Power Eiciency in IEEE 802.11a WLA with Cross-Layer Adaptation Jun Zhao, Zihua Guo, and Wenwu Zhu Microsot Research Asia 3/F, Beijing Sigma Center, o.49, Zhichun Road, Haidian District Beijing 100080, P.R.China Abstract Power consumption is one o the most stringent bottlenecks in mobile devices. In this paper, we investigate the joint eect o MAC and PHY layers on power eiciency in IEEE 802.11a WLA. Speciically, we study the link adaptation or a power eicient transmission by selecting a proper transmission mode and power level with the aid o our derived power eiciency model. This study addresses the undamental impact o the CSMA MAC protocol on the power eiciency o IEEE 802.11a WLAs. Some implications or system design are also discussed. In particular, we show that the non-radio-transmission power plays an important role in the power optimization o IEEE 802.11a WLA. I. ITRODUCTIO Power eiciency is an important issue or mobile devices, which are usually battery supplied. Compared with bandwidth, there is limited improvement in battery capacity. Thus, to reduce the power consumption, or equivalently, enhance the mobile device power eiciency is critical or wireless systems. IEEE 802.11a is a WLA standard or high-speed transmission in the 5GHz band [2]. It supports eight transmission modes and provides data rates up to 54Mb/s. Each mode uses dierent coding rate and modulation level. Generally, the higher the bit rate, the more vulnerable to channel errors and hence higher transmission power should be used to maintain the SIR to an acceptable level. One can deine the power eiciency as the transmitted traic per unit power. Then an interesting question arises: which physical mode/power level is the most power eicient? The answer to this will give us the guideline to reduce the power consumption or link adaptation. Although the mode selection itsel is in physical layer, it is closely related to the MAC layer since the MAC layer has great impact on the power eiciency, which will be shown in detail in next section. There is several reported work in literature on link adaptation regarding the mode selection or IEEE 802.11a, e.g., [3, 12, 13]. However, most o them targeted or throughput maximization without the consideration o power. There are several ways to improve power eiciency. In [7], the authors proposed an adaptive link layer strategy which uses hybrid ARQ/FEC to achieve power eiciency. In [8], a distributed mechanism is deined to improve power saving and channel utilization. However, these work only ocused on the link layer optimization without the physical layer consideration. Further more, previous works did not consider the impact o the number o stations on the power eiciency o a WLA. Actually, due to the characteristic o the MAC protocol o IEEE 802.11, the sensing and hearing or other stations will be a critical part in power consumption i the number o station is large. This will aect the PHY mode selection signiicantly. Thereore, in this paper, we study the power eicient transmission by jointly considering the PHY layer and the MAC layer interaction in IEEE 802.11a. With the proposed power consumption model, we will derive the link adaptation or mode selection. The rest o the paper is organized as ollows. Section II briely introduces the IEEE 802.11 PHY and MAC layers. The power consumption model is presented in Section III. The power eiciency is derived in Section IV. Section V presents numerical and analytical results. Finally, the paper concludes with Section VI. II. SYSTEM OVERVIEW A. PHY o IEEE 802.11a The PHY layer o IEEE 802.11a uses OFDM to combat the requency selective ading. The coding scheme in the IEEE 802.11a PHY is rate 1/2 convolutional code. Other coding rates (2/3, 3/4) are obtained by puncturing the mother rate ½ code. The combination o the modulation level and coding rate provides eight PHY modes supporting data rates ranging rom 6Mpbs to 54Mbps. The parameters are listed in Table I. From Table I, we can see that dierent data rates are associated with dierent power level requirements. The minimum sensitivity o 6Mbps mode only requires -82dBm while the mode o 54Mbps requires -65dBm, which is almost 50 times as that o mode 1. That is, the higher the data rate, the larger the receiver sensitivity, and thus the more transmitted power are needed to maintain the receiver power level. On the other hand, dierent PHY mode may encounter dierent transmission ailure probability due to channel error, which is determined by the power level. However, the total power consumption depends not only on the radio transmission, but many other actors, which will be described in Section III. Actually, the transmission power control (TPC) is now also discussed in IEEE 802.11 Task Group h [10, 11]. TABLE I MULTIPLE PHY MODES FOR IEEE 802.11a Mode Data rate Min (dbm) Modulation Code DBPS (Mb/s) rate (R) 1 6 --82 BPSK 1/2 24 2 9 --81 BPSK 3/4 36 3 12 --79 QPSK 1/2 48 4 18 --77 QPSK 3/4 72 5 24 --74 16-QAM 1/2 96 6 36 --70 16-QAM 3/4 144 7 48 --66 64-QAM 2/3 192 8 54 --65 64-QAM 3/4 216 DBPS : number o data bits per OFDM symbol B. DCF o IEEE 802.11 DCF is the undamental access method o IEEE 802.11 0-7803-7802-4/03/$17.00 2003 IEEE 2030
MAC. It is a carrier sense multiple access/collision avoidance (CSMA/CA) with binary exponential backo scheme [1]. The backo counter is a uniormly distributed number within [0, CW]. The CW doubles i transmission ails until it reaches the maximum CW size. A successul transmission is conirmed by acknowledgment packet (ACK). Thus, a transmission sequence comprises o DATA and ACK rame and this is called the basic access method. To reduce the collision cost and combat hidden terminal problem, a new access method, called RTS/CTS, is introduced. Under this access method, the station transmits RTS rame beore DATA rame transmission. The targeted station should response with CTS rame to the source station. Other station shall deer their channel access when they receive the RTS/CTS rame. One o the key characteristics o CSMA/CA scheme is that all the non-transmission stations should hear all the rames although most o them are not the destination stations. The overhead o power or hearing is considerable especially when the number o active stations is large. This is one o the undamental dierences rom the cellular system, e.g., GSM. III. POWER COSUMPTIO MODEL A. Power Consumption Components So ar, most o the work with regard to power optimization only considered the radio transmission power. However, as mentioned by [6], signiicant power may be consumed by other sources: operation o the card, computation, channel sensing and etc. This denotes that even in the idle slot, the card consumed some power. To investigate the overhead o these components on the power consumption, we model them as two parts at certain states: the base power sink, and the incremental part power sink; while the states include idle state, transmitting state and receiving state. Suppose the base power is P base which is the power required or basic operation. This includes the power consumed by the circuits, the power ampliier and etc. The incremental part due to transmission, receiving and channel sensing (idle) is given by P radio, Precv and P idle. Then the total power in each state, PTX, PRX and PI, are respectively given by PTX = Pbase + Pradio PRX = Pbase + Precv. (1) PI = Pbase + Pidle Dierent values o P base place dierent weight on the overhead o power consumption. For simplicity, P base is normalized to 1. The transmission power P radio is scaled based on the power level used to maintain dierent SIR; while the receiving and channel sensing power is invariant. B. State Duration and Energy Consumption The total consumed energy depends on how long the station stays in each state: transmission, receiving, and idle. The transmission can be data, control packet such as RTS/CTS, ACK, and etc. The card operates in receiving state i it does not transmit but other stations do so. The rest time is spent or idle. In IEEE 802.11a, the actual transmission duration o a rame depends on the rame size and the PHY mode used. The transmission duration o a rame with size L in PHY mode m is given by T ( L) = T + T + T ( L), (2) data PREAMBLE SIGAL SYM sym (m) where sym is the number o OFDM symbols needed to transmit L data bits and is given by sym ( L) = (16+ 8 L+ 6)/ DBPS. (3) Here unction x returns the smallest integer value greater (m) than or equal to x. DBPS is the number o data bits per OFDM symbol in PHY mode m and is given in Table I. Other notations are listed in Table II. TABLE II TIMIG RELATED PARAMETERS Parameter Value T SYM : Symbol interval 4us T PREAMBLE : Preamble duration 16us T SIGAL : SIGAL symbol duration 4.0us T slot : slot duration 9us T sis : short inter-rame space 16us T dis : DIFS duration 34us The data transmission in an IEEE 802.11 is completed by a sequence o rame exchanges. In a sequence o rame exchange, a station may be in dierent states. In the basic access method, the station is in transmission state during which it emits radio o data rame. Then, it changes to idle state at the end o the transmission or a time interval o T sis. Ater that, it may receive ACK rame, or still in idle time or T acktimeout, which depends on whether the data transmission succeeds or not. Fig.1 shows the transmission process o a station and (a) is the success case and (b) is the ailure case. For a tagged station, it may encounter ive types o events in next slot and the energy consumption or each event is: The tagged station transmits and succeeds: Etag, suc = PTX T ) + ; (4) PI ( TSIFS + TDIFS ) + PRX T ( Lack )) The tagged station transmits but ails: E tag, ail = PTX T ) + PI ( Tacktimeout + TSIFS ) ; (5) o station transmits: Eidle = PI Tslot ; (6) The tagged station does not transmit, but one o other stations transmits and succeed: E other, tx& suc = PRX [ T( Ldata) + T( Lack)] + PI ( TDIFS + TSIFS) (7) The tagged station does not transmit, and other station(s) transmit(s) but ail: E other, tx& ail = PRX T( Ldata) + PI ( TSIFS + Tacktimeout). (8) Here, the power consumption in each event may depend on the PHY mode and power level. For simplicity, we omit the super-script m which denotes the PHY mode in (4)-(8). 2031
Fig. 1 State in a transmission sequence. Obviously, the average energy consumption or a tagged station depends on the probability o each event, which will be derived in the next section. IV. POWER EFFICIECY MODEL In this section, we assume there are active stations and each with equal average packet size. A. Transmission Probability under Channel Errors The transmission may ail due to two causes: the collision between stations and the channel error. Since the ailure probability depends on the rame exchange sequences (basic access method as DATA+ACK, or RTS/CTS access method as RTS+CTS+DATA+ACK), we irst analyze the basic access method and then point out the dierences or RTS/CTS case. To characterize the DCF behavior with CSMA/CA in IEEE 802.11a, we use the p -persistent CSMA access model or the convenience o mathematical manipulation [4] [5]. Suppose the initial contention window is CW. For a tagged station, the transmission probability τ depends on the ailure probability o a rame exchange sequence, seq, which is given by ( ) 2 τ m ( seq) = (9) s 1 1 + CW + p ( seq) CW (2 p ( seq)) i i = 0 where p ( seq) = p, Lack) is the ailure probability with m denoting the PHY mode. To derive the ailure probability, note that in a WLA with two types o ailure causes (collision and channel error), the probability o a successul transmission is given by P s ( seq) = Ps, Lack ) = Ps ( Lack ) (10) where p ) s ) is the probability o successul transmission o data with length L data in PHY mode m, and p ) s ( L ack ) is the success probability o ACK rame. Since the length o ACK is very short compared with data rame, the success prob. o ACK is approximated to be 1. For the data rame, we have ps ( L) = (1 pc ( L))(1 per ( L)). (11) where p ) c ( L) is the probability that the tagged station collides with another transmission station and p ) er ( L) is the rame error rate (FER) o a data packet with length L. The probability that the tagged station sees a busy channel (collision occurs) is given by ( 1) c ( ) 1 (1 τ ( )) p L = seq. (12) Finally, the ailure probability is given by p ( seq) = 1 p ( seq). (13) By combining (9) ~ (13), we can derive the transmission probability o basic access method. Most o the results mentioned above still hold or RTS/CTS case, except the ollowing minor changes: P ( seq) = P ( L L L L ) s s rts, cts, data, ack = Ps ( Lcts ( Lack ) (14) Ps ) = (1 Pc ))(1 Per )) Here we assume that the CTS and ACK rame are always correctly received. Thus, in this access method, the RTS rame is only corrupted i collision occurs and the data rame is lost only when channel error occurs. B. Power Eiciency Expression We derive the power eiciency or the basic access scheme. Similar results can be obtained or RTS/CTS access method. As mentioned beore, or a tagged station, it encounters ive types o events in next slot. The probabilities are given as ollows, respectively. The tagged station transmits and succeeds: 1 ptag, suc = τ(1 τ) (1 per ) ; The tagged station transmits and ails: ptag, ail = τ ptag, suc ; o station transmits: p (1 ) idle = τ ; The tagged station does not transmit and one o other stations succeeds: 1 pother, tx& suc = ( 1) τ(1 τ) (1 per ); The tagged station does not transmit and other station(s) ail: Pother, tx& ail = (1 τ ) pidle pother, tx& suc 1 2 = (1 τ )[1 (1 τ ) ( 1) τ (1 τ ) (1 per )] Considering the results in Section III, the total energy consumed or the tagged station in the next slot is given by Eslot = ptag _ suc Etag, suc + ptag, ail Etag, ail + pidle Eidle + pother, tx& ail Eother, tx& ail. (15) + pother, tx& suceother, tx& suc The average duration o next (logical) slot, T interval, is the statistical sum o the durations o three events rom the system point o view: the successul transmission, the ailure transmission, and the idle state. That is, T = p T + p T + p δ. (16) interval s _ sys s _ sys idle _ sys where δ is the duration o a real physical slot and T s, T denotes the successul rame exchange duration o success and ailure case, respectively, which are shown in Fig. 1; while the probability o each transmission event can be expressed as: s 2032
1 1 ps _ sys = Cτ(1 τ) (1 per ) p _ sys = 1 ps _ sys pidle _ sys. (17) pidle _ sys = (1 τ ) The successully transmitted data is given by msg = p L. (18) s_ sys Hence, the bandwidth share o a station is 1 msg BW =. (19) T interval Then, the power eiciency can be expressed as the transmitted traic per unit energy. BW Ps _ sys L η = = (20) Eslot / Tint erval Eslot ote that, again, the super-script m or PHY mode is omitted here or the sake o simplicity. V. UMERICAL RESULTS AD AALYSIS In this section, we present numerical results to demonstrate: 1) optimal power eiciency can be obtained via adaptive power level and PHY mode selection based on the proposed model; 2) the characteristic o power eiciency and its dependence on system parameters (e.g., the number o stations, the non-radio-transmission power, etc.). Unless speciied, the parameters are speciied as the IEEE 802.11a standard [2]. The packet length L = 512 bytes. The basic power P base is normalized to 1 and the power in idle state and receiving state are given by P idle = 0 and P recv = 0.1 or all PHY modes. This relects the typical power consumption cases as noted in [6]. A. Power Eiciency Optimization The optimization o η is to select the power level l and PHY mode m to maximize η in (20) subject to some constraints. The constraints are to guarantee the received signal strength above the minimum receiver sensitivity listed in Table I. For example, by assuming noise and intererence power to be around 90 dbm, the receiver sensitivity o 82 dbm corresponds to an SIR=8 db. Obviously, the larger the received SIR is, the larger the transmitted power is proportionally needed. Thereore, in the ollowing, we use the received SIR to denote the transmitted power level or simplicity. ow, the problem is to search the optimal pair (, lm ) that maximize the η (l,m). ote that under a given power level, only part o eight PHY modes can be selected under the minimum receiver sensitivity constraint. For example, when the power level is 8dB, only PHY mode 1 can be used. When the SIR is 13 db, mode 1, 2, 3, and 4 can be used. For the FER o dierent PHY mode, we use the results in [9]. The radio power or SIR= 8dB is set to P radio = 0.5 and the transmission power o higher SIR is scaled accordingly. With dierent number o stations with CW=15, we put several sample results or the optimal (l, m) pair with (20) in Table III. Each η with the given power level in the table is the maximum value among all possible PHY modes. Finally, the best PHY mode can be ound among all the optimal (power, PHY) pairs, which is with the largest η and denoted by * in the table. We can see that power optimization is quite dierent rom throughput maximization. Throughput maximization always uses the highest transmission power to maintain high rate PHY mode while power eiciency needs to jointly consider the throughput and energy needed. TABLE III POWER EFFICIECY COMPARISO Power(dB) PHY η Throughput (Mb/s) 1 8 1 3.0 x 10 6 4.24 1 12 3 4.10 x 10 6 7.78 1 20 5 2.16 x 10 6 12.0 1 10* 3* 4.57 x 10 6 * 7.15 10 8 1 3.22 x 10 5 3.65 10 12 3 5.68 x 10 5 6.83 10 20 5 6.34 x 10 5 11.2 10 18* 7* 9.01 x 10 5 * 17.0 * denotes the optimal power level and PHY mode. B. Power Eiciency Analysis To see clearly the eect o the actors that inluence η, we simpliy the model in Section IV so that it only considers the energy consumption o data rames. Suppose the PSDU is L and or a tagged station, the energy consumption or transmission, receiving slot and idle slot can be approximated as in let side o (21) and the probabilities or transmitting/receiving/idle are given by the right side o (21) under the condition o τ << 1, (1 τ ) 1 τ. Etx = PTX T ( L) ptx = τ 1 Erx = PRX T ( L) prx = (1 τ)[1 (1 τ) ] ( 1) τ (21) Eidle = (1 τ ) PI δ pidle = (1 τ) 1 τ The data transmitted is given by 1 data = τ (1 τ) (1 Per ) L. (22) Then, we have the ollowing equation: data τ [1 ( 1) τ ] (1 Per ) L η =. (23) E τ Etx + ( 1) τ Erx + (1 τ ) Eidle 1) Impact o Power Consumption Model In Fig. 2, the ratio Pradio / P base ranges rom 0.1 to 10 to study its impact on the power eiciency in a WLA with = 10. As illustrated in Fig. 2, when Pradio / Pbase is small, the higher power schemes (22dB, 8) and (26dB, 8) achieve better eiciency than the lower power ones (8dB and 14dB). However, the conclusion reverses as the Pradio / P base ratio increases. Hence, we can see that the adaptive method preers lower power levels when Pradio / Pbase is large and vice versa. 2) Impact o Station umber In Fig. 3, η or several (l, m) modes are plotted vs. dierent station numbers (). From Fig. 3, we can see that η o lower power schemes degrades signiicantly when the becomes large. The reason is that when is large, the energy cost due to 2033
receiving state dominates. With lower power level, and hence lower PHY mode is employed, the energy cost per rame transmission is considerably greater than the higher power ones, which is clearly shown in (23). These results indicate that in WLA system, the power consumption needed or listening is considerable, which is dierent rom the power optimization in the cellular system which only needs to consider the transmitting and receiving one s own data. In Fig. 4, we plot the normalized duration o three states (Tx/Rx/idle) or two schemes: (8dB, PHY=1) and (26dB, PHY=8). We can see rom Fig.4 that the time or receiving increases when is large. Another observation is that the lower mode spends more time on receiving than the higher one. Since all stations except the transmitters need to hear the packet and consume some power, it is natural that the lower modes may exhibit ineiciency at high loads. VI. COCLUSIO In this paper, the power eiciency is investigated or the IEEE 802.11a WLA. By jointly considering the PHY layer and MAC layer, the mode selection or power eicient transmission is derived. The numerical results suggest that the power optimization is heavily dependent on the overhead introduced by non-radio-transmission power sinks. Hence, an adaptation scheme which selects proper mode can enhance the power eiciency considerably. REFERECES [1] IEEE 802.11, Part 11: Wireless LA Medium Access Control (MAC) and Physical Layer (PHY) Speciications, IEEE, Aug. 1999. [2] IEEE 802.11a, Part 11: Wireless LA Medium Access Control (MAC) and Physical Layer (PHY) speciications: high speed Physical Layer in 5 GHz Band, IEEE, Sept. 1999. [3] D. Qiao and S. Choi, Goodput enhancement o IEEE 802.11a wireless LA via link adaptation, in Proc. IEEE ICC2001, May 2001. [4] G. Bianchi, Perormance analysis o the IEEE 802.11 distributed coordination unction, IEEE J. on Selected Areas in Communications, vol. 18, no. 3, pp. 535-547, March 2000. [5] F. Cali, M. Conti and E. Gregori, Dynamic tuning o the IEEE 802.11 protocol to achieve a theoretical throughput limit, IEEE/ACM Transactions on etworking, vol. 8, no. 6, pp. 785-799, Dec 2000. [6] L. Feeney and M. ilsson, Investigating the energy consumption o a wireless network interace in an ad hoc networking environment, in Proc. IEEE Inocom2001, July 2001. [7] Paul Lettieri, Curt Schurgers, Mani B. Srivastava, "Adaptive link layer strategies or energy eicient wireless networking," Wireless etworks, vol.5, no.5, Baltzer, pp.339-355, 1999. [8] L. Bononi, M. Conti and L. Donatiello, A distributed mechanism or power saving in IEEE 802.11 wireless LAs, Mobile etworks and Applications, vol. 6, pp. 211-222, 2001. [9] A. Douexi etc., A comparison o the HIPERLA/2 and IEEE 802.11a wireless LA standards, IEEE Communications Magazine, pp. 172-180, May 2002. [10] IEEE 802.11h/D1.0, Part 11: Wireless LA Medium Access Control (MAC) and Physical Layer (PHY) Speciications: Spectrum and Transmit Power Management extensions in 5 GHz band in Europe, Drat Supplement to IEEE 802.11-1999 Ed., Drat 1.0, Jul. 2001. [11] J. Kim and C. J. Hansen, Power saving and intererence reduction with TPC under DCF, IEEE 802.11-01/519, Sept. 2001, available via http://ieee802.org/11. [12] Z. Lin, G. Malmgren and J. Torsner, System perormance analysis o link adaptation in HiperLA type 2, in Proc. IEEE VTC 2000 Fall, Sept. 2000. [13] M.K. Aziz and etc., Indoor throughput and range improvements using standard compliant AP antenna diversity in IEEE 802.11a and ETSI HIPERLA/2, in Proc. IEEE VTC 2001 Fall, Oct. 2001. Power Eiciency 10 7 10 6 10 5 Optimal Power=8dB, PHY=1 Power=14dB, PHY=3 Power=22dB, PHY=6 Power=26dB, PHY=8 10 4 10 1 10 0 10 1 Pradio/Pbase Fig. 2 η vs. Pradio / P base. The optimal (l, m) selection based on our model is compared with several ixed selection schemes. The optimal (power, PHY) selection sequences are: (22dB, 7),, (8dB, 3). Power Eiciency 10 7 10 6 10 5 Optimal Power=8dB,PHY=1 Power=14dB, PHY=4 Power=22dB, PHY=6 Power=26dB, PHY=8 10 4 0 5 10 15 20 25 30 35 40 45 50 Station umber Fig. 3 η vs.. The optimal (l, m) selection is compared with several ixed schemes, the optimal (l, m) selection sequences are: (10dB, 3),, (22dB, 7). Time Distribution 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 Power=8dB,PHY=1, rx Power=26dB, PHY=8, rx Power=8dB,PHY=1,tx Power=26dB, PHY=8, tx Power=8dB,PHY=1,idle Power=26dB, PHY=8, idle 0 0 2 4 6 8 10 12 14 16 18 20 Station umber Fig. 4 Time distribution o each state (tx/rx/idle) o dierent (power, PHY) pairs with respect to dierent station numbers. 2034