

, (1)

where the arrays from state[j][3] to state[j][K] record the latest K-2 frame arriving intervals from the traffic flow with the SA of state[j][1] and the SF of state[j][2]. Corresponding to a SA, if no frame is transmitted from this node when its LifeTime has elapsed, it is assumed that this node has left the WLAN. Corresponding to a SF, if no frame is transmitted from this node when its LifeTime has elapsed, it is assumed that this traffic flow (i.e., AC) has been closed.

Of course, the estimated values need to be updated periodically. Discussions in detail of the run-time estimation methods are beyond the scope of this paper.

B. Backoff Strategy After obtaining the number of competing nodes of ACi (ni),

each node tunes the value of its minimum contention window as follows [17]:

( )[ ] 10,2,38,7min, =×= irandnCW ii , (2)

where rand (x, y) returns a random value between x and y, [z] is the largest integer that is not more than z.

( ) =

=− elseCWCW

jCWCW

iji

iji

max,1,

min,, ,2min

0, (3)

where min(x, y) returns the smaller one of x and y. The maximum contention window of ACi, min,max, 2 i

mi CWCW i ⋅= ,

where the maximum backoff stage of ACi (mi) is defined in IEEE 802.11e.

Then the current backoff time is defined as:

( )upjilowjiji WWrandBackoff ,,,,, ,= , (4)

where the lower limit

=

=== ∑∑

+=+=

0,1,2

303

1,

3

1,,, iCW

iCWW

ikjk

ikjklowji , (5)

and the upper limit

∑=

=3

,,,ik

jkupji CWW . (6)

In this way, the transmission of the lower priority traffic is prevented from disturbing the transmission of the higher priority traffic.

Obviously, this is a cooperative game. As all the nodes can obtain the number of competing nodes of each AC, they make a binding commitment and thus play cooperative games among themselves. On the other hand, given that the higher priority traffic (i.e., ACi) set a smaller contention window, the best strategy for the lower priority traffic (i.e., ACj where i > j) is to set a larger contention window in order to make effective use of the channel.

IV. SIMULATION RESULTS To evaluate the performance of G-EDCA, extensive

simulations are carried out with NS-2 in the following. The values of the parameters used to obtain numerical results for simulations are specified in IEEE 802.11b. The channel is assumed to be ideal. The channel data rate is fixed at 11Mb/s. Assume there are 40 nodes and each node belongs to one AC. Traffic flows are classified into four ACs with equal number of nodes (i.e., n3 = n2 = n1 = n0). For simplicity, the frame will be discarded only as the retransmission time reaches the retry limit. Each node generates new packets under a Poisson process. The packet arrival rate is initially set to be lower than the saturation case, and it is subsequently increased so that, at the end of the simulation time, all nodes are almost in saturation conditions. The payload sizes for all ACs are fixed at 1023 bytes. For comparison, performance of G-EDCA, EDCA and DCF are compared under the same conditions.

To evaluate the performance of the estimation mechanism, the ideal performance of G-EDCA is simulated too, where it is assumed that each node has known the number of competing nodes of each AC. For simplicity, the two cases are named as Real G-EDCA and Ideal G-EDCA respectively.

When the load is light, Figs. 2-9 show that the performance of DCF, EDCA and G-EDCA are similar. When the load is heavy, they have distinct performances in ensuring QoS. In the following, we will focus on the comparison of saturation performance of the three protocols.

Fig. 2 shows that the saturated system throughput of G-EDCA is larger than that of DCF, which is much larger than that of EDCA. The system throughput of EDCA decreases very sharply under the large network load as collisions increase. In EDCA, to support service differentiation, compared with DCF, the high priority traffic has a smaller CW, which causes more collisions when the number of competing nodes is large. The more collisions there are, the more sharply the system throughput decreases. In G-EDCA, based on the incompletely cooperative game, each node adjusts its CW to the changing traffic loads (i.e., the number of competing nodes of ACs), so the system throughput of G-EDCA almost keeps constant.

Fig. 3-4 shows that in G-EDCA throughput of the high priority traffic (e.g., AC3 and AC2) is much larger than that of

3

the low priority traffic (e.g., AC1 and AC0). Throughput of AC3 and AC2 keep almost constant, and that of AC1 and AC0 decrease with the increasing network load. It indicates G-EDCA supports absolute service differentiation. Fig. 5 shows that in EDCA throughout of all ACs decrease with the increasing network load. It indicates EDCA supports relative service differentiation. In DCF throughput of each AC has the same throughput as shown in Fig. 6. It indicates DCF does not support service differentiation.

Figure 2. System throughput

Fig. 7-8 shows that the average access delay and jitter in G-EDCA are lower than those in EDCA.

Fig. 9 shows that the average packet-loss-rate of in EDCA increases very sharply due to the problem of selfish traffics with rational nodes.

Figure 3. Throughput of ACs in Real G-EDCA

In general, in G-EDCA each node adjusts its equilibrium strategy to the estimated game state, especially QoS requirements, so G-EDCA can solve the problem of selfish traffics with rational nodes in EDCA, and performs much better than EDCA in terms of system throughput and QoS.

Figure 4. Throughput of ACs in Ideal G-EDCA

Figure 5. Throughput of ACs in EDCA

Figure 6. Throughput of ACs in DCF

4

Figure 7. Access delay

Figure 8. Jitter

Figure 9. Packet loss rate

V. CONCLUSION In this paper, game theory is used in WLANs to solve the

problem of selfish traffics with rational nodes. To our best knowledge, no one has considered this problem. In G-EDCA, to implement the optimal equilibrium strategy, all the nodes play cooperative games based on the estimated game state, e.g., the number of competing nodes of each traffic priority. Moreover, a simple and run-time estimation algorithm is presented. The simulation results show that G-EDCA is an appropriate tool to improve throughput, and to decrease delay, jitter and packet-loss-rate.

REFERENCES [1] B. H. Walke, S. Mangold, and L. Berlemann: “IEEE 802 Wireless

Systems”, John Wiley & Sons, Ltd, 2006. [2] Hua Zhu, Ming Li, Imrich Chlamtac, B. Prabhakaran: “A survey of

quality of service in IEEE 802.11 networks”, IEEE Wireless Communications, Vol. 11, No. 4, Aug. 2004, pp. 6-14.

[3] Yang Xiao: “Performance Analysis of Priority Schemes for IEEE 802.11 and IEEE 802.11e Wireless LANs,” IEEE Transactions on Wireless Communications, Vol. 4, No. 4, July 2005, pp. 1506-1515.

[4] Engelstad P.E., Osterbo, O.N.: “Analysis of the Total Delay of IEEE 802.11e EDCA and 802.11 DCF,” IEEE International Conference on Communications (ICC 2006), Vol. 2, June 2006, pp. 552 – 559.

[5] Zhifeng Tao, Panwar, S., “Throughput and delay analysis for the IEEE 802.11e enhanced distributed channel access,” IEEE Transactions on Communications, Vol. 54, No. 4, April 2006, pp. 596 – 603.

[6] Jie Hui, Devetsikiotis, M.: “A unified model for the performance analysis of IEEE 802.11e EDCA,” IEEE Transactions on Communications, Vol. 53, No. 9, Sept. 2005, pp. 1498 – 1510.

[7] Naomi Ramos, Debashis Panigrahi, and Sujit Dey: “Quality of service provisioning in 802.11e networks: challenges, approaches, and future directions”, IEEE Network, July/August 2005, pp. 14-20.

[8] A. Akella, S. Seshan, R. Karp, et al: “Selfish behavior and stability of the Internet: a game-theoretic analysis of TCP”, SIGCOMM’02, August 2002, USA.

[9] S. R. Sandeep, and P. Nuggehalli, “QoS with selfish nodes in wireless networks”, COMSWARE’06, 2006.

[10] A. B. Mackenzie and S. B. Wicker, “Stability of multipacket slotted Aloha with selfish users and perfect information”, IEEE INFOCOM 2003.

[11] M. Cagalj, S. Ganeriwal, and I. Aad, et al., “On selfish behavior in CSMA/CA networks,” IEEE INFOCOM, Miami, FL, Mar. 2005.

[12] J. Shneidman, D. C. Parkes, “Specification Faithfulness in Networks with Rational Nodes,” PODC’04, July 25-28, 2004, St. Johns, Newfoundland, Canada.

[13] J. Shneidman, D. C. Parkes, “Rationality and Self-Interest in Peer to Peer Networks,” Proceedings of the Second International Workshop on Peer-to-Peer Systems (IPTPS’03), February 2003, Berkeley, CA.

[14] L. Zhang, H. Zhang, J. Zhang, and H. Zhang, “Selfish traffic with rational nodes in WLANs,” to appear in IEEE Communications Letters.

[15] P. D. Straffin, “Game Theory and Strategy,” Math. Assn. of America, 1993.

[16] Y. Xiao et al, “Game theory models for IEEE 802.11 DCF in wireless ad hoc networks,” IEEE Communications Magazine, March 2005.

[17] L. Zhao, J. Zhang and H. Zhang, “Using Incompletely Cooperative Game Theory in Wireless Mesh Networks,” IEEE Network, January/February 2008, Vol. 22, No. 1, pp. 39-44.

5