IEEE 802.11e Enhanced Distributed Channel Access (EDCA) Throughput Analysis Haitao Wu1 , Xin Wang2 , Qian Zhang3 and Xuemin (Sherman) Shen4 1
Microsoft Research Asia (MSRA), China,
[email protected] Tsinghua University, China,
[email protected] 3 Hong Kong University of Science and Technology, China,
[email protected] 4 University of Waterloo, Waterloo, Ontario, Canada,
[email protected] 2
Abstract—The upcoming standard IEEE 802.11e will provide Quality of Service (QoS) support in Wireless Local Area Network (WLAN). The contention based channel access method called Enhanced Distributed Channel Access (EDCA) is considered as the mandatory mode for Medium Access Control (MAC) in IEEE 802.11e. This paper presents an accurate throughput analytical model for EDCA in saturated situation. The analytical model is suitable for both basic access and RTS/CTS access mechanisms. It considers the features in EDCA such as different Arbitration Inter-frame Space (AIFS) and contention window for different Access Category (AC), and virtual collision for different priority queue in the same station; and can be easily extended to other features such as Transmission Opportunity (TXOP), etc. The analytical model is evaluated by extensive simulation results, and it provides deep understanding of the effect by different parameter setting to the throughput. Keywords- Enhanced Distributed Channel Access (EDCA), IEEE 802.11e, Analysis, Wireless Local Area Network (WLAN)
I.
INTRODUCTION
It has been experiencing a tremendous growth in recent years for WLAN as evidenced by the fast increasing popularity of WLAN hotspots deployed in residence, enterprise and public areas such as airports, campuses, etc. WLAN services are evolving from the best effort data services to real-time applications with a certain level of quality-ofservice (QoS) provisioning. To support QoS over WLAN, the standard IEEE 802.11e [2] is being proposed. In IEEE 802.11e, hybrid coordination function (HCF) is used as the MAC access method. It combines the contentionbased EDCA and contention-free HCF controlled channel access (HCCA) to provide QoS stations (QSTAs) with prioritized access. EDCA defines the prioritized carrier sense multiple access with collision avoidance (CSMA/CA) mechanism, and it is the mandatory mechanism in 802.11e. In this paper, we focus on a complete and accurate average throughput analysis achieved in the saturated situation by modeling the EDCA. Since 802.11e EDCA adopts most of the design principles in 802.11 DCF, such as exponential backoff, basic and *This work is finished when the second author was an intern at MSRA
RTS/CTS access method, etc., it is natural to leverage the modeling work for 802.11 DCF. In the literature, there have been extensive efforts to model 802.11 since the standard was proposed. In [4], a Markov model is developed to analyze the saturated throughput of 802.11, which captures the effect of the Contention Window (CW) and binary exponential backoff procedure used by DCF. In [5], the Markov model is refined by considering the frame retry limits and timeout after collisions. All of these models have been applied to the legacy DCF in 802.11, and can be extended to the scenario with different packet size and channel modulation rate. The new features of 802.11e EDCA [2] to provide QoS challenge the model, e.g., different Arbitration Inter-frame Space (AIFS) and CW for different Access Category (AC), and virtual collision for different priority queue in the same station. There are a lot of research works on 802.11e EDCA appeared in the literature. Based on the Markov model for saturated condition [4,5], an analytical approach is proposed for throughput and delay of 802.11e EDCA with different CW and retry factor for each AC [3,10]. In [6], a three-dimensional Markov model is developed for analyzing the performance of EDCA on AIFS and CW differentiation, but only the approximated results of iteration can be obtained. Recently, a concise thi-dimensional Markov model is proposed in [7] for EDCA. It partially gives the analysis for the effect of AIFS by introducing additional slot corresponding to the differentiation of AIFS as in [6]. Neither of them takes the different packet collision probability experienced by different AC into consideration. To address this issue, a novel notation called contention zone is introduced in [8], where the different collision probability for different AC in a random slot is analyzed by calculating the distribution for the location of the slot with last channel busy time. However, the calculation of the probability for contention zone in [8] does not take virtual collision into consideration. In addition, the analysis for the post collision period in [8] assumes a much less timeout value than that defined in standard [2] for stations in collisions. Therefore, all the previous works only partially address the challenges for modeling EDCA, which motivate our research on the modeling of EDCA. In this paper, we propose a Markov Chain based model for each AC, and differentiate the collision probability experienced by different AC through the refined contention
1-4244-0355-3/06/$20.00 (c) 2006 IEEE This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2006 proceedings.
zone. Different CW for each AC and the virtual collision defined in EDCA are also taken into account. Other factors such as different packet length, modulation rate, and Transmission Opportunity (TXOP) can be analyzed straightforwardly by extending our model. Our contribution is to address all the factors related to QoS introduced in EDCA and target at an accurate model. The remainder of this paper is organized as follows. In Section II, we briefly introduce EDCA. In Section III, we propose a Markov chain based model with contention zone [14] to analyze the average throughput performance of EDCA, taking all the QoS related factors in EDCA into consideration. The accuracy of our model is evaluated by extensive simulation results in Section IV. We conclude our paper in Section V. II.
Therefore, the backoff counter of low priority AC will be decreased much slower than that of the high priority AC. An interesting observation from this example is that, since the low priority AC can not access the channel in the interval introduced by AIFS difference, different AC experiences different channel busy probability, which makes AC with high priority beneficial. Most of the modeling [6,7] for 802.11e EDCA do not take this effect into account except that the novel notation of contention zone in [8].
IEEE 802.11 EDCA
In 802.11e EDCA, service differentiation is provided by assigning different contention parameters to different AC. A QoS station can support at most eight user priorities, which are mapped into four ACs. Each AC contends channel access with different AIFS and CW setting. Compared with DCF where DIFS is used as the common IFS for a station to access the channel, EDCF uses different AIFS for each AC to achieve the access differentiation, where the AIFS for a given AC is defined as AIFS[ AC ] = AIFSN [ AC ] × δ + SIFS .
The AIFSN denotes the number to differentiate the AIFS for each AC, and δ is the time interval of a slot for 802.11 standard, which is determined according to the physical medium used. Table 1 shows the default parameter settings defined for different ACs in 802.11e draft standard [2], where AC1 for voice is assigned the highest priority while AC4 for background is given the lowest priority. AC AC_VO(Voice) AC_VI(Video) AC_BE(Best Effort) AC_BK(Background)
respectively. The difference of AIFSN is 5, so the AC1 in STA1 will decrease its backoff counter 5 slots earlier than AC4 in STA2. In addition, the backoff counter of high priority AC may count to zero in this interval and transmit the packet, which results in channel busy due to high priority packet transmission and resynchronization after that.
CWmin 7 15 31 31
CWmax 15 31 1023 1023
AIFSN 2 2 3 7
Table 1 Default EDCA parameter set
Fig.2 Station with multiple priority queues
In a single QoS station supporting EDCA, each AC is implemented as a separate queue, as shown in Fig. 2. Each queue behaves like a virtual station and contends for the channel access independently. When a collision occurs among different queues of the same station, i.e., two backoff counters of the queues decrease to zero simultaneously, the highest priority queue always wins the contention, and the lower priority queues act as if a collision occurred. In the following, we develop a Markov chain based model to analyze the saturated throughput and channel access time. Note 802.11e EDCF also introduces the TXOP limit parameter allowing a transmission burst for different AC in contrast to a common restriction for one packet as in DCF, which can be analyzed by extending our model using similar methods proposed in [4,9] for different packet size and modulation rate. III.
Fig.1 Channel access in EDCA
To understand the service differentiation introduced by AIFS and CW, we use an example shown in Fig. 1, where there are two stations with packets in AC1 and AC4,
ANALYTICAL MODEL
Our analytical model for EDCF is based on Markov chain in [4,5]. We consider the saturated condition, i.e., if a station has packet arrival in a given AC, the AC will always have packets waiting for transmission. We also assume a perfect physical channel, i.e., we do not consider wireless transmission error, fading, capture effect, etc. By the Markov chain model, we obtain the relationship of collision probability and different contention window. With the transmission probability differentiation in contention zone, the difference in AIFS for different AC can be modeled. We combine the equations for Markov chain and contention zone with virtual
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2006 proceedings.
collision, and solve them numerically. With the calculated transmission probability and contention probability, we obtain the throughput for different AC.
0,W0[ AC ] − 2
1 − p[ AC ] p[ AC ]
0,W0[ AC ] − 1
] τ [ AC ] = ∑i =0 bi[,AC = 0 m
1 − ( p[ AC ] ) m+1 [ AC ] , b0 ,0 1 − p[ AC ]
(1)
] where bi[,AC is the stationary distribution of the Markov chain k ] in state (i, k) for AC respectively, and b0[ ,AC can be represented 0
by an equation of p[ AC ] using
[ AC ] −1 m Wi
∑ ∑b
[ AC ] i ,k
i=0
= 1 . So we can
k =0
figure out an equation for p[ AC ] by the conditional transmission probability τ [ AC ] respectively.
1 − p[ AC ] p[ AC ] i, Wi[ AC ] − 2
1 − p[ AC ]
i, Wi[ AC ] − 1
p[ AC ] p[ AC ] m, Wm[ AC ] − 2
m, Wm[ AC ] − 1
Fig 3. Markov chain model for each AC
A. Packet Transmission Probability The fundamental assumption for Markov chain based approach is that each station will transmit with a stationary probabilityτ in a generic (i.e., random chosen) slot, and each station collides with constant probability p regardless of the number of retransmissions suffered. However, the difference in AIFS of each AC in EDCA challenges such assumption since the number of competing stations may vary at different time slots and the collision probability can not be simply assumed as constant. Therefore, we use separate Markov chain for each AC, and the collision probability is assumed as an average value p[ AC ] for each AC, respectively. We also assume that each queue behaves as a virtual station, and it initiates a transmission in a random chosen slot available for this AC with a stationery conditional probability τ [ AC ] . Let b(t) be the stochastic process representing the backoff window size for a given station at slot time t, and the backoff stage is denoted as s(t). A discrete and integer time scale is adopted, where t and t+1 denote the two consecutive slot times, and the backoff counter of each queue decreases at the beginning of each slot time. Note different AC will experience different slot time due to the difference in AIFS. Thus, for each AC, the bi-dimensional process (s(t), b(t)) is a discretetime Markov chain, which is shown in Fig. 3. Here Wi[ AC ] − 1 denotes the contention window of the AC at backoff stage i, and m denotes the maximal backoff stage. Each backoff will double the value of Wi[ AC ] until the contention window is equal to the CWmax of the corresponding AC. The formation of the state transition in the Markov chain is similar to that in [5]. In the following, we briefly explain the formation procedure. The conditional transmission probability τ [ AC ] can be deduced as the sum of the probability
To solve the equation, the key issue is to obtain another formation for different collision probability p[ AC ] with conditional probability τ [ AC ] , which is introduced to address the difference of AIFS for each AC. Thus, the difference of AIFS creates different contention zones for AC to differentiate the access probability, which is illustrated in Fig. 4 by an example for the differentiation for two ACs, where AC-A stands for high priority and AC-B stands for low priority. We assume that for each queue, the probability that any transmission experiences a collision is constant within a contention zone, regardless of the number of retransmissions suffered. Comparing with the work in [8] for contention zone, we take virtual collision into consideration in calculating the collision probability. AIFS A
Zone 1
Zone 2
p A, zone1
p A, zone2
τA AIFS B
pB , zone2
τB
Fig. 4 Transmission probability in contention zones
To understand the effect of contention zone, we take two priority queues with different AIFS settings as an example. Note our analysis can be extended to the scenario with more priority queues as defined in standards. As shown in Fig. 4, in contention zone 1, only high priority queues contend for the channel access; while in contention zone 2, high priority queues contend with the low priority queues. In other words, a low priority queue never ‘sees’ contention zone 1. And due to virtual collision resolution, a high priority queue never ‘sees’ low priority queue in the same station. Under the saturated condition, the maximal slot number for all contention zones is L1 + L2 = min(CWmax [ AC ]) , where Lj denotes the number of slot in zone j. The transmission probability for a packet in each AC is restricted by the contention zones available for this AC, so it is a conditional probability. For example, τ B ,zone2 is the conditional transmission probability given that the random chosen slot is in contention zone 2 since a low priority queue can only access the medium in zone 2.
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2006 proceedings.
q1
1 − q1
q1
1
L1
2
1 − q1
L1 + L2
L1 + 1
1 − q1
For the case with two ACs, it can be simplified as
q2
(7)
Let π zone ( j ) denote the stationary distribution for a random slot in zone j, it can be calculated as
1 − q2 1
L1
1 − (q1 ) L1 +1 1 − (q2 ) L2 −1 −1 . z1 = [ + q1L1 q2 ] 1 − q1 1 − q2
L2
π zone( j ) =
∑z
(8)
k
z (k )= j
Fig.5 Transition of slots in contention zones
To calculate the stationary probability for the slot in each contention zone, we assume that transition of slots can be approximated by a Markov chain. Fig. 5 shows the relationship between different backoff slots for two ACs, where we use q j to denote the probability that no stations transmit in contention zone j. There are n stations in competition for channel access, and among them, nA stations have packets in AC A and nB stations have packets in AC B. Let µi, j denote the transmission probability for station i (1
