VIEWS: 26 PAGES: 6 CATEGORY: Education POSTED ON: 1/12/2010 Public Domain
Enhanced DCF of IEEE 802.11e to Support QoS Yang Xiao Computer Science Division The University of Memphis, Memphis, TN 38152 E-mail: yangxiao@ieee.org Abstract—In this paper, we will introduce the emerging IEEE In this paper, we focus us on EDCF, but not HCF. 802.11e standard to support quality of service at medium access Several priority studies have been reported in the literature control level. One of the most important functions in 802.11e is for DCF. Deng and Chang [2] proposed a priority scheme by the contention-based channel access mechanism called enhanced differentiating the backoff window: the higher priority class distributed coordination function (EDCF), which provides a priority scheme by differentiating the inter-frame space and the uses the window [0,2i+1 − 1] and the lower priority class uses initial window size. We propose an analytical model to evaluate the window [2i +1 ,2i + 2 − 1] , where i is the backoff stage. In [3], the EDCF priority scheme. Saturation throughput and saturation Aad and Castelluccia proposed a priority scheme by delay are derived analytically. Simulations are also conducted to differentiating inter-frame spaces (IFS’s), in which a higher validate analytical results. Our study shows that differentiating priority class uses IFS, whereas a lower priority class uses a the initial window size is better than differentiating the inter- frame space in terms of total throughput and delay. space that equals the sum of IFS and the maximum window size. In [4], Veres and Campbell et al. proposed priority Keywords- IEEE 802.11 protocol, performance evaluation. schemes by differentiating the minimum backoff window size and the maximum window size. In [22], Pallot and Miller I. INTRODUCTION proposed three priority schemes: static priority scheduling, prioritized DIFS time mechanism, and prioritized backoff time The IEEE 802.11 enables fast installation, with minimum distribution mechanism. In [23], Vaidya et al. proposed a fair management and maintenance costs, and is a very robust scheduling method. All the schemes [2-4, 22-23] are based on protocol for the best–effort service in the wireless medium, but simulations. it is unsuitable for multimedia applications with Quality of There have been many performance studies for IEEE Service (QoS) requirements. 802.11 [5-18]: both simulation [18] and analytical [14] studies. IEEE 802.11 medium access control (MAC) employs a Bing [12, 17] provided a good performance analysis by a mandatory contention-based channel access function called quantitative approach. Chhaya and Gupta [15] calculated the Distributed Coordination Function (DCF), and an optional throughput of CSMA/CA using a simple model with the centrally controlled channel access function called Point probabilities of capture and the presence of hidden stations. Coordination Functions (PCF) [1]. DCF adopts a carrier sense Bianchi [5, 11] and Ziouva et al. [6] presented an accurate multiple access with collision avoidance (CSMA/CA) with analysis to compute the saturation throughput performance. binary exponential backoff. DCF does not support QoS, and Huang, Ho, and Chen [13, 14] gave approximate models that even cannot support any priority scheme. PCF can support account for hidden terminals and capture effects. Cali, Conti very limited QoS. It is an optional function, and has not been and Gregori [9, 10] improved the performance by tuning implemented in most of current products. To support MAC- backoff strategies, and provided an excellent performance level QoS, the IEEE 802.11 standardization committee is analysis. Tay and Chua [15] provided a good approximation of saturation throughput. All theses schemes [5-18] are for currently working on IEEE 802.11e [20], a supplement to the original IEEE 802.11 DCF. original IEEE 802.11 MAC. The IEEE 802.11e MAC will There are very limited reports in the literature about IEEE support multimedia applications such as voice and video over 802,11e since it is new. Mangold and Choi et al. [21] the IEEE 802.11 WLANs. introduced IEEE 802.11e and provided performance The IEEE 802.11e MAC also employs a contention-based evaluation via simulations. channel access function called Enhanced Distributed In this paper, we will introduce EDCF priority scheme in Coordination Function (EDCF), and a centrally controlled section III, and based on Bianchi’s [5] and Ziouva’s [6] channel access function called Hybrid Coordination Function models, propose an analytical model to evaluate the EDCF (HCF). EDCF provides a priority scheme by differentiating priority scheme in section IV. Saturation throughput and the inter-frame space and the initial window size. Such a saturation delay are derived analytically. Simulation and priority scheme may not provide guaranteed QoS but analytical results are studied in section V. prioritized QoS due to the contention-based nature. Prioritized QoS will be useful for those multimedia applications that can II. CSMA/CA IN DCF live without rigid QoS. One advantage of prioritized QoS is We will briefly introduce DCF’s CSMA/CA mechanism, that it is simple to implement and looks like DiffServ model. which is useful for understanding our analytical model. 0-7803-7700-1/03/$17.00 (C) 2003 IEEE 1291 Readers can refer to [1] for details. In DCF, a station with a in the original IEEE 802.11 MAC. IEEE 802.11e allows a frame to transmit monitors the channel activities until an idle station to have fewer queues, but we will not address this period equal to a distributed inter-frame space (DIFS) is situation in this paper. detected. After sensing an idle DIFS, the station waits for a Assuming that a payload from a higher layer is labeled with random backoff interval before transmitting. The backoff time a priority value, it is pushed into the corresponding queue with counter is decremented in terms of slot time as long as the the same priority value. Each queue acts as an independent channel is sensed idle. The counter is stopped when a MAC entity and performs similar DCF function, introduced in transmission is detected on the channel, and reactivated when the previous section, with a different inter-frame space and a the channel is sensed idle again for more than a DIFS. The different initial window size. station transmits its frame when the backoff time reaches zero. At each transmission, the backoff time is uniformly chosen in the range (0, CW − 1) , where CW is the current backoff IV. AN ANALYTICAL MODEL window size. At the very first transmission attempt, CW equals Assume that any queue in any station always has frames the minimum backoff window size CWmin . After each ready to send. If further assuming that there are n active unsuccessful transmission, CW is doubled until a maximum stations in a Basic Service Set (BSS) and each station backoff window size value is reached. After the destination implements all 8 queues, there are total 8n queue entities, station successfully receives the frame, it transmits an which are equivalent to 8n stations in the original IEE 802.11 acknowledgment frame (ACK) following a short inter-frame DCF. We further lose above assumption so that we let space (SIFS) time. If the transmitting station does not receive ni (i = 0,..., 7) denote the number of active queues in the the ACK within a specified ACK Timeout, or it detects the priority i class. We only consider the basic access mechanism transmission of a different frame on the channel, it reschedules in this paper. The method in this paper can be easily applied to the frame transmission according to the previous backoff rules. the optional RTS/CTS mechanism. III. ENHANCED DCF (EDCF) PRIORITY SCHEME A. An Analytical Model The contention-based channel access function of IEEE For a given station in the priority i class (i = 0,...7) , b(i, t ) 802.11e is EDCF, in which multiple queues (up to 8) are used is defined as a random process representing the value of for different priorities (up to 8). Priorities are achieved by backoff counter at time t , and s(i, t ) is defined as the random differentiating the arbitration inter-frame space and the initial process representing the backoff stage j , where 0 ≤ j ≤ m window size. The traffic is classified into 8 priority and m is the maximum backoff stage. The value of the classes: i = 0,...7 . For the priority i class, the minimum backoff counter is uniformly chosen in the range backoff window size is CWmin [i ] , and the arbitration inter- (0,1,...Wi , j − 1) , where Wi , j = 2 j Wi ,0 and Wi ,0 = CWmin [i ] (we frame space is AIFS [i ] . If the priority i class has a lower priority than the priority j class, we have use a different notation for convenience). Let pi (i = 0,...7) CWmin [i ] ≥ CWmin [ j ] , and AIFS [i ] ≥ AIFS [ j ] , and at least one denote the probability that a transmitted frame collides, and pb denote the probability that the channel is busy. Similar to of above inequalities must be a real inequality. In other words, EDCF adopts AIFS [i ] ( i = 0,1,...,7 ) and CWmin [i ] Bianchi’s model [5] and Ziouva’s model [6], the bi- dimensional random process {s (i, t ), b(i , t )} is discrete-time ( i = 0,1,...,7 ) instead of DIFS and CWmin , respectively. If one Markov chain under the assumptions that the probability pi class has a smaller AIFS or CWmin , the class’s traffic has a better chance to access the wireless medium earlier. and the probability pb are both independent to the backoff Fig. 1 shows the EDCF timing diagram, where 3 priorities procedure. Therefore, the state of each station in the priority i are shown: i , j , and k . We can see that AIFS [i ] > PIFS , class is described by {i, j, k } , where i is just for an index where PIFS is point (coordination function) inter-frame purpose standing for the priority i class, j stands for the space. backoff stage and takes values (0,1,...m ) , and k stands for the AIFS[k] backoff delay and takes values (0,1,...Wi , j − 1) in timeslots. Immediate access when AIFS[j] Medium is free >= AIFS [i] AIFS[i] The state transition diagram for the priority i class is shown Contention Window from AIFS[i] PIFS 0 to CWmin[i] in Fig. 2, where the state {i, −1, 0} stands for the state that the Busy Medium SIFS Backoff-Window Next Frame station senses the channel idle after AIFS [i ] and transmits Slot time successfully without activating the backoff stage. Defer Access Select Slot and Decrement Backoff as long The non-null transition probabilities are listed as follows. as medium is idle Pr{i , −1, 0 | i , −1, 0} = (1 − pi )(1 − pb ) Fig. 1 EDCF timing diagram Normally, each station implements 8 queues, and each Pr{( i,0, k ) | ( i, −1,0 )} = 1 − (1 − pi )(1 − pb ) Wi ,0 , for 0 ≤ k ≤ Wi ,0 − 1 queue supports one priority, behaving as a single DCF entity Pr{( i, j, k ) | ( i, j , k )} = pb , for 1 ≤ k ≤ Wi , j − 1 and 0 ≤ j ≤ m 1292 Pr{( i , j , k ) | ( i , j , k + 1)} = 1 − pb , for 0 ≤ k ≤ Wi , j − 2 and 0 ≤ j ≤ m Substituting (8)-(9) to (7), we can solve unknown Pr{( i,0, k ) | ( i , j, 0 )} = (1 − pi ) pb Wi ,0 , for 0 ≤ k ≤ Wi ,0 − 1 and 0 ≤ j ≤ m parameters numerically. Then, we can calculate pb and pi Pr{( i, −1,0) | ( i, j,0)} = (1 − pi )(1 − pb ) , for 0 ≤ j ≤ m from (8)-(9). Pr{( i, j, k ) | ( i, j − 1,0 )} = pi Wi , j , for 0 ≤ k ≤ Wi , j − 1 and 1 ≤ j ≤ m B. Saturation Throughput Pr{( i, m, k ) | ( i, m,0)} = pi Wi ,m , for 0 ≤ k ≤ Wi ,m − 1 Let ps ,i (i = 0,...7) denote the probability that a successful (1-pi)(1-pb) transmission occurs in a slot time for the priority i class. We (1-pi)(1-pb) i,-1,0 [1-(1-pi)(1-pb)]/Wi, 0 have 7 nτ ps ,i = niτ i (1 − τ i ) n −1 ∏ (1 − τ h ) n = i i (1 − pb ) (1-pi )pb/Wi, 0 i (10) h 1-pb 1-pb h = 0, h ≠ i 1 −τ i 1-pb 1-pb 1-pb … i,0,0 pb i,0,1 i,0,2 i,0,Wi, 0-2 i,0,Wi, 0-1 Let Si (i = 0,...7) denote the normalized throughput for the pb pb pb pi/Wi, 1 priority i class. Let δ , L , Ts ,i , and Tc ,i denote the duration . . . . . . . . . . of an empty slot time, payload size, the average time that the . . . . . channel is sensed busy because of a successful transmission i,j-1,0 . . . . for the priority i class, and the average time that the channel pi/Wi, j 1-pb 1-pb has a collision for the priority i class, respectively. We have, E ( payload information transmitted in a slot time for the priority i class ) 1-pb 1-pb 1-pb i,j,0 i,j,1 i,j,2 … i, j,Wi, j-2 i, j,Wi, j-1 pb pb Si = (11) pb pb E ( length of a slot time ) p s ,i E ( L ) = . pi/Wi, j+1 . . . . (1 − pb ) δ + ps ,iTs ,i + pb − ps ,i Tc,i . . . . . . . . pi/Wi, m . . Let TH , TE ( L ) , TACK , SIFS , L * , TE ( L*) , and γ denote the 1-pb 1-pb 1-pb … 1-pb 1-pb i, m, Wi, m-1 time to transmit the header (including MAC header, PHY i, m, 0 i, m, 1 i, m, 2 i, m, Wi, m-2 pb pb pb header, and/or tail), the time to transmit the payload, the time pb to transmit the ACK, the time of SIFS, the length of the pi/Wi, m longest frame in a collision, the time to transmit a payload Fig.2 The state transition diagram for the priority i class with length E ( L*) , and the time of the propagation delay, Let bi , j ,k = lim Pr{s (i , t ) = j , b(i , t ) = k } be the stationary respectively. We have, t →∞ distribution of the Markov chain. In steady state, we can Ts ,i = TH + TE ( L ) + SIFS + γ + TACK + AIFS [i ] + γ (12) derive following relations through chain regularities. Tc ,i = TH + TE ( L*) + AIFS [i ] + γ (13) bi , j ,0 = pi j bi ,0,0 0 ≤ j ≤ m -1 (1) C. Saturation Delay pm bi ,m,0 = i bi ,0,0 (2) The conditional probability p s ,i pb ( i = 0,...7) is the 1 − pi probability that a transmitted frame for the priority i class is Wi , j − k 1 bi , j ,k = bi , j ,0 0 ≤ j ≤ m,1 ≤ k ≤ Wi , j -1 (3) successfully transmitted. One over this probability is the Wi , j 1 − pb average number of retransmissions for the priority i class, i.e. 1 − (1 − pi )(1 − pb ) (4) pb ps ,i . Let N c ,i (i = 0,...7) denote the random variable bi ,0,0 = bi , −1,0 1 − pb representing the number of collisions before transmitting a m Wi , j −1 frame for the priority i class. We have bi ,−1,0 + ∑ ∑b =1 (5) E ( N c , i ) = p b p s ,i − 1 i , j ,k j =0 k =0 (14) Let α1 = 2 (1 − pb ) 2 (1 − pi )(1 − 2 pi ) and α 2 = [1 − (1 − pi )(1 − pb )] . The average backoff delay depends on the value of a Plugging (1)-(4) into (5), we have, station’s backoff counter and the duration when the counter α1 (6) freezes due to others’ transmissions. Let X i ( i = 0, ...7) denote bi ,−1,0 = α1 + α 2 (1 − 2 pi ) + (1 − pi + 2m pi m+1 − pi m+1 2 m+1 )Wi ,0 the random variable representing the time interval during which the counter reaches zero without considering the case Let τ i be the probability that a station in the priority i when the counter freezes. We have class transmits during a generic slot time. We have m Wi , j −1 bi ,0,0 Wi ,0 2 1 − pi − 3 pi ( 4 pi ) + 4 pi − 1 m (15) E ( Xi ) = ∑ m 2 (1 − pb )(1 − 2 pi ) (7) ∑ kb = 6 (1 − pb ) (1 − 4 pi )(1 − pi ) τ i = ∑ bi , j ,0 = i , j ,k j = 0 k =1 (1 − 2 pi ) + (1 − pi + 2m pi m+1 − pi m+1 2m +1 ) Wi ,0 α1 + α2 j =−1 Let Fi (i = 0,...7) and N F (i = 0,...7) denote the time that the i 7 pb = 1 − ∏ (1 − τ h ) nh (8) backoff counter of a station freezes and the number of times h =0 that the backoff counter freezes, respectively, for the priority i −1 7 pi = 1 − ∏ (1 − τ h ) nh (1 − τ i ) ni −1 ∏ (1 − τ h ) nh (9) i class. The mean number of consecutive idle slot times is h =0 h =i +1 1293 (1 − pb ) pb . Therefore, we have, results match pretty well. As the number of active queues E( X i ) increases, saturation throughputs for both classes decrease, ( ) E N Fi = −1 (16) and saturation delays for both classes increase. Fig. 4 also max((1 − pb ) pb ,1) shows that class 1 is more sensitive to the number of active Since ps ,i is the probability that a successful transmission queues than class 0. As illustrated in the figures, the class 0 occurs in a slot time, the probability that the transmitted frame has a much better saturation throughput and saturation delay is successful is ps ,i pb . Therefore, we have than the class 1. Therefore, the EDCF priority scheme is quite effective. p p (17) ( ) E ( Fi ) = E N Fi s ,i Ts ,i + (1 − s ,i )Tc ,i x 10 6 pb pb 4 Let Bi (i = 0,...7) denote the backoff delay of a station for 3.5 the priority i class before accessing the channel under busy 3 channel condition. We have 2.5 Priority Priority 0 0 class: class: Simulation Analytic E ( Bi ) = E ( X i ) + E ( Fi ) Priority 1 class: Simulation (18) Priority 1 class: Analytic 2 Let Di (i = 0,...7) denote the random variable representing the 1.5 frame delay for the priority i class. Let To denote the time 1 0.5 that a station has to wait when its frame transmission collides before sensing the channel again. Let Ttimeout denote the 0 5 10 15 20 25 30 35 40 45 Number of active queues duration of the ACK timeout in time slots. We have Fig. 4 Saturation Delay (µ sec.) E ( Di ) = E ( N c ,i ) E ( Bi ) + Tc ,i + To + E ( Bi ) + Ts ,i (19) 0.55 To = SIFS + Ttimeout (20) 0.5 0.45 Priority 0 class: Simulation V. NUMERICAL AND SIMULATION RESULTS 0.4 Priority 0 class: Analytic Priority 1 class: Simulation Priority 1 class: Analytic 0.35 The simulation models had been developed based on the 0.3 IEEE 802.11e draft [20], the IEEE 802.11a standard [8], and OPNET Wireless LAN simulation model version 8.0A. 0.25 Parameters of IEEE 802.11a can be found in [7-8], as well as 0.2 how to calculate TH + TE ( L ) accurately [7]. Both the data rate 0.15 15 20 25 30 35 CWmin of class 1 40 45 50 and the control rate are 6Mbps. The frame size is fixed as Fig. 5 Saturation Throughput (Normalized) 1024 bytes. For demonstration purpose, we adopt two priority B. Effects of the initial window size classes instead of 8 priorities. Fig. 5 and Fig. 6 have following parameters: A. Effectiveness of the EDCF priority scheme AIFS [0] = AIFS [1] = PIFS , W = 16 and n0 = n1 = 10 . Fig. 5 (Fig. 6) 0,0 0.55 shows saturation throughputs (saturation delays) over the minimum window size of class 1, W1,0 , which changes from Priority 0 class: Simulation Priority 0 class: Analytic Priority 1 class: Simulation 0.5 Priority 1 class: Analytic 0.45 16 to 48. As illustrated in Fig. 5 (Fig. 6), when W1,0 = 16 , 0.4 saturation throughputs (saturation delays) are the same for 0.35 both classes. As W1,0 increases, the saturation throughput of 0.3 class 1 decreases, the saturation throughput of class 0 increases, the saturation delay of class 0 decreases, and the saturation delay of class 1 increases dramatically. Therefore, 0.25 0.2 saturation delay is very sensitive to the minimum window size. 0.15 5 10 15 20 25 30 35 40 45 We also observe in Fig. 6, as W1,0 increases, the saturation Number of active queues Fig. 3 Saturation Throughput (Normalized) delay of class 0 does not change a lot, and it is good for delay Fig. 3 and Fig. 4 have following parameters: sensitive applications. An interesting observation in Fig. 5 is [ AIFS [0], AIFS [1]] = [ PIFS , DIFS ] , W0,0 ,W1,0 = [16,32] , and n0 = n1 , that the throughputs of class 0 and class 1 are symmetric along a line parallel to x-axis. This phenomenon indicates that class where PIFS = SIFS + SLOT = 25 µ sec. and 0 can steal throughput from class 1 as the initial window size DIFS = SIFS + 2* SLOT = 34 µ sec. [1, 8]. Fig. 3 (Fig. 4) shows of class 1 increases, whereas the total throughput of all classes saturation throughputs (saturation delays) for two priority does not change much. The reason that class 0 can steal classes over the number of active queues ( n0 or n1 ). As bandwidth from class 1 is stated as follows. As the initial illustrated in both figures, analytical results and simulation window size of class 1 increases, there are two direct effects. 1294 The first effect is that entities in class 1 will delay accessing the saturation throughput and the saturation delay of class 0 the channel so that the saturation throughput of class 1 will will remain the same. decrease and the saturation delay of class 1 will increase. The We can conclude that differentiating the initial window size second effect is that as the initial window size of class 1 is a better than differentiating the inter-frame space in terms of increases, collision probabilities of class 0 will decrease so total throughput and delay. However, differentiating the inter- that the probability that a transmitted frame from class 0 is frame space gives a very fast way to access the channel in successful becomes larger. Therefore, the saturation favor of the class with a short AIFS. throughput of class 0 will increase and the saturation delay of 0.3405 class 0 will decrease. In summary, differentiating the initial Priority 0 class: Simulation Priority 0 class: Analytic Priority 1 class: Simulation window size has both the function of reducing collisions and 0.34 Priority 1 class: Analytic the function of providing priorities, 0.3395 6 x 10 2.5 0.339 0.3385 2 0.338 1.5 0.3375 0.337 1 25 26 27 28 29 30 31 32 33 34 AIFS[1] (class 1) Priority 0 class: Simulation Priority 0 class: Analytic Fig. 7 Saturation Throughput (Normalized) Priority 1 class: Simulation 0.5 Priority 1 class: Analytic 5 x 10 4.995 0 4.99 15 20 25 30 35 40 45 50 CWmin of class 1 Fig. 6 Saturation Delay (µ sec.) 4.985 C. Effects of the arbitration inter-frame space (AIFS) 4.98 Fig. 7 and Fig. 8 have following parameters: 4.975 AIFS [0] = PIFS , W0,0 = W1,0 = 16 and n0 = n1 = 10 . Fig. 7 4.97 Priority 0 class: Simulation Priority 0 class: Analytic Priority 1 class: Simulation (Fig. 8) shows saturation throughputs (saturation delays) over Priority 1 class: Analytic 4.965 the arbitration inter-frame space of class 1, AIFS [1] , which 4.96 changes from PIFS to DIFS . As illustrated in the figures, 25 26 27 28 29 30 31 32 33 34 AIFS[1] (class 1) when AIFS [1] = PIFS = 25 µ sec. , saturation throughputs Fig. 8 Saturation Delay (µ sec.) (saturation delays) are the same for both classes. As AIFS [1] D. Effects of the number of active entities increases, the saturation throughput of class 1 decreases Fig. 9 and Fig. 10 have following parameters: linearly, and the saturation delay of class 1 increases linearly. [ AIFS [0], AIFS [1]] = [ PIFS , DIFS ] , W0,0 ,W1,0 = [16,32] , and n0 = 10 . An interesting observation is that both the saturation Fig. 9 (Fig. 10) shows saturation throughputs (saturation throughput and the saturation delay remain constant as delays) over the number of active queues of class 1, n1 , which AIFS [1] increases. This fact indicates that the arbitration inter-frame space of class 1 does not affect the saturation changes from 5 to 45. As illustrated in the figures, as n1 throughput and the saturation delay of class 0 as long as increases, the saturation throughput of class 0 decreases, and AIFS [0] ≤ AIFS [1] . Compared to effects of the initial window the saturation delay of class 0 increases. An observation is that size in Fig. 5 and Fig. 6, effects of the arbitration inter-frame the saturation throughput of class 1 increases as n1 increases. space in Fig. 7 and Fig. 8 show very different characteristics. The reason is that there are more active entities of class 1 so The reason is that the initial window size and the arbitration that the total saturation throughput of class 1 increases. inter-frame space have different effects on priorities: the initial Another interesting observation is that the saturation delay of window size has both the function of reducing collisions and class 1 decrease at first and then increases later. From the the function of providing priorities, whereas the arbitration parameters, we know that class 0 has a higher priority than inter-frame space has the function of providing priorities by class 1, but how come that class 0 gets worse total throughput accessing channel earlier/later, but not the function of and delay than class 1 in Fig. 9 and Fig. 10, respectively, when reducing collisions. Therefore, contrast to Fig.5 and Fig. 6, as the number of active queues of class 1 is large? The reasons the arbitration inter-frame space of class 1 increases, class 0 are stated as follows. There are more active entities of class 1 cannot steal bandwidth from class 1. The reasons are stated as than class 0 so that the total saturation throughput of class 1 follows. As the arbitration inter-frame space of class 1 may be larger than that of class 0 even though the throughput increases, entities in class 1 will delay accessing the channel per entity in class 1 is smaller than that in class 0. To so that the saturation throughput of class 1 will decrease and understand the delay part becomes a little tricky. In fact, due the saturation delay of class 1 will increase. However, to different AIFSs and different initial window sizes for two collision probabilities of class 0 will remain the same so that classes, current windows of two classes have an overlap, and 1295 the size of such an overlap is changing dynamically. When the [2] D.-J. Deng and R.-S. Chang, “A priority Scheme for IEEE 802.11 DCF Access Method,” IEICE Trans. Communications., Vol. E82-B, No.1, number of active entities of class 1 is very large, the number Jan. 1999, pp.96-102. of collisions in the overlap portion become much more severe [3] I. Aad and C. Castelluccia, “Differentiation Mechanisms for IEEE than the non-overlap portion. Therefore, the delay 802.11,“ IEEE INFOCOM 2001. performance for class 0 depends on the ratio of the non- [4] A. Veres, A. T. Campbell, M. Barry, and L.-H. Sun, “Supporting Differentiation in Wireless Packet Networks Using Distributed Control,” overlap portion with the overlap portion in class 0. If the non- IEEE J-SAC, Vol. 19. No. 10, Oct. 2001, pp. 2081-2093. overlap portion in class 0 is relatively small and the non- [5] G. Bianchi, “IEEE 802.11--Saturation Throughput Analysis,” IEEE overlap portion in class 1 is relatively large, the average delay Communications Letters Vol. 2, No. 12, Dec.1998, pp. 318-320. of class 0 could be larger than class 1. [6] E. Ziouva and T. Antonakopoulos, “CSMA/CA performance under high traffic conditions: throughput and delay analysis,” Computer 0.6 Priority 0 class: Simulation Communications, 25 (2002), pp.313-321. [7] Y. Xiao and J. Rosdahl, “Throughput and Delay Limits of IEEE Priority 0 class: Analytic 0.55 Priority 1 class: Simulation Priority 1 class: Analytic 0.5 802.11,” IEEE Communications Letters, Vol. 6, No. 8, Aug. 2002, pp. 0.45 355-357. 0.4 [8] IEEE 802.11a WG, Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) specification: High-speed Physical 0.35 Layer in the 5GHz Band, Sep. 1999. 0.3 [9] F. Calì, M. Conti, and E. Gregori, “Dynamic Tuning of the IEEE 802.11 0.25 Protocol to Achieve a Theoretical Throughput Limit,” IEEE/ACM 0.2 Trans. Networking, Vol. 8, No. 6, Dec. 2000, pp. 785-790. 0.15 [10] F. Cali, M. Conti, and E. Gregori, “IEEE 802.11 Protocol: Design and Performance Evaluation of an Adaptive Backoff Mechanism,” IEEE J- SAC, Vol. 18, No. 19, Sep. 2000, pp. 1774-1786. 0.1 5 10 15 20 25 30 35 40 45 Number of active queues in class 1 Fig. 9 Saturation Throughput (Normalized) [11] G. Bianchi, “Performance Analysis of the IEEE 802.11 Distributed x 10 5 Coordination Function,” IEEE J-SAC, Vol. 18, No. 3, Mar. 2000, pp. 18 Priority 0 class: Simulation 535-547. [12] B. Bing and R. Subramanianb, “A novel technique for quantitative Priority 0 class: Analytic 16 Priority 1 class: Simulation Priority 1 class: Analytic 14 performance evaluation of wireless LANs,” Computer Communications, Vol. 21, No. 9, July 1998, pp.833-838. [13] K. C. Huang and K.-C. Chen, "Interference analysis of nonpersistent 12 10 CSMA with hidden terminals in multicell wireless data networks", Proc. 8 IEEE PIMRC, Toronto, Canada, Sept. 1995, pp. 907-911. 6 [14] T. S. Ho and K. C. Chen, "Performance evaluation and enhancement of the CSMA/CA MAC protocol for 802.11 wireless LAN's", Proc. IEEE 4 PIMRC, Taipei, Taiwan, Oct. 1996, pp. 392-396. 2 [15] H. S. Chhaya and S. Gupta, "Performance modeling of asynchronous 0 5 10 15 20 25 30 35 40 45 data transfer methods of IEEE 802.11 MAC protocol", Wireless Number of active queues in class 1 Networks, vol.3, pp. 217-234, 1997. Fig. 10 Saturation Delay (µ sec.) [16] Y. C. Tay and K. C. Chua, “A Capacity Analysis for the IEEE 802.11 MAC Protocol,” Wireless Networks 7, 2001, pp. 159-171. VI. CONCLUSIONS [17] B. Bing, “Measured Performance of the IEEE 802.11 Wireless LAN,” IEEE LCN’99. In this paper, we introduced the emerging IEEE 802.11e [18] G. Bianchi, L. Fratta, and M. Oliveti, “Performance evaluation and standard to support quality of service at medium access enhancement of the CSMA/CA MAC protocol for 802.11 Wireless control level. An analytical model was proposed to study LANs,” Proc. PIMRC 1996, Taipei, Taiwen 1996, pp. 392-396. [19] Y. Xiao and J. Rosdahl, “A Performance Analysis of IEEE 802.11a EDCF priority scheme in terms of saturation throughput and Wireless LAN,” Proc. of The 6th World Multi-Conference on saturation delay. Our study shows following results: SYSTEMICS, CYBERNETICS AND INFORMATICS, (SCI 2002), • The EDCF priority scheme is quite effective. July 14-18, 2002, Orlando, Florida, U.S.A, pp. 243-248. [20] IEEE 802.11 WG, Draft Supplement to Part 11: Wireless Medium • Differentiating the initial window size is a better than Access Control (MAC) and physical layer (PHY) specifications: differentiating the inter-frame space in terms of total Medium Access Control (MAC) Enhancements for Quality of Service throughput and delay. However, differentiating the inter- (QoS), IEEE 802.11e/D2.0, Nov. 2001. [21] S. Mangold, S. Choi, P. May, O. Kein, G. Hiertz, and L. Stibor, “IEEE frame space gives a very fast way to access the channel in 802.11e Wireless LAN for Quality of Service,” Proc. European Wireless favor of the class with a short AIFS. ’02, Florence, Italy, February 2002. • Simulation results match analytical results very well. [22] X. Pallot and L. E. Miller, "Implementing message priority policies over The results of this paper are beneficial in designing good an 802.11 based mobile ad hoc network," IEEE MILCOM 2001. N. Vaidya, P. Bahl, and S. Gupta, “Distributed Fair Scheduling in a Wireless prioritized QoS parameters. LAN,” ACM MOBICOM 2000. REFERENCES [1] IEEE 802.11 WG, Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) specification, Aug. 1999. 1296