A CSMA-based MAC protocol for WLANs with automatic synchronization capability to provide hard quality of service guarantees

Computer Networks 127 (2017) 31–42

Contents lists available at ScienceDirect

Computer Networks journal homepage: www.elsevier.com/locate/comnet

A CSMA-based MAC protocol for WLANs with automatic synchronization capability to provide hard quality of service guarantees Yaw-Wen Kuo∗, Jane-Hwa Huang Department of Electrical Engineering, National Chi Nan University, Nantou, Taiwan, ROC

a r t i c l e

i n f o

Article history: Received 15 April 2016 Revised 12 July 2017 Accepted 17 July 2017 Available online 27 July 2017 Keywords: Wireless LAN Quality of service Medium access control protocol

a b s t r a c t The carrier sensing multiple access with collision avoidance (CSMA/CA) protocol is a widely-adopted MAC protocol in the current wireless networks, but the quality of service (QoS) cannot be guaranteed due to random access. Investigations reveal that the collision avoidance mechanism which relies on the binary exponential backoff scheme is the root cause of QoS issue. Therefore, this paper first proposes a CSMA with automatic synchronization (CSMA/AS) MAC protocol to mitigate the collision problem caused by random access. By CSMA/AS, all the stations can be synchronized and then served in a round-robin fashion without contention collisions. Even if a new station joins, the wireless network can also quickly converge and go back to the synchronized state. The simulation results show that the proposed CSMA/AS protocol can fully mitigate the issues caused by random access, such as the severe contention collisions and large delay variation. In addition, this paper demonstrates how to provide hard QoS guarantees, such as fairness, rate guarantee, and delay guarantee, which cannot be achieved by the existing CSMA-based protocols. Because CSMA/CA does not rely on any additional control message, the implementation complexity of CSMA/AS is similar to that of legacy CSMA/CA protocols. © 2017 Published by Elsevier B.V.

1. Introduction The increasing prevalence of real-time applications has created a strong demand for quality of service (QoS) support in network infrastructures. A centralized contention-free medium access control (MAC) protocol [1–5] has frequently been considered the only method capable of serving real-time applications with hard QoS guarantees; however, the implementation of this protocol is highly complex because it requires the dynamic allocation of system resources in real time. In the IEEE 802.11 wireless local area network (WLAN) standard [1], the centralized contention-free MAC protocol was considered optional, and only a simple reference scheduler was presented. In practice, all commercial products are currently based on a contention MAC protocol. The primary MAC protocol of IEEE 802.11 [1] is called the distributed coordination function (DCF) based on the carrier sensing multiple access with collision avoidance (CSMA/CA) protocol with the slotted binary exponential backoff (BEB) scheme. The performance of the DCF protocol is extensively analyzed and studied in the literature. The works in [6–10] were based on the two-dimensional Markov chain

∗

Corresponding author. E-mail address: [email protected] (Y.-W. Kuo).

http://dx.doi.org/10.1016/j.comnet.2017.07.007 1389-1286/© 2017 Published by Elsevier B.V.

model for computing performance metrics such as throughput, average access delay, or distribution of access delay. Several previous studies [11–19] have investigated the performance of nonsaturated networks under the assumption of homogeneous traffic. To enhance the performance of the DCF protocol, various algorithms [20–23] have been developed to adaptively adjust the initial backoff window size based on the estimated number of active stations or to control the amount of traffic entering the network by employing a shaper at each station. Collectively, these results indicate that QoS support is poor because random access generates large variations in packet delay. The enhanced distributed channel access (EDCA) protocol [1] is aimed at improving QoS guarantee in WLANs by assigning unique parameters to individual traffic classes. The EDCA protocol can provide only service differentiation without QoS guarantees, because it is also based on the CSMA/CA protocol for channel contention. Hard QoS guarantees cannot be provided to real-time applications with a strict delay bound by a random access MAC. Packet delay plays a crucial role in the service quality of realtime applications, but the definition of delay depends on the studied scenarios. In previous studies on saturated WLANs [6–10], the delay of a successfully transmitted packet has been defined as the time interval from the moment the packet reaches the head of its queue to the moment it is transmitted to the destination (i.e., ac-

32

Y.-W. Kuo, J.-H. Huang / Computer Networks 127 (2017) 31–42

cess delay). In studies investigating nonsaturated WLANs [11–19], wherein packets were generated according to a particular model, the delay has been defined as the interval from the moment when the packet is generated to the moment it is transmitted to the destination (i.e., end-to-end delay). Because of random access, a station can be infinitely interrupted whenever other stations are active. Consequently, only the statistics of delay are studied in the literature. The problem of poor QoS capability arises from the BEB-based collision avoidance feature. The fundamental goal of BEB is to reduce collision probability under heavy loads by doubling the size of the contention window of collided stations. Under this design, various stations have equal share of the wireless channel in the long run, but packets from a particular station are not fairly treated. The doubled contention window forces collided packets to wait for long time, thus resulting in a lengthy delay. By contrast, non-colliding packets can be transmitted quickly. Consequently, the delay variation of a station is typically substantial, which is unfavorable for real-time applications. Our previous research [24,25] started with parameter optimization with delay constraint under the DCF protocol, and we proposed the fixed contention window backoff (FCWB) protocol to achieve a better performance. Because we believe that further improving the delay performance of the DCF protocol is difficult, we focused on modifying it. In [26], a novel protocol called the delay contention DCF (DC-DCF) protocol was proposed to assign higher priority to collided stations by adding several additional backoff slots in the first transmission attempt of a new head of line packet. The simulation results showed that the DC-DCF protocol possesses a trafficshaping feature that was considered favorable because it provides traffic isolation between stations. Basically, DC-DCF can be considered as the general form of the CSMA/CA protocol. The performance and behavior differ a lot with parameter configuration. In this paper, we first focused on optimizing the parameters of the DC-DCF protocol for a saturated WLAN based on the Markov chain model. The most critical result is that the optimal initial contention window size for the DC-DCF protocol should have a value of 1, which differs considerably from the values typically used in the DCF protocol. Subsequently, we conducted several simulations to validate the optimization, and the simulation results differed from what we anticipated. Through careful observation, a fundamental assumption of the Markov chain model was identified as invalid for the DC-DCF protocol with the above configuration. Nevertheless, we found that saturated WLANs become synchronized quickly when the value of the initial contention window is set to 1. Through careful adjustment of the backoff mechanism to operate under nonsaturated conditions, an appropriate version called CSMA/AS was proposed to form a synchronized WLAN without collisions. Consequently, a minimum service rate is guaranteed for each active station, and the bound of end-to-end delay can then be derived for a given type of traffic in order to provide hard QoS guarantees. The proposed CSMA/AS makes the traditional contention-based protocol able to provide a station level guarantee. We believe that it is cost-effective than Point coordination function (PCF) [1] which requires a customized scheduler. This remainder of this paper is organized as follows. Section 2 provides a brief overview of the DCF, FCWB, and DC-DCF protocols. Section 3 describes our method for designing the proposed CSMA/AS protocol. The properties of CSMA/AS protocol are derived and explained in Section 4. Simulation results demonstrating the advantages of the proposed protocol are given in Section 5. Finally, Section 6 presents the conclusions of this study.

2. Preliminaries To clarify the concepts presented in this paper, this section briefly introduces the fundamental concepts of WLANs in the MAC layer. 2.1. DCF protocol When a station transmits a packet, it must first detect the wireless medium. If the medium is busy, it defers transmission until the medium is in idle. When the medium becomes idle, the source station can initiate a backoff operation only after an additional idle time interval, which is referred to as the DCF interframe space (DIFS). The backoff counter, which has a uniformly selected initial value, is decreased by one after an idle slot time, and is frozen when the source station detects that the medium is busy. When the backoff counter reaches zero, the source station starts transmitting the packet. When the backoff counters at multiple stations reach zero at the same time, a collision occurs when packets are transmitted simultaneously. When the destined station successfully receives the packet, it transmits a positive acknowledgment (ACK) to the source station after a time interval, which is known as the short interframe space (SIFS). After the source station receives the positive ACK, the transmission is successfully completed. If the source station does not receive the positive ACK, it schedules a retransmission, and the backoff operation restarts. For each transmission attempt, the initial backoff count value is uniformly selected from [0, Wi − 1], where Wi is the current contention window size, and i denotes the backoff stage (i.e., the number of failed transmissions for a given packet). Initially, i = 0 for each packet, and this value is increased by one when transmission failure is detected. The contention window size Wi at stage i is controlled according to the BEB scheme. At the first transmission attempt, W0 is equal to the minimum contention window size, which is denoted as W. When a station detects a failed transmission, it doubles Wi until a maximum value Wmax = 2m W is reached, as shown in (1). Subsequently, Wi remains constant until the packet is successfully transmitted or dropped. The contention window size Wi is defined as



Wi =

2iW, 0 ≤ i ≤ m, 2mW ,m < i ≤ R,

(1)

where R is the retry limit, and m is the maximum number of times that Wi can be doubled. When a station fails to transmit the packet at backoff stage R, it drops the packet and initiates a new transmission for the head-of-queue packet where W0 = W. To avoid the hidden node problem, a four-way handshake protocol, called the request to send (RTS)/clear to send (CTS) access mechanism, is typically used. When the backoff counter reaches zero, the source station transmits an RTS packet instead of the original data packet. After the destined station receives the RTS packet, it delays transmission of the CTS packet according to the SIFS. After successfully receiving a CTS packet, the source station delays transmission of the data packet according to the SIFS. If no CTS packet is received, then the source station must schedule a retransmission. RTS and CTS packets also contain a network allocation vector (NAV) field to notify other stations how long the source station requires to complete the transmission. 2.2. Transmission opportunity (TXOP) By adopting the DCF protocol, a station can transmit a packet only after completing the backoff operation. Hence, when the stations perform backoff operations, there exist idle periods in the wireless medium, resulting in high protocol overhead. The IEEE

Y.-W. Kuo, J.-H. Huang / Computer Networks 127 (2017) 31–42

WLAN standard in [1] introduced a burst transmission feature, called transmission opportunity (TXOP), to reduce the protocol overhead. This feature enables the consecutive delivery of multiple packets within the time interval TTXOP when a station successfully completes an RTS/CTS handshake. The TXOP feature is easy to implement for calculating the NAV. After overhearing an RTS/CTS packet, all stations except the transmission pair suspend their backoff counter for a period equal to the NAV. By doing so, the transmission pair can continue delivering multiple packets without collisions occurring. The duration of a TXOP interval is specified in milliseconds. Generally, a longer TXOP interval for an active station can reduce the protocol overhead because fewer backoff operations are required. However, the other stations may be blocked for too long. Consequently, a tradeoff exists between channel utilization and packet delay.

τ=

1 − pR+1

33

for all i. Through application of the FCWB protocol and appropriate adjustment of the contention window size, the collision probability p in a saturated network can be adequately controlled by adjusting W, and the delay variation can be markedly reduced. To maximize the channel utilization, the recommended value for p is 0.17 [25]; substituting p = 0.17 into (2) and (3) enables determining the optimal contention window sizes for various network sizes, as described in Section 5. 2.5. Delayed contention DCF (DC-DCF) In [11], Zhai et al. observed that the maximum throughput can be achieved when the collision probability p is approximately 0.196 for the DCF protocol. Therefore, a method for maintaining high throughput for large networks by controlling user traffic was pro-

,

(1 − ⎧ p) f m+1 m+1 ⎪ ⎨ W (1 − (2 p) )(1 − p) + (1 − 2 p)(1 − p ) , 2(1 − 2 p)(1 − p) where f = R+1 m+1 R R+1 (m−R ) ) ⎪ ⎩ W (1 − (2 p) )(1 − p) + (1 − 2 p)(1 − p ) + W 2 p (1 − 2 p)(1 − p , 2(1 − 2 p)(1 − p)

R ≤ m.

(2)

R > m.

2.3. Worst case analysis In a queueing system with light traffic, packet delay is short because of the short queue length. Nevertheless, as the network load increases, the packet delay increases considerably, and an appropriate scheduling algorithm is indispensable to provide certain level of QoS guarantees for real-time applications. Classic scheduling algorithms have been developed and analyzed based on the worst-case condition (i.e., when all users are greedy) in order to provide hard QoS guarantees. For further detail regarding the operations and properties of classic scheduling algorithms, the reader is referred to [27] and [28]. A WLAN can be regarded as a special queueing system in which the system randomly selects an active station to serve. In the literature, the worst case where all stations are active is considered in [6–10], but only the statistics of access delay can be obtained. These analyses were based on the Markov chain model, wherein a state represents the current retry stage and the backoff counter for a station. Assume that each transmission attempt has the same independent collision probability p. The transmission probability τ of a station for a given slot can be calculated using (2) [7]. Let N denote the number of stations in the network. The collision probability can be calculated using the following equation:

p = 1 − (1 − τ )N−1 .

(3)

Eqs. (2) and (3) form a nonlinear system with two unknowns (τ and p), which can be solved numerically to obtain a unique solution. Subsequently, the first moment, second moment, and distribution of access delay can be calculated. For further detail, the reader is referred to [6–10]. Although the saturated condition is a special operation case, the analysis results in this condition provide a quick way to evaluate and optimize the worst case performance. 2.4. Fixed contention window backoff (FCWB) protocol In [24] and [25], we have investigated the DCF protocol with the aim of maximizing channel utilization when the TXOP feature is enabled. We formulated an optimization problem to determine the optimal parameters. The near optimal solution found in those studies led to the development of the FCWB protocol, where the contention window sizes of all backoff stages were fixed at Wi = W

posed, although a method for controlling user traffic was not addressed in [11] and remains an open issue. A traditional method for controlling user traffic is to add a traffic shaper for each station, but this would be impractical for current low-cost WLAN equipments. In [26], we proposed a novel protocol for controlling contention collision. The basic concept of that method was to reduce the number of stations contending for the wireless medium during the backoff process of a tagged station. To protect the collided stations, applying our approach postpones the initial transmission attempt of each packet for numerous slots. Because this delays the contention cycle of each packet, the proposed scheme was named the delayed contention DCF protocol, in which the collided packets have a higher probability of being transmitted; thus, their delay can be shortened. This concept can be implemented very easily by modifying the calculation of the backoff counter in a station, as follows:



Backoff counter =

C + rand(W0 ), for 1st transmission, rand(Wi ), for retransmission,

(4)

where rand(Wi ) is a function that generates a uniformly distributed random integer within the range [0, Wi − 1], and C denotes the number of additional slots in the backoff process for the initial packet transmission attempt. Moreover, the DC-DCF protocol can be considered as a general form of CSMA/CA, because the DCF protocol is equivalent to the DC-DCF protocol when C = 0. Let τ be the transmission probability of a station in a given slot. With the same procedure in [7] to solve the state probability of the Markov chain model of DC-DCF [26], τ can be calculated as follows:

τ=

1 − pm+1 , (1 − p)(C + f )

(5)

where f is given in (2). A simple method for determining the operation point is to set p to the value suggested in [11] (i.e., p = 0.196). Subsequently, from (3) and (5), we can calculate C as follows:

C = round (

1 − pm+1

(1 − p)(1 −



N−1

1 − p)

− f ).

(6)

The suggested value of C is given in Section 5 for various network sizes N.

34

Y.-W. Kuo, J.-H. Huang / Computer Networks 127 (2017) 31–42

The simulation results in [26] showed that the DC-DCF protocol can successfully regulate the traffic without the need of an additional traffic shaper, resulting in traffic isolation between stations. Although we successfully achieved the goal of controlling the collision probability and reducing the delay variation in [26], a method for determining the optimal parameters of C and W for the DC-DCF protocol remains unsolved

the time wasted when a collision occurs, respectively. If a tagged station experiences k collisions and l backoff slots before it successfully transmits a burst, the burst delay d can be expressed as (X1 + X2 + … + Xl + kTc + Ts ), where Xj for j = 1, 2, …, l is a random variable representing the time consumed in the jth backoff slot. The random variables Xj are i.i.d. and the duration of Xj can be expressed as

3. Design of the proposed CSMA/AS protocol

Xj =

First, we describe how to derive the optimal values of system parameters for the DC-DCF protocol based on the traditional Markov chain model in a saturated WLAN. The obtained optimal value for the initial contention window was 1, which differs considerably from the frequently used values in the DCF protocol. Moreover, the simulation results of a saturated WLAN with the optimized DC-DCF protocol were quite different to what we expected, and this revealed a critical property that the network can become synchronized. Finally, a new mechanism of sustained backoff process is added to make the protocol able to operate in a nonsaturated network. Our approach to design the proposed protocol is described in the following subsections. 3.1. DC-DCF parameter optimization Channel utilization and packet delay are major factors in designing high-throughput WLANs with QoS guarantee. We started with investigating the performance of the DC-DCF protocol in a saturated network with the TXOP feature enabled. Within a TXOP interval, multiple packet transmissions in one TXOP interval form a TXOP burst. The TXOP burst delay d represents the time period between the end of two successive TXOP bursts of a station. As mentioned in Section 1, the delay is unbounded because of random access. Let D denote the desired bound of access delay, and any burst delay longer than D is considered a violation event. The QoS goal is controlled to maintain the violation probability below a user-specified threshold, denoted as Pv . Additionally, packet loss caused by exceeded retransmission is also considered a violation event. The QoS constraint can be expressed as 1 − P(d ࣚ D) ≤ Pv . We can calculate the violation probability by using the method in [10], which is summarized as follows. All successfully transmitted bursts can be classified into several groups based on the total number of backoff slots and the number of collisions. The probability P(d ࣚD) is equal to the summation of the conditional probability of each group. Let k and l denote the number of collisions and total number of backoff slots for a tagged burst, respectively. The violation probability can be expressed as

P (d ≤ D ) =

s (k ) R C+   k=0

=

P (d ≤ D|k, l )P (l |k )P (k ),

Ts , with the probability Ps , Te , with the probability Pe , Tc , with the probability Pc ,

(8)

where Ps is the probability of successful transmission in a backoff slot of a tagged station, Pe is the probability that all other stations are idle, and Pc is the probability of a collision occurring. Consider a wireless network comprising N stations. These probabilities can be calculated by Ps = (N − 1)τ (1 − τ )N −2 , Pe = (1 − τ )N −1 , and Pc = 1 − Ps − Pe , respectively. To reduce the computation complexity, d is further approximated as a Gaussian random variable with mean md and standard deviation σ d . Subsequently, the conditional violation probability can be easily calculated by using the error function. The next step is to calculate the channel utilization, denoted as μ. Let Ptr represent the probability that at least one transmission is in the considered backoff slot, and Ptr = 1 − (1 − τ )N . Let Psuc denote the probability of a successful transmission, and Psuc = Nτ (1 − τ )N −1 . The actual channel utilization is dependent on the packet size in a transmission burst and the traffic arrival pattern, but this is too complex if the traffic model is involved. For computational simplicity, μ can be approximated by

μ≈

Psuc TTXOP

(1 − Ptr )Te + Psuc Ts + (Ptr − Psuc )Tc

.

(9)

When the packet is large or the frame aggregation feature in IEEE 802.11 is enabled, the approximation error becomes small. As defined in (1)–(9), channel utilization and QoS are markedly influenced by the system parameters, including the additional backoff slots C, the initial contention window size W, and the duration time TTXOP of TXOP interval. Let C∗ , W∗ , and T∗ TXOP be the optimal values for C, W, and TTXOP , respectively. To maximize the channel utilization under the QoS constraint, the optimization problem can be formulated as follows: ∗ (C ∗ , W ∗ , TTXOP ) = argmax μ, subject to 1−P (d < D ) ≤ Pv .

(10)

C,W,TTXOP

To clarify the method for finding the solution, we first introduce two properties of the analytic model before describing the optimization procedure. Property 1. If R, N, C, and W are given, μ increases with TTXOP . Proof. Ts is equal to TRTS + TCTS + 2SIFS + TTXOP + DIFS, where TRTS is the transmission time of an RTS packet, and TCTS is the transmission time of a CTS packet. We can rewrite (9) as

l=C

s (k ) R C+   k=0

P (d ≤ D|k,l )P (k, l )



(7)

l=C

where s(k ) = kj=0 (W j − 1 ) and P(k) = pk (1 - p). Because l is the summation of discrete random variables, the probability mass function (pmf) of l given k collisions can be calculated by P (l |k ) = ( f0 ∗ f1 ∗ · · · ∗ fk )(l ), where fj is the pmf of a uniform distribution in [0, Wj - 1]. Three possible cases exist in a randomly selected backoff slot observed by a tagged station: 1) empty slot, where other (N−1) stations are idle; 2) a station successfully transmits a burst; and 3) a collision occurs. For the latter two cases, the tagged station suspends its backoff, and the consumed time will be longer than the duration of a slot time. Let Te , Ts , and Tc denote the empty slot time, the time for a burst to be transmitted successfully, and

μ≈

1

(1−Ptr )Te +Psuc (TRT S +TCT S +2SIF S+DIF S ) + (Ptr −Psuc )Tc Psuc TTXOP

.

(11)

+1

In (11), varying TTXOP does not affect Ptr , Psuc , Te , and Tc . Thus, as shown in (11), μ increases with TTXOP . Property 2. If R, N, C, W, and D are given, the violation probability increases with TTXOP . Proof. For the given combination of (k, l), the mean md and standard deviation σ d of burst delay can be calculated according to the probability mass function of X in (8) as follows:

md = l (Ps Ts + Pe Te + (1 − Ps − Pe )Tc ) + kTc + Ts .

(12)

Y.-W. Kuo, J.-H. Huang / Computer Networks 127 (2017) 31–42



σd = l E (Xk2 ) − X¯k2  = l Ps Ts (Pe (Ts − 2Te ) + Pc (Ts − 2Tc )) + ITs, Pe Te2

Pc Tc2

− Pe2 Te2

(13)

− Pc2 Tc2

where the last term ITs = + − 2Pe Te Pc Tc includes all the terms irrelevant to Ts . Because Ts is considerably larger than Te and Tc , (Ts − 2Te ) and (Ts − 2Tc ) are positive. Consequently, (12) and (13) show that both the mean md and standard deviation σ d of d increase with Ts , implying that the violation probability also increases with TTXOP .  Based on these two properties, we developed the following optimization procedure. For a given value of p, τ and C can be determined by using (5) and (6), respectively. Because both the channel utilization and the violation probability increase with TTXOP , once C and W are given, the optimal T∗ TXOP is the maximum of all feasible TTXOP values that meet the QoS requirement of (1 - P(d < D)) ≤ PL . Finally, the channel utilization μ by using (9). The graphical plot of the channel utilization μ versus p shows that smaller W values are more favorable. Consequently, the optimal W∗ = 1. In addition, we also found that the channel utilization is less sensitive to the collision probability when W = 1, implying that selecting an appropriate value for p for all configurations is reasonable. Based on the numerical data of various configurations, we determined that setting p = 0.32 improves channel utilization. So far, the performance analysis is based on the analytic Markov chain model. After conducting several experiments of a saturated WLAN by using NS-2 [29] to validate the obtained optimal value, we observed that the simulation results differed considerably from the performance evaluation derived by the analytic model. Unexpectedly, when W = 1, all stations quickly become synchronized (i.e., each station sequentially transmits packets without any collision). Consequently, the measured collision probability was much smaller than the expected value 0.32. By assuming that each station operates independently, the analytic model reduces the complexity by decoupling the entire network into N independent Markov chains. Although this assumption is invalid when the network is synchronized, the optimal W∗ obtained through the analytic model leads to the critical feature of synchronizing a WLAN. Accordingly, we proposed the CSMA/AS protocol, which is detailed in the following subsection. It is worthy to point out that this subsection describes how we found this interesting property. The proposed CSMA/AS protocol does not rely on this analytic model to determine the parameter value. 3.2. The CSMA/AS protocol In practice, the queue of a station may be empty when all packets are sent. Thus, we also investigated the performance of the proposed method in nonsaturated networks. Simulation results show that the collision probability p increases considerably in nonsaturated cases, because a station loses synchronization when it is idle. As a result, we need a method to solve this issue. The complete operation of the proposed CSMA/AS protocol is summarized as follows. The backoff counter of a packet is calculated through (4), and the contention window is updated through (1), wherein W = 1. In addition, we developed a new mechanism of a sustained backoff process to render the CSMA/AS protocol effective under any traffic load. The sustained backoff mechanism can make the stations always behave as they would under the saturated condition, thereby maintaining synchronization among the stations. Each station must restart the backoff process immediately after completing a transmission attempt, regardless of whether its queue is empty or not. When the backoff countdown of a station ends, there are two possible cases:

35

Case I: If packets are in queue, the station can initiate the four-way handshake transmission procedure. After completing the current transmission, it restarts the backoff process again. Case II: If the queue is empty, the station also restarts the backoff process. According to the CSMA/CA protocol, before a station starts counting down, it must wait for an idle DIFS to ensure that no transmission is ongoing. In this case, however, because the station successfully counts to zero (i.e., the wireless medium is idle), waiting for an idle DIFS is unnecessary. Consequently, the station restarts the backoff process without waiting for the DIFS in this case. Fig. 1 depicts the operation of the CSMA/AS protocol in a network comprising three stations and C = 10. Because W0 = 1 was adopted for the CSMA/AS protocol, the backoff counter for the first transmission attempt is fixed and equal to C. For retransmission, Wi is updated according to the BEB scheme. The basic unit in the timeline is a slot. Without loss of generality, assume that DIFS and a collision event separately consume two slots as shown in Fig. 1. We define a station backoff cycle as the time interval consisting of a backoff countdown and the subsequent transmission attempt. In case no packet is queued when a station counts down to zero, the transmission attempt is null, and a new cycle starts. Suppose that Stations 1 and 2 collide at Slot 0, and the remaining backoff counter of Station 3 is equal to 8. The contention window W1 for Stations 1 and 2 is doubled and equals to 2 after the first collision occurs. Suppose that the two stations draw an identical backoff value and collide again at Slot 5. At the third transmission attempt, Stations 1 and 2 successfully draw unique numbers. Station 2 transmits at Slot 11, and Station 1 transmits at Slot 18. At Slot 28, Station 3 counts down to zero, but there is no packet queued for transmission. Because Stations 1 and 2 are in backoff, Station 3 restarts its backoff process without waiting for DIFS in order to maintain synchronization with other stations. For time after Slot 28, these stations are well synchronized in the order of Station 2, Station 1, and Station 3. The number of empty slots in a backoff cycle of any station is also fixed and equal to the number of additional backoff slots, C. In summary, a backoff cycle in the synchronized state consists of C empty slots and N transmissions when all N stations are active. The first contribution of this paper is proposing the CSMA/AS protocol to achieve a collision-free CSMA network, by adopting (4) with W0 = 1 and integrating the sustained backoff mechanism. The authors in [30–32] also proposed the protocols to achieve collision-free operation. Barcelo et al. developed the learning-BEB (L-BEB) protocol [30]. At first, the station follows the BEB scheme to contend for the channel. Once the station successfully sends a packet, the station switches from the BEB mode to the deterministic backoff mode until a collision occurs. There are two differences between L-BEB and CSMA/AS. First, CSMA/AS uses a deterministic backoff value in the first transmission attempt all the time, but L-BEB switches between random access and deterministic access. Second, CSMA/AS adopts the sustained backoff mechanism to keep the inactive stations still synchronized, but stations in a L-BEB network become out of sync when they are idle. The zero collision MAC (ZC-MAC) protocol in [31] tries to form a time division multiple access network. In the ZC-MAC protocol, a new or collided station should monitor the channel activity, and then select an empty time slot to send data. If the transmission attempt is successful, the station can periodically use the selected time slot until a collision occurs. Fang et al. [32] proposed the adaptive versions of ZC and L-BEB protocols to adjust the cycle length based on the network loading. The number of stations is likely to be larger than the number of slots in one cycle when the traffic is light. Though the throughput and protocol efficiency increases, the complexity

36

Y.-W. Kuo, J.-H. Huang / Computer Networks 127 (2017) 31–42

Fig. 1. Example illustrating the operation of the proposed CSMA/AS protocol (C = 10).

also increases and the network may be disturbed by both station arrival and traffic activity. During the period of convergence, transmission is no longer collision free and hard QoS requirement cannot be guaranteed for collided stations. Because we exploit that the TXOP mechanism can significantly boost the throughput as demonstrated in Section 5.1, this paper focuses on increasing the convergence speed as shown in the next paragraph under the condition that a cycle can accommodate all stations. As a result, new-station arrival is the only possibility to make a synchronized station out of sync. Even when collisions occur, the collided stations can quickly become synchronized again to lower the impact on QoS. In practice, the proposed CSMA/AS protocol integrated with the TXOP feature is a simple and effective way to achieve a collision-free wireless network with high throughput. One major concern in the CSMA/AS network is how fast the network can converge again and return to the synchronized state, after a new station joins. Therefore, the final step in designing the CSMA/AS protocol is determining the appropriate number of additional backoff slots C which is the key parameter affecting the convergence time. The analytic model discussed in Section 3.1 is invalid for the proposed protocol, and the method for determining the appropriate value is described as follows. The parameter C in the CSMA/AS protocol affects the number of collisions before the network becomes synchronized. Mathematically analyzing the complex effect of a new-station arrival on the synchronized network is difficult. However, we suggest a simple rule for control the successful transmission probability, based on the following observation in the saturated case. In a synchronized network, the timeline can be divided into the repeating patterns, each of which contains (C + N) events and N DIFSs. An event could be an empty slot or a successful transmission. For example, in Fig. 1, the time interval from Slot 44 to 71 is a pattern with 3 successful transmissions, 10 empty slots, and 3 DIFSs. Let pj denote the probability that the newly arriving station transmits successfully on the jth attempt. The new station can successfully transmit its packet in the first try, as long as the transmission occurs at the event of an empty slot. Thus, p1 can be calculated by

p1 = C / (C + N ).

(14)

Suppose that the first attempt is unsuccessful, and the transmission by the new station collides with that of station i. The other (N − 1) transmissions of other stations spread randomly in the following (C + N − 1) events. Because W1 = 2, the probability that the two collided stations draw unique backoff values at the second attempt is equal to 0.5. The probability that any one of the other (N − 1) stations also occupies the slot selected by the new station is equal to (N − 1) / (C + N − 1). Accordingly, p2 can be calculated by





p2 = ( 1 − p1 ) 0.5∗ 1 −

N−1 (C + N − 1 )



.

(15)

For a given network of size N, the value of C can be adjusted to control the probability of successful transmission, as shown in (14) and (15). For example, if C is triple of N, (p1 + p2 ) is approximately 0.84. Thus, by setting C = 3 N, the probability that a new station would encounter no collision or only one collision is approximately 0.84. Based on this simple rule, we can determine the value of additional slots C to ensure that the network can stabilize quickly after a new station joins. In practice, a new station can monitor the wireless channel to determine the location of unused slots for communicating with the AP to avoid collision with existing stations. In an ad hoc network where the network size known a priori, each station can be appropriate configured before deployment. However, for an infrastructure network that is managed by an access point (AP), C can be dynamically adjusted. Once the AP adjusts the value of C, it broadcasts the updated value. After receiving the broadcasted packet, each station immediately updates parameter C and its backoff counter. Let C’, bci , and bci denote the new value of C, the original backoff counter value, and the updated value of station i, respectively. Each station updates its backoff counter value by



bci + (C  − C ), bci = bc + (2C  − C ), i

if (bci + (C  − C )) ≥ 0, otherwise.

(16)

According to the operation of CSMA/AS protocol, if the backoff counters at all the stations simultaneously decrease or increase by an identical value, the network can maintain synchronization. 4. CSMA/AS properties The second contribution of this paper is to fully utilize the collision free property for providing hard QoS guarantees. This section presents the properties of the proposed CSMA/AS protocol in a synchronized network, including the saturated throughput, guaranteed service rate per station, end-to-end delay bound for constant bit rate (CBR) service, end-to-end delay bound for variable bit rate (VBR) service, and advantage of the contention-free feature. For ease of presentation, several symbols are defined in Table 1. A method for quickly measuring the system capacity is to calculate the saturated throughput. In a synchronized CSMA/AS network, a pattern with C empty slots, N bursts, and N DIFSs repeats. The packet transmission time T1 with payload size L is equal to (Th + (L/PR ) + 2SIF S + TACK ). For a given TXOP duration TTXOP , the maximum number of packets in a TXOP burst, denoted by n, equals  TTTXOP . Subsequently, the saturated throughput STH can be calcu1 lated as follows:

ST H =

NnL . C Te + N (DIF S + TRT S + SIF S + TCT S +TT XOP )

(17)

We can see that STH increases with TTXOP , as expected, but we should limit the size of TXOP bursts when considering the QoS

Y.-W. Kuo, J.-H. Huang / Computer Networks 127 (2017) 31–42

37

Table 1 Definition of symbols. Terms

Definition

PR L N T1 Th STH R Tp dCBR and dVBR

PHY data rate Packet size (bits) Maximum number of packets in a TXOP burst Total transmission time of a packet Transmission time of PHY and MAC headers Saturated system throughput Guaranteed service rate for station i Period for a CBR flow Delay bound for CBR/VBR flow

guarantee for each station. Similarly, STH can be calculated for the case without TXOP feature, as follows:

ST H =

NL . C Ts + N (DIF S + TRT S + SIF S + TCT S +T1 )

(18) Fig. 2. Calculation of the delay bound for VBR traffic.

A WLAN can be considered as a multiqueue system, in which several stations share the wireless medium. Because the stations are served in a random order with the traditional DCF protocol, the QoS is hard to guarantee. By adopting the CSMA/AS protocol, each station is adequately synchronized and served in a round robin fashion. Thus, fairness is easy to maintain because the service order is fixed. The guaranteed service rate of an active station can be calculated as follows:

r=

ST H . N

(19)

In the following, we explain how to provide hard QoS guarantees for CBR and VBR services in a synchronized CSMA/AS WLAN. Given the required end-to-end delay bound, the maximum value of TTXOP can be derived to maximize the saturated throughput. A CBR flow requires a rigid delay bound. For every packet generating period Tp , a CBR flow generates a packet that should be served within Tp . In legacy WLANs, ensuring the delay of a CBR flow is difficult, because of the large delay variation caused by random access. By contrast, the CSMA/AS protocol can guarantee a service opportunity for each station every (C + N) events. Because the maximum delay occurs when all of the stations are greedy, the end-to-end delay bound dCBR can be computed by

dCBR = C Te + N (DIF S + TRT S + SIF S + TCT S + TTXOP ).

(20)

To meet the hard QoS requirement that dCBR ≤ Tp , the maximal TTXOP can be calculated as follows:

TTXOP =

Tp − C Te − DIF S − TRT S − SIF S − TCT S . N

(21)

The CSMA/AS protocol also can provide hard QoS guarantees for the VBR traffic. In general, the arrival traffic of a VBR flow within a given time interval can be modeled using the token bucket algorithm with the parameters σ and ρ , where σ denotes the “burstiness” and ρ denotes the token rate. In the token bucket algorithm, the accumulated arrival traffic when a station is busy is bounded by the envelop function (σ + ρ t), as shown in Fig. 2. Because the CSMA/AS protocol guarantees a service rate r to each station, the entire network can be described based on the model of a latencyrate (LR) server [27] with two parameters: latency and guarantee rate. Latency, denoted as , is defined as the worst-case delay for the first packet transmitted in a busy period for a given station. By the same argument for calculating the end-to-end delay bound for CBR traffic,  can also be computed using (20). With the guaranteed service rate r, the accumulated served traffic of a station can be expressed as r(t − ) as shown in Fig. 2. Then, as proven in [28], when r ≥ ρ , the end-to-end delay bound provided by an LR server can be computed by

dV BR =  +

σ / r.

(22)

Table 2 System parameters. Parameter

Value

PHY rate, PR Basic rate SIFS DIFS TRTS TCTS , TACK Slot time, Ts C for CSMA/AS W for DCF & DC-DCF W for CSMA/AS L Tp

216 Mbps 24 Mbps 16 μs 34 μs 14.7 μs 12.7 μs 9 μs 3N 32 1 10 0 0 bytes(80 0 0 bits) 40 ms

The token rate ρ should not be smaller than the arrival rate when modeling the traffic. Because σ decreases as ρ increases, setting ρ = r produces the tightest bound. Usually CBR and VBR flows coexists at the same time. In practice, TTXOP is determined by (21) based on Tp first, and then r and dVBR can be calculated by (19) and (22), respectively. Another unique advantage of the CSMA/AS protocol is that it can render the network collision-free when all of the stations are synchronized. Traditional DCF-based protocols are subject to collision problems. Even when the queue is not full, packets may be dropped when the number of transmission attempts exceeds the retry limit in conventional WLAN MAC protocols. By contrast, in the CSMA/AS protocol, once the network is synchronized, a packet is dropped only when queue overflow occurs. In addition, the contention-free capability of the CSMA/AS protocol improves the reliability of wireless transmissions, particularly for multicast/broadcast traffic. Unlike unicast packets, which are delivered with a positive ACK, multicast/broadcast packets are delivered directly to all destination stations without an ACK in order to reduce overhead and delay. The source station cannot determine whether the transmission is successful. Once a collision occurs in a traditional WLAN, the multicast/broadcast packet is lost. This problem can be mitigated by adopting the proposed CSMA/AS protocol. 5. Performance evaluation This section evaluates the performance of the proposed CSMA/AS protocol by using typical system parameters listed in Table 2. The data packets were transmitted at 216 Mbps, while all control packets were transmitted at the basic rate. All simulations were conducted using the NS-2 simulator [29]. Simulations of the DCF, FCWB, and DC-DCF protocols were also conducted for compar-

38

Y.-W. Kuo, J.-H. Huang / Computer Networks 127 (2017) 31–42 Table 3 Parameters for various N values. N

TXOP (μs)

W for FCWB

C for DC-DCF

10 15 20 25 30 35 40

3895 2562 1895 1495 1229 1038 895

97 151 204 258 312 365 419

25 54 82 111 139 168 196

protocol is more than 10% higher than that of the other protocols because, in addition to empty slots, collisions also degrade the protocol efficiency. However, compared to the physical layer (PHY) rate of 216 Mbps, the protocol efficiency is poor, even for the proposed protocol. As anticipated, the STH of the DCF protocol decreased with N because the collision probability increases with N. Heavy collisions in the DCF protocol resulted in redundant transmissions and degraded the efficiency of the protocol. Regarding the FCWB and DC-DCF protocols, the collision probability was adequately controlled by adjusting W or C, respectively, resulting in flatter curves. Because the optimal parameters for the FCWB and DC-DCF protocols were determined not only based on throughput, the STH of these protocols was smaller than that of the DCF protocol when N < 35 for the case of the DC-DCF protocol or when N < 40 for the case of the FCWB protocol. Fig. 3(b) shows the saturated throughput STH with the burst transmission feature enabled. Comparing it with Fig. 3(a) shows that applying the burst transmission feature doubles the throughput. The improvement from applying this feature is considerably greater than that from reducing the number of collisions. The results of this experiment also demonstrate the crucial role of the MAC protocol in achieving an efficient WLAN. Considering both hard QoS guarantee and throughput, TTXOP was adjusted based on the delay bound and network size by computing (21). Because TTXOP decreases as N increases, the protocol overhead increases with N, implying that the throughput decreases with N, as shown in Fig. 3(b). With the burst transmission feature enabled, the performance gap between the CSMA/AS protocol and the other protocols becomes smaller, but the capability of QoS guarantee differs considerably, as demonstrated in the following results. For the remaining experiments discussed in this section, the burst transmission feature was enabled. 5.2. Guaranteed rate

Fig. 3. Saturated throughput STH versus N: (a) without TXOP, and (b) with TXOP.

ison. The built-in 802.11 MAC and queue modules were modified to support the CSMA/AS, FCWB, and DC-DCF protocols, as well as the burst transmission feature. When the burst transmission feature was enabled, the length of TXOP interval TTXOP was calculated using (21), in which Tp = 40 ms (as listed in Table 3), and the same value was used for all protocols. As listed in Table 3, the values of W for FCWB were determined for the given p = 0.17, and the values of C for the DC-DCF protocol were determined for the given p = 0.196. 5.1. System capacity The first experiment was conducted to measure the saturated throughput STH under the scenario where all stations are greedy. Fig. 3(a) shows the results for the case without TXOP. For the CSMA/AS protocol, because C = 3 N, the value of STH is unrelated to the network size, as indicated by (17). After the network has been synchronized, only C empty slots are wasted in one backoff cycle. For all cases, the saturated throughout STH of the CSMA/AS

In a multi-queue system, a crucial purpose of QoS guarantee is to provide a minimum service rate to each active user. For the second experiment, a network comprising 30 stations was considered. All stations were greedy, although the data rate of Station 1 was varied. Fig. 4(a) shows the throughput of Station 1 versus its data rate. According to Fig. 3(b), STH is more than 86 Mbps for all of the protocols when N = 30, implying that each station can share more than 2.8 Mbps on average. However, the simulation result shows that the throughput of Station 1 under the DCF protocol was considerably less than the arrival rate, even when the arrival rate was 1 Mbps. Aggressive contention by the DCF protocol made it difficult to provide any rate guarantee. A station can gain more bandwidth only by transmitting more, but many packets were dropped. However, the FCWB and DC-DCF protocols made the stations less aggressive because of the large initial contention window or delayed contention. The collision rate was controlled and additional bandwidth was reserved for Station 1. When the arrival rate was greater than 2 Mbps, the throughput of the FCWB protocol could no longer follow the arrival rate and packet loss occurred. Although the DC-DCF protocol provided better traffic isolation between stations, packet loss continued to occur when the arrival rate was near the saturated point (i.e., 2.8 Mbps in this experiment). Serving stations sequentially under the CSMA/AS protocol provided perfect isolation between stations. As shown in Fig. 3(b), in a CSMA/AS network where N = 30, STH = 88.59 Mbps and r = 2.953 Mbps. The simulation result matches the calculated value. Fig. 4(b) shows the instantaneous throughput of Station 1, which was measured every second when the arrival rate was 2 Mbps for the first 30 s. The unstable throughput of the DCF protocol indicated that it is impossible to provide any rate guarantee. By contrast, the instantaneous throughputs of other three protocols

Y.-W. Kuo, J.-H. Huang / Computer Networks 127 (2017) 31–42

Fig. 4. (a) Throughput of Station 1, and (b) instantaneous throughput of Station 1 when input rate = 2 Mbps.

were more stable, implying that they can support a service rate of 2 Mbps for Station 1. The rate of variation under the FCWB protocol was more than that under the DC-DCF or CSMA/AS protocols. Few packets dropped under the FCWB protocol in this case, as indicated by the trace file. 5.3. CBR service This experiment, showing the worst end-to-end delay performance of a CBR flow for various network sizes, was conducted under the scenario where Station 1 had a CBR flow, whereas the other stations were greedy. The inter-packet arrival time was 40 ms, which is also the strict packet delay bound. The arrival rate of the CBR flow was 200 Kbps for a packet size of 10 0 0 bytes. Fig. 5(a) shows the average packet delay for all protocols. Thus, the DCF protocol is inadequate for serving such a low-rate flow for all cases because the average delay was approximately double the delay bound. For the other three protocols, all of the average end-toend delays were smaller than the delay bound. A CBR flow requires a hard delay bound guarantee; to validate this, Fig. 5(b) shows the cumulative distribution function of the packet delay when N = 20. We can see that the packet delay of the CSMA/AS protocol was within the delay bound for all packets, which demonstrates the capability of this protocol for providing hard QoS guarantee. On the other hand, packet delay might exceed the delay bound under the other three protocols. The probabilities of the delay being larger than the delay bound under the DCF, FCWB, and DC-DCF protocols were 0.32, 0.075, and 0.014, respectively. Although the DC-DCF protocol can markedly reduce the

39

Fig. 5. (a) Average CBR packet delay versus network size and (b) CDF of packet delay for N = 30.

delay variation, a packet might be preempted by too many packets from other stations because of random access, let alone the DCF and FCWB protocols. 5.4. VBR service This experiment was conducted in a network where N = 30 and each station has one VBR flow generated by a Poisson process. The average arrival rate of Station 1 was fixed at 2 Mbps, and we varied the arrival rate of other stations, denoted as r’, from 1 to 3.5 Mbps to simulate various network loads. Fig. 6(a) shows the average endto-end delay of Station 1. When the load was light, the average delay under the DCF protocol is the smallest, but the maximum delay was the worst one, as indicated by the raw data. As the network load increased, the delay under the DCF protocol increased considerably and with no limit, and packet loss became severe when r’ > 2 Mbps. For the other schemes, the average delay also increased with r’, but became stable when the network was saturated. Fig. 6(b) shows the CDF of the end-to-end delay when r’ = 3.5 Mbps. The capability of delay guarantee for VBR traffic can be measured by the probability of packet delay exceeding dVBR , which can be easily calculated by 1-CDF(dVBR ). Similar to Fig. 5(b), the tail spread widely except under the CSMA/AS protocol, because of burst traffic and random access. Regarding the calculation of the delay bound, the traffic must first be modeled. As mentioned in Section 4, ρ is equal to r, which is 2.953 Mbps in this case. Analyzing the arrival traffic with the above token rate showed that σ = 1400 bytes. Because TTXOP was determined using (21) with Tp = 40 ms,  was also equal to 40 ms. Substituting these values into (22) yielded dVBR = 78 ms. The probabilities of the delay ex-

40

Y.-W. Kuo, J.-H. Huang / Computer Networks 127 (2017) 31–42

bound can also be derived for a given type of traffic. In addition, the network capacity also increases because the transmission becomes collision free. These properties were evaluated through simulation. A simple and efficient MAC protocol is crucial to supporting QoS, but this paper shows only station-level QoS because each station had only one flow in all of the simulations. In the future, we will investigate a more general scenario in which each station has mixed traffic and downlink traffic dominates the network. Typically, multiqueues with an appropriate scheduler should be adopted in each station to provide class-level QoS. In addition, how to manage overlapped WLANs or hidden nodes remains an open issue. Acknowledgement This work was supported in part by the Ministry of Science and Technology, R.O.C., under grant no. MOST 105-2221-E-260-006-. References

Fig. 6. (a) Average VBR packet delay at different loadings and (b) CDF of packet delay when r’ = 3.5 Mbps.

ceeding dVBR under the DCF, FCWB, DC-DCF, and CSMA/AS protocols were 0.3, 0.15, 0.021, and 0, respectively. In summary, synchronizing a wireless network is the only way to guarantee a strict delay bound for a CBR or VBR flow. 5.5. Convergence test In the final experiment, we recorded the number of total collisions before the network (N = 30) was synchronized with random seeding in the random number generator of NS-2. Each station started in sequence with a time gap of 0.15 s. In addition to user traffic, routing messages were generated after the first user’s packet arrived. The simulation was repeated 40 times. The average number of total collisions was 20 and the maximum was 58. In other words, less than one collision (on average) occurred at each station before the network is synchronized. Subsequently, all stations had contention-free transmissions until a new station joined the network. 6. Conclusion In this paper, a novel MAC protocol for WLANs is proposed to fully mitigate the problems caused by random access such as collisions and poor QoS guarantee. By adding the delayed contention feature, adding the sustained backoff mechanism, and setting the initial contention window size to 1, we easily modified the CSMA/CA protocol to CSMA/AS, by which all stations were adequately synchronized without relying on any additional signaling message. Once the network was synchronized, the minimum service rate for a station can be guaranteed, and the end-to-end delay

[1] IEEE Standard for Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications, IEEE Std 802.11, 2007. [2] A. Grilo, M. Macedo, M. Nunes, A scheduling algorithm for QoS support in IEEE802.11 networks, IEEE Trans. Wirel. Commun. 10 (3) (2003) 36–43. [3] D. Gao, J. Cai, K.N. Ngan, Admission control in IEEE 802.11e wireless LANs, IEEE Netw. 19 (4) (2005) 6–13. [4] M.M. Rashid, E. Hossain, V.K. Bhargava, Controlled channel access scheduling for guaranteed QoS in 802.11e-based WLANs, IEEE Trans. Wirel. Commun. 7 (4) (2008) 1287–1297. [5] X. Yu, P. Navaratnam, K. Moessner, Resource reservation schemes for IEEE 802.11-based wireless networks: a survey, IEEE Commun. Surv. Tut. 15 (3) (2013) 1042–1061. [6] G. Bianchi, Performance analysis of the IEEE 802.11 distributed coordination function, IEEE J. Sel. Area Commun. 18 (3) (20 0 0) 535–547. [7] P. Chatzimisios, A.C. Boucouvalas, V. Vitsas, IEEE 802.11 packet delay – a finite retry limit analysis, in: Proc. IEEE Globecom, vol. 2, 2003, pp. 950–954. [8] P. Chatzimisios, A.C. Boucouvalas, V. Vitsas, IEEE 802.11wireless LANs: performance analysis and protocol refinement, EURASIP J. Appl. Signal Processing (2005) 67–78. [9] O. Tickoo, B. Sikdar, Queueing analysis and delay mitigation in IEEE 802.11 random access MAC based wireless networks, in: Proc. IEEE INFOCOM, 2004, pp. 1404–1413. [10] A. Banchs, P. Serrano, A. Azcorra, End-to-end delay analysis and admission control in 802.11 DCF WLANs, Comput. Commun. 29 (7) (April 2006) 842–884. [11] H. Zhai, X. Chen, Y. Fang, How well can the IEEE 802.11 wireless LAN support quality of service? IEEE Trans. Wirel. Commun. 4 (6) (2005). [12] T. Sakurai, H. Vu, MAC access delay of IEEE 802.11 DCF, IEEE Trans. Wirel. Commun. 6 (5) (May 2007) 1702–1710. [13] K. Duffy, D. Malone, D. Leith, Modeling the 802.11 distributed coordination function in non-saturated conditions, IEEE Commun. Lett. 9 (8) (August 2005) 715–717. [14] C. Foh, M. Zukerman, J. Tantra, A Markovian framework for performance evaluation of IEEE 802.11, IEEE Trans. Wirel. Commun. 6 (4) (April 2007) 1276–1265. [15] D. Malone, K. Duffy, D. Leith, Modeling the 802.11 distributed coordination function in nonsaturated heterogeneous conditions, IEEE/ACM Trans. Netw. 15 (1) (February 2007) 159–172. [16] R.P. Liu, G.J. Sutton, I.B. Collings, A new queueing model for QoS analysis of IEEE 802.11 DCF with finite buffer and load, IEEE Trans. Wirel. Commun. 9 (8) (August 2010) 2664–2675. [17] Q. Zhao, D.H.K. Tsang, T. Sakurai, A simple and approximate model for nonsaturated IEEE 802.11 DCF, IEEE Trans. Mob. Comput. 8 (11) (November 2009) 1539–1553. [18] R.P. Liu, G.J. Sutton, I.B. Collings, A new queueing model for QoS analysis of IEEE 802.11 DCF with finite buffer and load, IEEE Trans. Wirel. Commun. 9 (8) (August 2010) 2664–2675. [19] L. Dai, X. Sun, A unified analysis of IEEE 802.11 DCF networks: stability, throughput, and delay, IEEE Trans. Wirel. Commun. 12 (8) (August 2013) 1558–1572. [20] A.L. Toledo, T. Vercauteren, X. Wang, Adaptive optimization of IEEE 802.11 DCF based on Bayesian estimation of the number of competing terminals, IEEE Trans. Mob. Comput. 5 (9) (September 2006) 1283–1296. [21] D. Deng, C. Ke, H. Chen, Y. Huang, Contention window optimization for IEEE 802.11 DCF access control, IEEE Trans. Wirel. Commun. 7 (12) (December 2008) 5129–5135. [22] J. Choi, J. Yoo, S. Choi, C. Kim, EBA: an enhancement of the IEEE 802.11 DCF via distributed reservation, IEEE Trans. Mob. Comput. 4 (4) (July/August 2005) 378–390. [23] J. Choi, J. Yoo, C. Kim, A distributed fair scheduling scheme with a new analysis model in IEEE 802.11 wireless LANs, IEEE Trans. Veh. Technol. 57 (5) (September 2008) 3083–3093.

Y.-W. Kuo, J.-H. Huang / Computer Networks 127 (2017) 31–42 [24] Y.W. Kuo, T.L. Tsai, Fixed contention window backoff scheme for the contention-based IEEE 802.11 MAC, in: Proc. of IEEE TENCON, Fukuoka, Japan, 2010, pp. 847–851. [25] Y.W. Kuo, T.L. Tsai, Design and evaluation of a contention-based high throughput MAC with delay guarantee for IEEE 802.11 WLANs, J. Commun. Netw. 15 (6) (2013) 606–613. [26] Y.W. Kuo, W.F. Lu, Delayed contention DCF MAC protocol for IEEE 802.11 wireless LANs, in: Proc. of IEEE VTC-Spring, Yokohama, Japan, 2012. [27] D. Stiliadis, A. Varma, Latency-rate servers: a general model for analysis of traffic scheduling algorithms, IEEE/ACM Trans. Netw. 6 (5) (October 1998) 611–624.

41

[28] Y. Jiang, Relationship between guaranteed rate server and latency rate server, Comput. Netw. 43 (3) (October 2003) 307–315. [29] The network simulator – NS2, http://www.isi.edu/nsnam/ns/index.html. [30] J. Barcelo, B. Bellalta, C. Cano, M. Oliver, Learning BEB: Avoiding Collisions in WLAN, Eunice Summer School, 2008. [31] J. Lee and J. Walrand, Design and analysis of an asynchronous zerocollision MAC protocol, Arxiv preprint 0806.3542v1 [cs.NI], 2008. [32] M. Fang, D. Malone, K.R. Duffy, D.J. Leith, Decentralised learning MACs for collision-free access in WLANs, Wirel. Netw. 19 (1) (2013) 83–98.

42

Y.-W. Kuo, J.-H. Huang / Computer Networks 127 (2017) 31–42 Yaw-Wen Kuo received the B. S. and the M. S. degrees in Electrical Engineering from National Tsing Hua University, HsinChu, Taiwan, ROC, in 1992 and 1994, respectively, and the Ph.D. degree in Communication Engineering from National Chiao Tung University, HsinChu, Taiwan, ROC, in 20 0 0. After three months of military service, he joined ZyXEL Communications Corps. in HsinChu Science Park, Taiwan as a hardware project leader in central-office equipment BU. Since 2006, he joined the National Chi Nan University, Nantou, Taiwan, ROC, where he is currently an Associate Professor of Electrical Engineering. His research presently focuses on wireless networks and embedded system.

Jane-Hwa Huang received the B.S., M.S., and Ph.D. degrees from the National Cheng Kung University, Tainan, Taiwan, R.O.C., in 1994, 1996, and 2003, respectively, all in electrical engineering. Since 2004, he has been with the Department of Communication Engineering, National Chiao Tung University, Hsinchu, Taiwan. Currently he is an assistant professor in the Department of Electrical Engineering, National Chi Nan University, Nantou, Taiwan, ROC. His current research interests are in the areas of performance evaluation, wireless networks, and multi-hop communications.