You are on page 1of 11

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. XX, NO.

Y, MONTH 2002

1

Impact of ARQ Protocols on QoS in 3GPP Systems
Carla-Fabiana Chiasserini, IEEE Member, Michela Meo, IEEE Member
Abstract— Traffic transmission over the radio channel requires appropriate techniques to preserve information integrity while maintaining the desired QoS and energy constraints. In this paper, we focus on a possible configuration of the ARQ protocol defined in 3GPP (Third Generation Partnership Project) technical specifications. A Markov model is developed in order to study the protocol behavior as channel characteristics change. Through this model, we evaluate the impact of the 3GPP ARQ scheme on the QoS of traffic services such as data file transfer and IP telephony. In the case of data transfer, a trade-off between loss probability and energy efficiency is derived. In the case of IP telephony, we investigate the possibility to guarantee low maximum delay and low jitter by adopting the ARQ protocol as error-recovery scheme. Keywords— Wireless Quality of Service, ARQ Protocols, Third Generation Wireless Networks, Performance Analysis.

I. I NTRODUCTION Recently, a great effort has been made to provide a common standard for next generation wireless networks. The Third Generation Partnership Project (3GPP) is currently producing technical specifications for a third generation mobile system based on both an evolved GSM network and novel radio access technologies [1]. Besides offering all the anytime, anywhere GSM features, 3GPP specification based systems will offer advanced multimedia applications, mobile Internet services, and data communication with rates as high as 2Mbps. Here, we focus on the information transfer techniques specified in 3GPP at the link layer and evaluate their performance in the presence of different classes of traffic. Several papers dealing with ARQ (Automatic Repeat Request) schemes have appeared in the literature. Performance analysis of ARQ protocols is presented in [2], [3], [4]. In [5], the authors study the trade-off between energy constraints and average traffic delay that can be achieved by dynamically adapting ARQ schemes to channel characteristics and traffic QoS (Quality of Service). Energy efficiency at the data link layer is addressed also in [6], [7], where the behavior of ARQ protocols is studied when transmissions in the presence of disadvantageous channel conditions are avoided. A similar approach is used in [8], where channel conditions and packet deadline constraints determine whether transmissions are deferred, as well as the best error-correcting code to select. In this paper, we consider an ARQ scheme configuration obtained by applying the control mechanisms defined for the acknowledged transfer mode in the 3GPP Radio Link Control (RLC) specification [9]. In addition, we consider that the transmitter is allowed to operate in two different modes: greedy or saving mode. In greedy mode, the transmitter is always active, regardless of the radio channel conditions. On the contrary, in saving mode the transmitter defers the information transfer whenever disadvantageous channel conditions are detected [6],
This work was supported by the Italian Ministry for University and Research through project RAMON and by the Center for Wireless Multimedia Communications (CERCOM), Torino, Italy. C.-F. Chiasserini and M. Meo are with the Dipartimento di Elettronica, Politecnico di Torino, Torino, Italy. Email: chiasserini,michela @polito.it
¡  

[7]. We introduce the saving mode as an enhancement to the 3GPP specification to let the system ‘save’ radio resources as well as energy in the case of bad channel conditions. The impact of the ARQ scheme on the traffic QoS of both loss-sensitive and delay-sensitive services is evaluated by developing an analytical model based on standard Markov techniques. For loss-sensitive services, such as data file transfer, QoS requirements are expressed in terms of throughput and traffic loss probability. QoS of delay-sensitive services is mainly based on parameters such as maximum delay and jitter; conversational and interactive real-time services belong to this class. In order to estimate the impact of the ARQ scheme on the QoS of loss- and delay-sensitive traffic, we identify two case-studies. First, we consider data file transfer over a wireless link and, besides traditional QoS metrics for data traffic, we study the energy efficiency of the ARQ scheme by using a realistic model of the battery behavior. Then, as an example of delay-sensitive traffic, we consider IP telephony. Typically, such a traffic is supported by RTP (Real-time Transport Protocol), which is implemented on top of UDP as a transport protocol. Since UDP and RTP do not have error control functions, efficient mechanisms for loss recovery should be implemented at the link layer [10]. We therefore evaluate the capability of the 3GPP ARQ protocol to satisfy delay-sensitive QoS requirements. The choice of separately analyzing loss-sensitive and delaysensitive traffic derives from the indication that the design of near future multi-service networks will most likely be based on traffic segregation [11]. Since traffic segregation is based on the acknowledgment that different services exhibit highly different sensitivity to the system performance parameters, it achieves efficient system resource utilization while providing QoS guarantees. The remainder of the paper is organized as follows. Section II presents the ARQ protocol configuration; Section III describes the system and the assumptions introduced in order to develop the analytical model. Section IV presents the Markov model of the ARQ scheme. The QoS metrics and the numerical results showing the impact of the ARQ scheme on the system performance are presented in Sections V and VI in the case of data file transfer and IP telephony, respectively. Finally, Section VII concludes the paper. II. T HE ACKNOWLEDGED T RANSFER M ODE
IN

3GPP

The reference network scenario assumed in the 3GPP specification of the RLC (Radio Link Control) protocol [9] consists of a mobile station and an access point to the terrestrial core network, communicating with each other through the radio channel. Here, we focus on the acknowledged information transfer mode [9], which implements a selective ARQ scheme. According to [9], we denote by PDU (Protocol Data Unit) the data unit exchanged between the RLC layer and the MAC layer, and by STATUS PDU the data unit used by the receiver

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. XX, NO. Y, MONTH 2002
TX RX
PDU1 PDU2
Status PDU

2

Fig. 1. Temporal diagram of the ARQ scheme when one or more consecutive PDUs are lost and the sender is in saving mode.

traffic generator

PDU flow

transmission buffer retransmission buffer

While constructing the ARQ model, the following assumptions are introduced: 1. A single buffer models the presence of both the transmission and retransmission buffers. 2. There is not a maximum number of retransmissions per PDU. 3. STATUS PDUs are always correctly received. 4. The inter-polling time is equal to one PDU transmission time. The model, that we develop, focuses on the following aspects of the system: i) the traffic generator, which represents the arrival process of PDUs to the RLC layer; ii) the transmission and retransmission buffers, although the retransmission buffer is only implicitly represented in the model as is described below; iii) the radio channel; iv) the energy consumption and the battery behavior at the mobile station.

P

entity to require the transmission of missing PDUs and to send feedback information to the transmitter. The loss of a PDU is detected by the receiver either because an out-of-sequence data unit is delivered or no data unit is received by a certain time, . At the sender, the acknowledged mode involves the use of transmission and retransmission buffers. Upon the reception of a negative acknowledgment, the sender retransmits only the PDUs that have been indicated as missing by the receiver, giving to retransmissions higher priority than data units transmitted the delay between the end for the first time. We denote by of a PDU transmission and the reception of the STATUS PDU notifying that the PDU was lost, and by the number of PDUs transmitted during time (including the PDUs whose transmission started during ). Moreover, the sender can check on the receiver’s status by polling the receiver until it receives a reply; the receiver must reply to the sender’s poll with a STATUS PDU. In addition to the transfer mechanisms described above, we consider that whenever the sender detects disadvantageous channel conditions, it can enter two different operational modes: greedy or saving mode. In the case of greedy mode, upon the reception of a STATUS PDU notifying that some PDU was lost, the sender retransmits the missing data units at once and then goes on with the information transfer. In the case of saving mode the sender stops transmitting and starts periodically polling the receiver to probe the channel status. The sender starts retransmitting the missing data units only when good channel conditions are detected; i.e., when the sender receives a reply to its poll. The evolution of the protocol in the case of PDU loss and considering the sender in saving mode and , is illustrated in Fig. 1, where dashed lines represent lost data units. In the diagram on the left side, PDU is lost and a STATUS PDU transmission is triggered at the receiver by the expiration of timer . After a time interval equal to , the STATUS PDU arrives, the sender ends the transmission of PDU and stops transmitting. the sender receives a STATUS PDU inBy time interval

III. S YSTEM D ESCRIPTION

AND

A SSUMPTIONS

Let us focus on the information transfer taking place over the uplink radio channel, where, with reference to the ARQ scheme described above, the mobile station acts as sender and the access point acts as receiver. (Notice however that the developed model can represent the reverse situation as well.) We represent the system as sketched in Fig. 2, and we study its behavior by developing a discrete-time Markov chain (DTMC) model, in which the time is slotted according to the PDU transmission time, denoted by .
B greedy Th saving wireless channel

Fig. 2. System under analysis.

G IF

G HF

4 D2 1

#$

STATUS PDU transmission time + 2 (propagation time)

PDU4

$ #$#

PDU transmission time

!"#$

PDU3

! !"

"!"

A 9 B@8

8

E

46 731

46 731

C

46 731

¥¦

PDU4

trx

¡ '(¡'( '(¡ ¡( '(¡(' ¡' '(¡ ¡ %&¡&'( &¡% % ¡& %&¡%&% %&¡ )0¡0%& ¡) )0¡ 0¡0)0 0)¡) )¡ 0¡0 )¡ ¡)0)

¥ ¥¦

¦¥¦

46 731

4 76 1

¨ ¨ ¨ §¡§¡§¡ ¡¡¡§¨

£¤¡£ £¤¡ £¤¡££ £¤ ¢¡£ ¢ ¡  ¢¡     ¡  ¢ ¡ 

ttx

PDU3
Status PDU

PDU3

ttx ttx

Poll Poll

PDU2

ttx

Status PDU

PDU2

trx trx

C

Status PDU

46 731

ttx

ttx

PDU2 

trx

trx

C 



PDU1

E

TX

RX

dicating that PDU arrived successfully; thus, the information transfer is resumed and PDU is retransmitted. In the diagram on the right side, two consecutive PDUs are lost. In this case, the sender receives only the first STATUS PDU indicating that PDU is missing. After time period , not having received any other STATUS PDUs, the sender starts polling the receiver until the poll transmission goes through and the receiver replies with a STATUS PDU. The missing data units are then retransmitted. Saving mode allows for a better usage of the radio resources as well as of the available energy, since the energy cost of polls and STATUS PDUs are negligible compared to the cost of transmitting PDUs. However, saving mode is less reactive than greedy mode to changes in the channel conditions, and therefore it may cause QoS degradation. In order to dynamically trade-off between QoS provisioning and energy saving, we introduce a control mechanism on the sender such that, if the number of PDUs in the transmission buffer is less than a given threshold , the sender operates in saving mode, otherwise it enters greedy mode. Under low load conditions which are not critical for QoS provisioning, we privilege energy saving; on the contrary, when queue fills up and QoS deteriorates, we favor QoS provisioning. By varying the value of , we can easily establish the desired trade-off between energy saving and QoS as radio channel conditions change. 

¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ 

    

¡ ¡ ¡  ¡ ©¡© ¡©© ©©¡ © ¡©

42 531

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. XX, NO. Y, MONTH 2002

3

A. Traffic Generator and Transmission Buffer The traffic generator is modeled as a discrete-time activeidle Markov process with time granularity equal to the PDU transmission time, . A graphical representation of the process is shown in Fig. 3. During active periods, one PDU is generated during a time slot with probability ; whereas, no PDUs are generated during idle periods. Let active and idle denote the average active and idle periods, respectively, and active and idle denote the active and idle periods in slot units: and idle . Then, the parameters active active idle of the traffic generator are

TABLE I C HANNEL BEHAVIOR :
PROBABILITIES AND AVERAGE ERROR PROBABILITY, TRANSITION

AND AVERAGE TRANSMIT POWER .

idle

idle

active

active

α 1−α idle β PDU flow active
probability λ

active

1−β

idle

active

Fig. 3. Traffic generator.

active

(2)

The radio channel is modeled as a Gilbert channel. Two states, good and bad, represent the state of the channel during the transmission time of one PDU: the PDU transmission is successful if the channel is in state good, and unsuccessful otherwise. We denote the transition probability from state good to state bad by and the transition probability from state bad to state good by ; and have been obtained by using the results in [12]. The procedure adopted is as follows. We consider the channel traces derived in [12] for a WCDMA air interface. As described in [12], channel traces are obtained through a UMTS

C. Energy Consumption and Battery Behavior We consider both the energy consumption at the mobile station due to the transmission of data units and the phenomenon of

%

B. Radio Channel

The average error probability, denoted by , and the average length of a burst of errors, denoted by , are derived as and , respectively. The results obtained by using the above channel representation were found in good agreement with simulation results, as shown in [14]. As an example, the average error probability, the values of and , the average length of a burst of errors, and the average transmit power for the uplink channel are reported in Table I, for different mobile stations. These results were obtained by assuming thermal noise equal to -132 dBW, the exponent in the path loss model equal to 3.5, log-normal shadowing with standard deviation equal to 4 dB, Doppler frequency set to 6 Hz, the SINR threshold fixed to 6 dB, and Rayleigh fast fading [12]. 

 ¨  ¢  ©

¥ 

A 2 A 2

Q P # "  0 % G D!A@ C 2 # " ©  ¢ A )(© ¨ 0 H%  68 5 0 %4 8 D6C BA5@ 3A #¢ "  7I(& © % G D)(C !A© @ C 97§4 # " ¢  ¢ A I(© ¨ )(© E% F8 E©)(6 5 E%4 C 9D7BA6C8 §4 3A & H & 5@ 2 ©

4 76 

9

9    

¢A

where the ratio active active is the probability that the idle generator is active. Observe that the presence of a single traffic generator is justified by the assumption that a traffic segregation policy is employed. Thus, different kinds of traffic generators (in particular loss-sensitive and delay-sensitive traffic generators) are separately treated. Instead of describing both the transmission and retransmission buffers, we model a single buffer from which PDUs are removed only after they are successfully delivered to the receiver. This assumption has negligible impact on the performance measures since retransmissions have priority over new transmissions and, therefore, only a few PDUs are waiting in the retransmission buffer at the same time.  

idle

active

# $" 

 )(©)(© '1%0& %

(

Given that the maximum service capacity of the system (with noiseless channel) is one PDU per time slot, we denote by the traffic load normalized to the system capacity, i.e., the average number of PDUs generated per time slot,

(

¦

¥

F

9 

9

and

(1)

0.075287 0.076295 0.122688 0.128128 0.159866 0.175809 0.199287 0.231275 0.265145 0.293261 0.334144 0.372904 0.420459 0.421107 0.455887 0.567886

0.029687 0.022623 0.045430 0.046739 0.056124 0.058466 0.064498 0.077703 0.080407 0.093798 0.099239 0.110713 0.121956 0.122935 0.128343 0.176178

0.364635 0.273903 0.324859 0.318043 0.294946 0.274087 0.259145 0.258275 0.222850 0.226047 0.197755 0.186181 0.168098 0.168998 0.153181 0.134057

2.742469 3.650927 3.078258 3.144229 3.390451 3.648476 3.858843 3.871842 4.487323 4.423859 5.056763 5.371117 5.948912 5.917230 6.528224 7.459514

[W] 0.019412 0.019633 0.022188 0.021043 0.022464 0.021279 0.022618 0.021183 0.022584 0.022314 0.022623 0.023179 0.022993 0.023151 0.023001 0.023240

UTRA-FDD simulator, which accounts for signal propagation, open and closed loop power control, intra- and inter-cell interference, Doppler frequency, and interleaving. The parameters in input to the simulator are set according to [13]. By using the UMTS simulator, the Signal to Interference plus Noise Ratio (SINR) trace is derived for each simulated user; then, by fixing a SINR threshold, the corresponding sequence of PDU error probabilities is computed for each SINR trace. Let denote the number of samples per sequence (i.e., the simulation duration expressed as number of time slots), the probability of receiving the -th PDU correctly, and the probability of receiving the -th PDU in error. For each sequence of PDU error probability, and are derived as, (3) (4)

!      

, AVERAGE LENGTH OF THE

BURST OF ERRORS , 

F

P

 

F

A

P ¢

¡

F

F 9

¤

¨

F

F

¨ F ¢ ©   §9 ¦

F

¡ 

P

F

P

P ¢ 9 



F

A 

¡

¡
9

F 9

£

¡ ¡

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. XX, NO. Y, MONTH 2002

4

2.1

1.4

0.92

0.65

0.43

where

(6) Changes in the state of the traffic generator (see Fig. 3) determine changes in the value of ,

According to our traffic model, when the generator is active, a new PDU arrives at the buffer with probability , if if

Table III reports the transitions from state to its successors , with . In the Table, the third column denotes the conditions that states and have to satisfy in order to . Notice that not all the possible transitions have

 

¥

%9 H  & 9 H 

#

H B©

 

‡ … ˆ†9 H  d

#

©

A 9 B©  © e H  

©

A B9  H ©  # © ©  # ©

e

H B©

where is the state of the traffic generator; can assume two values only which represent the activity and the idle periods of the traffic generator, respectively denoted by and : ; is the transmission buffer occupancy in number of PDUs (remember that includes the PDUs in the retransmission buffer), can assume all integer values between 0 and ; is the state of the transmitter, either active or idle; when active, the transmitter is sending PDUs over the radio channel; on the contrary, the transmitter is idle when it has no PDUs to transmit or when, being in saving mode, it has detected bad channel conditions, ;

¥

%9 &9 %9 & 9

#

#

if if if if

(7)

(8)

¥ 

x % H (f„9  x % H (f„9  x & H (¥„9  x & H (¥„9 

x D F H ƒ¥BG‚9  x D F H ƒ¥BG‚9 y¥CCqw9 x DAA   x DAA  y¥CCqw9

#

#

#

#

¤ 3A 2 ¤ 2£ £ 3A  

3A 2  

stt

ttv

u

ttv

u

9 2 H cV CSf  U R )

9 2 H WU TS@Q   R )

#

#

%

%

In this section, we describe the DTMC model for the case equal to 1; i.e., . Extensions to large values with of will be discussed at the end of this section. In order to model the system behavior, the following dynamics have to be carefully described: i) the PDU generation process, ii) the buffer occupancy, iii) the working mode of the transmitter, and iv) the channel state. Accordingly, let the DTMC state be defined by the vector

if if if if

D F BG9 DAA  CCq€9 D F BG9 H DAA  CCB€9 H

# 

3A 2

stt

IV. A NALYTICAL M ODEL

is the probability that the channel state , with ; is the probability that the traffic generator changes from to , with ; is the probability that a PDU is generated in state ; is an indicator, which is equal to 1 if state is a possible successor of state ; and equal to 0 otherwise. As already mentioned, , and are independent of the state of the system. The value of changes according to the two-state model of the channel, changes from to 

B© 8# © © Xe  H  © H 

(5) 

81

H 

H 3©

2 H iU C(hQ 2 H rU TS@Q  R )   R ) 

H   8H  H   8H 1  H 1  H P H  © I© H 8©# © © % 9  8#   #  8# 1  # 1  #  #  © 9 # ©  HI© 8# © © %

#

d I2 H cb !a`Q Y2 H WV TS8Q 9 B©  © X  U R ) X  U  R ) H 

2 D F DAA ) 5IhECCB@g' H    

2  %) 5& 10' H p  

%

#

#

%

#

#

%

#

H 

H

H  q© 8 © e H  d $ $  2 H iU Ca`f  R ) $  2 H Wb TS8f  U  R )

#

#

%  # © % 

 # ©©

#

#

8 

H ©

%

%

$

d

charge recovery, which takes place in real batteries during idle periods. The energy consumption per PDU transmission is computed as , where is the power consumed by a mobile station in transmitting mode. Clearly, the energy consumption per information data unit transfered over the radio channel increases with the increase of the number of PDU retransmissions. On the contrary, charge recovery may lead to a significant improvement in battery lifetime when a pulsed discharge is implemented instead of a continuous discharge [15], [16]. As shown by several experimental tests [17], [18], [19], [16], the performance gain due to this phenomenon depends on the amount of time spent by the mobile station in idle state. The lower the radio terminal activity rate, the greater the improvement in battery lifetime obtained through pulsed discharge. To account for charge recovery, we define the energy gain as the ratio of the total amount of available energy at the radio terminal when a pulsed discharge is applied to the amount of energy available in the case of continuous discharge. By considering that the transmission of a packet is the operation with the highest energy cost, we take the terminal activity rate to be equal to the number of PDU transmissions per time slot. In Table II, obtained in the case of a lithiumwe report the values of ion battery as the PDU transmission rate of the radio terminal varies; as an example, a value of equal to 1.4 corresponds to a gain of 40% in battery lifetime relatively to the case where a continuous discharge is performed.

6

Transmission rate

0.1

0.2

0.3

0.4

0.5

6 

$

RATE VARIES .

6

$

¤ ¥£

2 & %) ' 310( &

4

" #

%  

   8  !  8 1  1 9

¤ £

421 § 46 53B¨731

¤ ¦£

¡ ¢ 

VALUES OF ENERGY GAIN

AS THE MOBILE STATION TRANSMISSION

6

TABLE II

in slot the value of denotes the state that the transmitter assumed in slot , ; is the channel state, ; in slot the value of denotes the state that the channel assumed in slot , . Observe that, in order to properly take into account the delay between the transmission of a PDU and its acknowledgment, the introduction of some memory about the past history of the system is necessary; therefore, we introduced the variables and . denote the probability that the chain moves in Let one-step from source state to desti. In determining probanation state bilities ’s notice that the behaviors of the channel and of the traffic generator are independent of the rest of the system; thus, their stochastic description can be a-priori specified by input parameters. On the contrary, changes in the transmitter mode and in the buffer occupancy can be deterministically derived from the actual state of the system and from the channel and generator behavior. Hence, we write the following expressions

2 F DAA ) 5DBGECCB@7' 8  A 2  2 D  DAA ) ' 5BFGECCB@78 ¥ 2 % 5& 98)7' 8 A 2 1 81

$

2  %) 5& 10'

©

4 76

¡ ¢ 

% 

1

P

8

4 76

8

% 

1 $ 

$ $

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. XX, NO. Y, MONTH 2002

5

TABLE III

From the steady-state probabilities many interesting performance metrics can be derived, as will be shown in the following sections. V. Q O S P ERFORMANCE
OF

L OSS -S ENSITIVE T RAFFIC

In this section, we employ the proposed model to assess the impact of the ARQ scheme on the QoS of loss-sensitive data traffic services such as data file transfer. In order to model data 

# ©©

G FX

(9)

‡

ˆ

x w9

u `



8

¨

G ¡F

4 #t9 X s

‡

x €9

r

G F

¥

A 9 ©

q

‡

v €9

ˆ 

© ‰ •¢’ 4 ˆ t” 9 “ ‘ˆ …

  3A 2

©

…

r ‡†

8

4

r

q '

‡

v w9

…

© 

©© ‰ ˆ H ©©   H ©  # © % ©

x t9

G FX

4

ƒ   ©„ BrTA ‚ ©

s

8

¨

…

4X y 9©A

are listed in the Table. Transitions corresponding to losses are explicitly indicated. The state variables and are, by definition, determined by the value of and in the previous slot, so that for each transition from state to , we have and . Changes in the value of and depend deterministically on the channel and generator behavior, as well as on the current state. Let us first focus on the saving mode of the transmitter. Since , the transmitter chooses its working mode according to the most recent estimate of the channel state, which is the value of . If during slot , while the transmitter is active and sending a PDU, a STATUS PDU arrives notifying that the PDU sent in slot is missing, the transmitter completes the current PDU transmission, and switches to idle at the beginning of slot . Thus, as reported in Table III, in saving mode from states with and the chain moves to states with . On the contrary, if the transmitter is idle in slot , it sends a short poll message at the beginning of slot , receives a STATUS PDU still in slot and switches to active at the beginning of slot . These are the transitions from states with and , to states with . When in greedy mode, the transmitter is always active. Therefore, states with and are not possible. The buffer dynamics are governed by the PDU generation process and by the successful transmission of PDUs over the radio channel. In particular, during a transition from to a PDU is generated with probability if the traffic generator in the destination state is active (states with ). A PDU is successfully transmitted if the transmitter is active and the channel is good. However, as mentioned above, a successful PDU transmission in slot is notified during slot , and the PDU is removed from the buffer only at the beginning of . Thus, during a transition from to , a PDU is slot removed from the buffer if and . Finally, we adopt the convention that at the beginning of a slot, the generation of a new PDU precedes the removal of a PDU from the buffer. This implies, for example, that the transition from state

to state corresponds to a successful transmission and a PDU loss with probability , while it corresponds to a successful transmission and no PDU loss with probability . By eliminating from the state space the states that are never entered, the total number of states is given by . In most of the numerical results shown in Section V the buffer size is equal to ; in this case and by assuming , the state space comprises 480 states. The model can be extended to take into account values of which are greater than 1. In these cases, the state definition has to include the information about the transmitter and the channel state for the past time the state space cardinality results to be equal slots. With to , which for and is equal to 960. The steady-state probabilities are computed by employing standard techniques. After choosing an ordering of states, denoted by , which maps state into an integer value in the range , the transition probabilities can be collected in the matrix whose size is . The transition probability corresponds in matrix to the element in row and column . Under the same state ordering the vector collects the steady-state probabilities; we denote the elements of by with . Vector is derived by solving the following system of linear equations, 

#   H 1% 9% YA 2   

$ ¡ ¨ ¨ 9¡ " U%Uf R P E ¨ 9¡ h G ¡ $ ¡ ¨ " IG ¡ IgeP D$ ¨ ¡ ¨ $ ¨ ¡ E ¥   h  f R ¥ $ ¨ ¨ ¡ ($ ¦ ¨ ¡ " U%Uf R P E ¨ 9¡ h G ¡ $  ¡ ¦ ¨ " IG ¡ IgeP D$ ¨ ¡ ¨ $ ¨ ¡ E ¥   h  f R ¥  ¡  ¨ $  ¡ ¦ ¨ &¨ 9¡ " d£ ! E 9¡ X ¡ $¨ " d£ 9 E " @ C R `YUa2 ¥ E 9¡ X ¡ P 5G ¡ G $¨ " 5 c£ ! ¥ E " R `YUa2 E ¥ ($ 9¡ X ¡ P G ¡ G $ ¨ ¨ 5 ¥ ¥ ¦ " Xb£ ! E " R `YUa2 E 9¡ ¡ P G ¡ G $¨ 5 " @ C ¥ R P`YUGT2 E ¥ ¨ 9¡ G ¡ $ ¡ ¨ ¥   " RSQG ¡ UT2 ED$ ¨ ¡ ¨ $ ¨ ¡ P G ¥ ¦ " R PQ%UGT2 E $¨1¡A($ 9¡ G ¡  ¨ ¨ ¥ ¦ " SPQG ¡ IH@ EF¡ ¨ $ ¨ ¡ ¨ $ ¨ ¡ R G    ¥ $ ¨ ¡ ($ ¦ ¨ ¡ ¨ ¦ $ ¡ ¨ ¨ 9¡

4  H 

©

   CCqpC#  9% 1%a48#  © DAA     

¦ ¦ ¦ ¦ ¦

¦ 8

¦ 8

¦ 8

¦ 8

¦ 8

¦ 8

¦ 8

¦ 8

¦ 8

¦ 8

¦ 8

¦ 8

8

8

8

8

8

" ()¨ ¨ ¨ @ C )¨ ¡ ¨ $ h " A)0¨ 6 ¨ ©¨ ¡  ¨ $6  ¨ ¡ " ¡ ¨ $ ¨ ¨ ¨ 6 @ D¡ ¨ $ C6    " ¡ )06 ¨ 6 ¨ a)($ ¨ $ ¨ h¨ " (¨ ¨ 6 ¨ @ 6 V ¨ ¡ ¨ $   " ¡ ¨ $ )¨ 6 ¨ ¨ ¡ ¨ $  6  " )0)¨ 6 ¨ 6 ¨ R ¨ ¡ ¨ $  P " 0)(¨ 6 ¨ R ¨ X  ¨ $  ¨ 36 P " 5 &0()¨ @ 6 V ¨ X¨ $¨ 3¨ 3 ¡ 5 " )&0(©($ X¨ $¨ 3¨ 3¨ ¡¨ " V  ¨ $¨ 3 (5 )0¨ ¨ @ W¡ ¨ 5 !4!4 "©¦ ¨ $  ¨63 ¨ ¡  ¨ 00¨ ©($ 5¨ B4 ¦ ¨ 3 " ¡!(¢)¨ ()¨ @ 6 V ¨  4 $ ¡ " ¡ )¨ $ 6 ¨ (¨ ¡ ¨ $   ¨ 3  " ¡ )¨ $ ¨ ¨ ¨ 6 @ W¡ ¨ V   " ¡0)¨ 6 ¨ ©($  ¨ $6  ¨ ¡¨ " 0)©¨ 6 ©¨ ¡  ¨ $  ¨ ¡6  ¨ ¡  " ¡ ¨ $ )¨ ¡ ¨ ¨ @ D¡ ¨ $ C6    " ¡ ()00¢($  ¨ $6 ¨ 3 ¨ 3 ¨ 2 ¨ " A)()0¨ ¨ ¨ ¡ ¨ $¨ 3 @ " &0()¢¨ ¡  ¨ $  ¨ 3 ¨ 36 ¨ 2 $  

#

H ©

9

% 9 H 

H B©

6

8H 

DCACB 9  A H © ©

#

1

# # 8

6

  9  H ©

D F BG

9

DAA CCB 9

67¥ 67¥ ¦ §¥ ¦ §¥ 67¥ ¦ §¥ 67¥ 67¥ 67¥ ¦ §¥ 67¥ ¦ §¥ 67¥ ¦ §¥ 67¥ ¦ §¥ 67¥ ¦ §¥ ¦ §¥ 67¥ 6 7¥

8H 1

9

" ¨ ¡  1)¨ 5 4 4 !!©¦ " ¨ ¡ )¨  5 " 4 4   !B¢¦ )¨ ¡ )¨ X 5 " )¨ ¡ )¨ X  " ¨ )¨ ¨ ¨ X  ¡ 5  ¡  "  ¡  X   ¡  ¨ 36 #5 !¨ ¨ 0¨ 5

‡

8# 

" ¡  X  ¡  ¨ 3 ¡  ¨ D!¨ ¨ 0)¨ )1¡ 5 "¤!)¨ ¡  ¨ 0)¨ )1¡ ¡  ¨ 3 ¡  ¨ 5 4 4 !!©¦ " )¨ ¡ )¨ ¨ ¨ ¡ )¨ ¡ X   5 6 6 " ¨ (¨ ¨ )1¡ ¡ ¨ 3 ¡¨ 5!!©¦ 4 4 6 " ¨ ¡ ¨ ¨ ¡ ¨ ¡ )¨ ¡    5" 4 B!4©¦ ¨ ¡ ¨ 6 0)¨ ¨ ¡  ¨ 3 @ 5 4 !B4¢¦ 6 ¨ (0)¢)1¡ ¡ ¨ 3¨ 3¨ 2¨ ¡ ¢ 

A

d

% 9

# ¨

i p

"

5 4 4 !!©¦

State

Successor

Condition

(possible loss)

(possible loss)

" !)0)¨ )('&&$ $  ¨ $  $  ¨ $  ¨ $  ¨
saving mode greedy mode

£ $ %© 

" ¡  ¡  ¡  ¨ ¡  ¨ ¡  ¨ #!¨ ¨ ©©©¡

£ ¡ ¤¢ 

P OSSIBLE ONE - STEP STATE CHANGES FROM STATE

TO STATE

¦ §¥

¦ §¥

.

A

8

6 

6

8# 1

2

H ©

%

& 9

6

& ‚9

¨ T1¡ ¨ h¨ 6 6 ¨ ¨ ¡ )¨ ¡  6 ¨ 6 T¨ ¡ ¨ h 6 6 ¨ ¨ ¡ )¨ ¡  6 @ C 6 R 1¡ P¨ @ C R 1¡ P¨

1 # ©  81

9

1

#1

&

#1

A

2

9

6

¦ §¥ ¦§¥ ¦ §¥ ¦ §¥ ¦ §¥ ¦ §¥ ¦ §¥ ¦§¥ ¦§¥ ¦ §¥ ¦§¥ ¦ §¥ ¦ §¥

1

H1 #© ¨ A

6

8
A

6

A

¨

% 9 H 1

6

6

9

8

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. XX, NO. Y, MONTH 2002

6

(11) where is the set of transitions from to , which cause losses (see Table III). Similarly, we compute the average number of total PDU transmissions (unsuccessful included) per time slot as the sum of the probabilities of the states in which a PDU transmission is attempted, (12) with and . In a time interval of slots, the number of transmission atand the number of successful transtempts is equal to mission is equal to . Therefore, we compute the average number of transmissions (retransmissions included) per PDU as,

At the mobile station, the energy consumption per data unit correctly transmitted over the radio channel, is

A. Performance Metrics for Loss-Sensitive Traffic In the case of data traffic, performance of the ARQ scheme can be expressed in terms of throughput, average PDU delay, PDU loss probability, and energy consumption. Throughput is defined as the average number of information bits that are successfully transmitted in the time unit. The sum of the probabilities of the states in which a successful PDU transmission occurs provides the average number of PDU successfully transmitted in a time slot. By properly normalizing this sum, we obtain the throughput as,

with being the power consumption in transmitting mode. As mentioned before, the energy gain due to charge recovery effect [15], [16] depends on the average time between two subsequent PDU transmissions. Table II reports the values of the energy gain as the average number of PDU transmissions per time slot varies. We compute the average data unit delay, , from the average buffer occupancy by applying Little’s formula,

(10)
i b

where the throughput is computed as in (10). The factor makes the average data unit delay be expressed as millisecond. B. Numerical Results Access to Internet-based services via radio mobile terminals has recently become a reality. We assume that the average size of the file transferred over the radio channel is equal to 12 KB, the peak rate is equal to 384 Kbps, and is equal to 0.25. The values of the parameters which characterize the channel behavior and the values of power consumption at the mobile station in transmitting mode are reported in Table I. We first validate our model by comparing analytical results to simulation results. The main objective of the validation is to assess the impact of the fixed-point procedure, that is used for the

By considering that a PDU transmission time is equal to 10 ms and the peak data rate is equal to 384 Kbps, we have bytes.

¦

with and ; the PDU size is equal to , is the slot duration. By expressing as bits and as milliseconds, is the throughput in Kbps. The PDU loss probability is the probability that a PDU as the can not be accommodated in the buffer. We compute fraction of PDUs which are lost. The sum of the probabilities
£ D G 2 `H@ 2 C(F 

# © ‰  B© C# © © % H 
©

(13)

(14)

(15)

¦

¨ §% 

©©

¦ 06 X
H B©
‰ 1X ƒ

‡

F‚  

c da ¦ 9 ! A

” “ ’ •¢t

i g

£

¥

F

6 Y W TP Si  9 F  6 pURXV 9 h

4 

©©

q 4 h P 76 r9 

¥ £   ¦¤¡ t9 ¦¤¡ sE‘" ‚ ¥ £   Xƒ

¦ y6 X  # ©© ‰  H © # ©© %
2
‰ 

8  C 8 1 96 % P  ©   

f”“ X•¢’ 9 F

!

ƒ d‚

&

£

"

X 06 F f6 X

6

£

ca db4

£

9 © ) 9 %q g

6

9

6

ƒ

Y UTR W Q `SXV 9 ¨ P

¨e

4 76 q

£

¥ £¡ ¦¤§ 

%

traffic, we let the active period of the traffic generator represent the time that the generator needs to successfully transfer a file and we set ; idle periods represent time intervals between subsequent file transfer requests. In Section III, we described the dynamic of the traffic generator as independent of the system performance. However, an accurate description of the system for data services has to take into account the implicit feedback that the system performance has on the traffic generator. In particular, the traffic generator is going to be active until all the PDUs corresponding to the current file are successfully transferred; i.e., if system performance is low and many PDU losses occur, the generator is going to be active for a longer time period. We take this phenomenon into account by neglecting losses due to the wired network and assuming that all PDUs, that are lost due to buffer overflow, have to be regenerated. Let be 1 the length of the file to be transferred and the PDU size; is the number of PDUs corresponding to the file and active is the duration of the active period in time slots ( active is also equal to the actual number of generated PDUs). Let also be the probability that a PDU is lost, due to buffer overflow. At steady-state we expect: active . is a-priori specified, both active and deWhile pend on the system performance. We derive them by means of a fixed-point procedure. We initially set active and . We then solve the model and derive new estimates of and active , say and active . This procedure is repeated deriving and active at the -th step. When convergence on the parameter estimates is reached, the steady-state behavior is analyzed. Furthermore, in order to have an even more realistic description of a data traffic generator, in loss states we forbid that the generator switches to the idle state. In these cases, in fact, the generator has at least one data unit to regenerate.

of the transitions in which a PDU is lost represents the average number of PDU which are lost during a time slot. In a time interval of slots, the average number of lost PDU is equal to this sum multiplied by . In the same time interval, the average number of PDUs arriving at the buffer equals , where is the normalized traffic load as defined in (2). Thus, is given by,

 

¨ %

  ¥ £ ¦¤¡   9

¡

¨ §%

¥ £¡   ¡ ¨ ¦¤§  9  ©%

DB@ 8 642 ECA9753£ 1#¦' 0)(

‡

¢

¢

¥ £¡ ¦¤¢  

©©

2 

¥  8  ECCB  8 1 1P %¦¤©£ ¡  DAA    

¡

‰

2 3A ©

"”“ #•¢’ P 9 ¥ £  ! ¦¤¡   

G  

¨

%

¡

¡

A 

¨  %  G 

9

P

¡

 

¥ £¡ ¦¤§   

 

9 © ) 9 %q $ 



¢

¨

%

¡‡

¥ £¡ ¦¤§   ¡ ¥ £ ¦¤¡  
9 ¨

2 4 9I6

¢

&

¨ ©% ¨ ©%

%

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. XX, NO. Y, MONTH 2002

7

1e-02

0.38 0.36 ε=0.42 ε=0.37 ε=0.33 ε=0.29 ε=0.26

Energy consumption [mJ]

Loss probability,PL

0.34 0.32 0.3 0.28 0.26 0.24

Th = 0 Th = B

1e-03

analysis, m=1 simulation, m=1 simulation, m=2 simulation, m=3 1e-04 0 5 10 15 20 25 30 35 Threshold, Th 40 45 50

0.22 0.001 0.01 Loss probability, PL

1e-01

Loss probability,PL

1e-02
Energy consumption [mJ]

1e-03

1e-04

Th=0 Th=20 Th=40 Th=50 0.1 0.15 0.2 0.25 0.3 0.35 0.4 Error probability, ε

1e-05 0.05

Fig. 5. Data traffic loss probability as a function of the average error probability over the wireless channel and for different values of threshold .

‡

¨

v

%

9 @4

A 9 8

G ¡F

¨ A%

‡

v

G F

9 HF G

‡ 9 HF G

G ¡F

G F

evaluation of the steady-state behavior of the system in the case of data traffic. A synchronous simulator was developed for the analytical model validation. The simulator accurately describes the proposed system configuration. It comprises a detailed description of the transmission and retransmission buffers and of the regeneration mechanism of lost PDUs; while it shares with the analytical approach all the other assumptions listed in Section III. Analysis and simulation results are compared in Fig. 4. PDU loss probability, , is plotted versus threshold , with buffer capacity PDUs, and average error probability over the wireless channel . Analytical results derived for are quite accurate when compared to simulation predictions. In the absence of the fixed-point approach, as it is the case for delay-sensitive traffic, analysis and simulation plots are indistinguishable (comparisons will not be shown here for the sake of brevity). As already mentioned, due to the exponential growth of the state space, the model becomes exceedingly complex for . However, from simulation results with it can be seen that does not have a relevant impact on system performance.

A 9 8

x ‡i

8

R P

R P

Fig. 4. Data traffic loss probability as a function of threshold and simulation results are compared.

. Analytical

Fig. 6. Trade-off between energy consumption at the mobile station and loss probability for data traffic, as the average error probability over the wireless channel varies.

0.38 0.36 0.34 0.32 0.3 0.28 0.26 0.24

Th = 0

ε=0.42 ε=0.37 ε=0.33 ε=0.29 ε=0.26

Th = B

0.22 100 120 140 160 180 200 220 240 260 280 300 Average delay E[T], [ms]
Fig. 7. Trade-off between energy consumption at the mobile station and average delay for data traffic as the average error probability over the wireless channel varies.

only. Also, because of Thus, we show results for the case space limitation and since loss probability is typically one of the most critical performance metrics, we do not validate the model for other performance measures. When designing a system with the objective of providing QoS guarantees for data traffic, we usually aim at satisfying constraints on loss probability. Fig. 5 shows the impact of the average error probability over the wireless channel on the loss proband . Notice that the ability for different values of values and correspond to the case where the transmitter operates in greedy mode only and in saving mode only, respectively. As expected, the error probability over the radio channel has a remarkable impact on . However, from Fig. 5 we observe that, under the considered traffic scenario, may turn into significant changes in changes of the value of the loss probability. Given the channel conditions, we can reduce the loss probability by almost one order of magnitude by decreasing from 50 to 0. The selection of small values of providing small values of , however, may not always be a

G ¡F

  ¡9

8

x

¥

‡ 9 ‡

¨

P

%

v €9

4

8

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. XX, NO. Y, MONTH 2002

8

(17)

Then, we have,

Combining (19) with (16), we write
” “ ‰ •(’

where the denominator makes the probabilities be conditioned to the entrance of a PDU in the buffer. By substituting (21) in

© 

© # © %  # © ‰ 9  © ©  £  © © ©  £ © ‰ a 4

#© 

e

‰

©

Some of the most significant metrics for delay-sensitive traffic are related to the distribution of the delay experienced by PDUs. QoS constraints for delay-sensitive traffic are typically expressed in terms of the jitter , which is the delay variance normalized to the mean delay, and in terms of the probability that a PDU reaches the receiver after a maximum admissible delay, . By exploring the possible paths of the Markov chain, we compute the delay distribution. In order to simplify the analysis, we

‰



A. Performance Metrics for Delay-Sensitive traffic

(20) where collects the probabilities ordered according to . Finally, we need to derive the probabilities . Again, we proceed by observing the Markov chain evolution. Consider the set of transitions from state to state , which correspond to a PDU entering the buffer and making the state of the system be . The probability is equal to, 

©©  £ 

g 

G D8  

£ 9 g    !

¥

¥

if otherwise
 t9

g  D8   © © 9 2 © ¨©9 F ) G   §¥ 

©©  £ 

©© 

G H8    © ©   © ©  £ 

‰

2

(

9

We define the row vector

as a state indicator,

(18)

(19)

(21)

¥ 

if otherwise
 „9 

©©  2 H iUV R   R # # ! aH ` ) 5%&X5&2  H ‡WH b©U # Y¥q)H © %     9

9

…

DCACAqp 1% E& YA    9    2 H cU Ca`)  R  

#

 

#

@Q)

r  X X X  x  !G``@q!TA 

# 

©© 

% ¦©X© 2 H r( U # IR ) %   

‡ A … 9

DAA   % &  CCqp8  15TA 8  ‡ 9 G F

2@

9  9 ) B 

@Q

©

(

!

%

9 2 9 F ) ¥

#

(

‡ 

£ !

s

v

u 

©©

9

@

%

#© 

QoS requirements that have to be fulfilled for delay-sensitive traffic can be expressed for the basic information unit of the RLC layer in terms of average and maximum delay, jitter, and loss probability [20]. We study the performance of IP telephony over a wireless link, when UDP is adopted at the transport protocol and the ARQ scheme is used at the RLC layer. The voice source is modeled as the traffic generator model proposed in Section III-A, with the alternate active and idle periods representing the talkspurts and the silence gaps, respectively [21]. The mean active time is taken to be 1 s and the mean idle time is assumed to be 1.35 s [22]; in the active state, in each time slot the voice source . The voice data generates one packet with probability rate is equal to 12.2 Kbps [23]. Notice that in this case we do not introduce the fixed-point procedure described for data traffic, since voice traffic is delay-sensitive and we do not expect that PDU losses influence the PDU generation process. As widely accepted, we consider that a good quality voice traffic is obtained when the number of dropped voice packets is less than 1% of the total number of generated packets, and that the voice quality is still acceptable if the packet dropping probability is less than 3% of the total number of generated packets. We take 150 ms and 250 ms as constraints on the maximum delay to obtain an optimal and a good voice quality, respectively; for the delay jitter, the requirements for optimal and good voice quality are 75 ms and 200 ms, respectively [24].
v

‡ 9

 

VI. Q O S P ERFORMANCE

OF

D ELAY-S ENSITIVE T RAFFIC

In order to compute , we focus on the behavior of the system after the arrival of a PDU. Since is the state entered by the system immediately after the PDU arrival, we need to evaluate the probability of all the paths which make the Markov chain move in slots from state to the states corresponding to the successful deliver of the PDU. By assuming that after the considered PDU arrival no other PDUs are allowed in the system, is the probability that the system starts in and empties in slots. The Markov chain which describes this transient behavior of the system can be derived from the previously described Markov chain by simply imposing that , i.e., by forbidding the arrival of new PDUs. Let be the transition matrix derived by substituting 0 to in the matrix obtained for the case . The buffer empties when a transition from state to state occurs; this happens with probability . as in (9), we collect in By ordering states according to the row vector the probability that the buffer empties from state , with ,

© 

©©  £ 

‰

¥  ©©  £ 

©

‰ ©2

‡ 9 G F 4 9 IF G

§ ¥ © ¢9 F )

¥ 2 © §¨¦9 F ) ¥ ‡ 9 ¡F G ©

2 © ¢9 F ) § ¥

%

” “ •¢’

%

!

9 2 ©9 F ) ¥

%

¥ 2 © ¨9 F ) § ¥

¥

% %

©

good designing choice since it results in large energy consumption. Figures 6 and 7 show how energy consumption can be tradedoff with QoS performance by varying . Figures 6 and 7 present the energy consumption per PDU, as expressed in (14), as a function of the loss probability and of the average PDU delay, respectively. In the plots, different curves correspond to different values of the average error probability over the wireless channel; different points on each curve are obtained by fixing the value of and varying . The results can therefore be used to find the optimal trade-off between energy consumption and QoS as the channel conditions change. For example, looking at we can obtain a decrease of 18% in enFig. 6, for ergy consumption at the expense of the increase of one order of magnitude in loss probability. For , an increase of the loss probability as small as 4.5% allows for 32% of reduction in energy consumption. Likewise, Fig. 7 shows that a significant reduction in energy consumption can be achieved provided that a small additional delay in the data transfer is accepted.
x u

focus on the two boundary cases: , i.e., the system is working in saving mode only, and , i.e., the system is working in greedy mode only. We first consider the case . Denote by the probability that a PDU arriving at the buffer makes the system enter state and denote by the probability that, in this case, the delay is equal to slots. We can write, (16)

¥‡ 9  

G F

¥‡ 9 P

 

G ¡F

y &x

¥‡

9

P

P

4

£ ¡ ¤¢F

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. XX, NO. Y, MONTH 2002

9

Discard probability, PD

(20) for all the values of , we obtain the delay distribution. This procedure can be repeated for the case where the transmitter operates in saving mode only, by properly setting matrix , i.e., by substituting 0 to in the matrix obtained for . From the delay distribution we can derive the jitter, , defined as the variance of delay divided by the mean delay,

We can also compute the probability that a PDU is discarded , at the receiver because its delay exceeds a target value

, is the probability that a PDU is The total loss probability, lost either because of buffer overflow or because it is discarded at the receiver for exceeding delay,

B. Numerical Results

We consider IP voice traffic as a case study for delay-sensitive traffic. The values of the parameters characterizing the channel behavior and the values of power consumption at the mobile station in transmitting mode are reported in Table I. Fig. 8 shows the discard probability computed as in (23) for different values of the target delay for greedy mode only (solid lines) and for saving mode only (dashed lines) when two different cases are considered for the error probability over the radio channel, and ; the buffer capacity is set to 20. The curves of greedy mode only and saving mode only, and , respectively, are which correspond to the bounds of the discard probability. For values of between 0 and the discard probability lies inside the region defined by the bounds. In the considered scenario, the loss probability due to buffer overflow is negligible; packets are lost mainly because they are discarded at the receiver due to exceeding delay. Consider the case with . If the maximum delay constraint is larger than 175 ms, the service can be provided with acceptable quality (i.e., packet losses smaller than 3%) by using saving mode, and with good quality (i.e., packet losses smaller than 1%) by using greedy mode. Fig. 9 shows the discard probability versus the average error probability over the radio channel, when the target delay is fixed at 200 ms. Again, the overflow probability is much smaller than the discard probability which almost coincides with the total loss probability . As the channel conditions worsen, the discard probability increases significantly. For values of the average error probability greater than 0.25, greedy mode is preferred to saving mode in spite of its larger energy consumption; in fact, the delay perceived by using saving mode is too large for providing the service with acceptable quality. When the average error probability over the radio channel is larger than 0.32, the service can be guaranteed with acceptable quality only if the QoS constraint on the maximum delay is relaxed. The impact of the buffer size on the total loss probability can be observed in Fig. 10 for average error probability equal to 0.17

Discard probability, PD

Total loss probability, PT

4

£ ¤¡ F

  4 G ‚9 IF  G F

¥ 2 ¨ 2 2    % 3A ©  % 3A © 3A

¥

2 ¥ 9 F )

4

¥

C ƒ F‚ £

9 HF G

  ¡%

ƒ

…

y

g %  % G A 3A 2

©

2 ģ C

§ ¥ £! ¨¦¤g

‡ ‡

  ¥‡ 9 P

¥

x

‡

9

9 ¡F G

£

P

¥

1e+00

ε=0.23, greedy ε=0.23, saving ε=0.076, greedy ε=0.076, saving

9   9 ¢   9 % g

 

1e-01

(22)

1e-02

1e-03

  ¥‡ 9 P

%
x

(23)
50 100 150 200 250 300 350 400 450 500 Target delay [ms]
Fig. 8. Discard probability for IP telephony as a function of the target maximum delay, in the case of greedy mode only and saving mode only.

(24)

1e+00 greedy saving

%

1e-01

1e-02

1e-03 0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

Error probability, ε
Fig. 9. Discard probability for IP telephony as the average error probability over the wireless channel varies, in the case of greedy mode only and saving mode only.

g

%

4

1e-01

greedy saving ε=0.17 ε=0.23

1e-02

1e-03 10 11 12 13 14 15 16 17 18 Buffer capacity
Fig. 10. Total loss probability for IP telephony versus buffer capacity for different values of average error probability over the wireless channel, in the case of greedy mode only and saving mode only.

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. XX, NO. Y, MONTH 2002

10

0.42 0.4

Energy consumption [mJ]

0.36 0.34 0.32 0.3 0.28 0.26 0.24 0.22 50 60 70

VII. C ONCLUSIONS In this paper, we considered the ARQ protocol configuration defined in 3GPP specifications, and we introduced a control mechanism of the transmitter activity that allows us to establish an optimal trade-off between energy saving and quality of service provisioning. According to the proposed mechanism, the transmitter enters greedy or saving mode depending on a threshold value on the buffer occupancy. A Markov model was developed to study the system performance as the channel conditions vary and in the presence of different quality of service requirements. Results derived in the case of both data file transfer and IP telephony showed the trade-off between energy consumption and quality of service performance, that can be achieved by changing the buffer threshold. Also, in the case of IP telephony, we found that the proposed ARQ configuration is able to guarantee the required values of maximum delay and jitter by choosing an appropriate buffer threshold. By plotting the total loss probability versus the buffer capacity, we were able to determine the optimal value of buffer capacity, as the channel conditions and the operational mode of the transmitter change. ACKNOWLEDGMENTS
0.2 0.25 0.3 0.35 0.4 Error probability, ε

Th = B
80 90 100 Jitter, J [ms] 110 120 130

Fig. 11. Trade-off between energy consumption and jitter for IP telephony, for different values of average error probability over the wireless channel.

1.4 1.3 1.2

Energy gain

1.1 1 0.9 0.8 0.7 0.6 0.05 Th=0 Th=5 Th=10 Th=15 Th=20 0.1 0.15

The authors would like to thank Prof. Michele Zorzi and Dr. Michele Rossi for providing the results reported in Table I. R EFERENCES
Third Generation Partnership Project (3GPP), “Technical specifications,” March 2000, http://www.3gpp.org. [2] R. Cam and C. Leung, “Throughput analysis of some ARQ protocols in the presence of feedback errors,” IEEE Trans. on Communications, vol. 45, no. 1, pp. 35–44, January 1997. [3] H. Lee and B. Ngo, “Queuing analysis of the selective repeat ARQ scheme,” IEE Proceedings, vol. 138, no. 6, pp. 487–93, December 1991. [4] M. Chiani, “Throughput evaluation for ARQ protocols in finite-interleaved slow-frequency hopping mobile radio systems,” IEEE Trans. on Vehicular Technology, vol. 49, no. 2, pp. 576–81, March 2000. [5] J. Redi, C. Petrioli, and I. Chlamtac, “An asymmetric, dynamic, energyconserving ARQ protocol,” in Proc. VTC’99, Houston, TX, May 1999. [6] M. Zorzi and R.R. Rao, “Error control and energy consumption in communications for nomadic computing,” IEEE Transactions on Computers, vol. 46, no. 3, pp. 279–89, March 1997. [7] B.R. Badrinath and P. Sudame, “To send or not to send: Implementing deferred transmissions in a mobile host,” in Proc. the 16th ICDCS, 1996, pp. 327–33. [8] M. Elaoud and P. Ramanathan, “Adaptive use of error-correcting codes for real-time communication in wireless networks,” in Proc. INFOCOM’98, 1998, pp. 548–55. [9] Third Generation Partnership Project (3GPP), “Radio link control (RLC) protocol specification,” March 2000, http://www.3gpp.org. [10] G. Carle and E.W. Biersack, “Survey of error recovery techniques for IPbased audio-visual multicast applications,” IEEE Network, vol. 11, no. 6, pp. 24–36, November-December 1997. [11] P. Bender, P. Black, M. Grob, R. Padovani, N. Sindhushayana, and A. Viterbi, “CDMA/HDR: A bandwidth-efficient high-speed wireless data service for nomadic users,” IEEE Communications Magazine, vol. 38, no. 7, pp. 70–77, July 2000. [1]

Fig. 12. Energy gain versus average error probability over the wireless channel for IP telephony, for varying values of threshold .

and 0.23. The maximum acceptable delay is equal to 200 ms. Observe that the loss probability is dominated by the discard probability: increasing the buffer capacity beyond 14 does not have any effect on the loss probability. When is equal to 0.17, a good quality IP traffic can be provided in greedy as well as in saving mode. On the contrary, when is equal to 0.23, saving mode can only guarantee an acceptable quality of service. Using greedy mode involves higher energy consumption than using saving mode. Fig. 11 shows the trade-off between energy consumption and jitter, that is obtained by varying the value of threshold . Different curves refer to different values of average error probability over the radio channel. For example, when , a 25% reduction in energy consumption can be achieved at the expense of a 17% increase in the jitter; when , a 17% reduction in energy consumption involves a 40% increase in the jitter. In Fig. 12, the energy gain is plotted as the average error probability over the radio channel increases and as the buffer threshold varies. The step behavior of some curves is due to the discretization of the values of the transmission rate and of the cor-

G ¡F

0.38

Th = 0

P

ε=0.42 ε=0.37 ε=0.33 ε=0.29 ε=0.26

responding energy gain (see Table II). Since the transmission rate in greedy mode is larger than that in saving mode, the energy gain increases as the time spent by the transmitter in saving mode increases, i.e., as threshold grows. We observe that, for large values of , the energy gain obtained when the transmitter operates in saving mode only is twice the energy gain achieved in greedy mode only.

P

R P

P

   ¥ ‡ 9 P
y •x

G F

¥‡

9

P

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. XX, NO. Y, MONTH 2002

11

[12] A. Giovanardi, G. Mazzini, V. Tralli, and M. Zorzi, “Some results on power control in wideband CDMA cellular networks,” in Proc. of the IEEE WCNC 2000, Chicago, IL, USA, September 2000. [13] Third Generation Partnership Project (3GPP), “UE radio transmission and reception (FDD),” September 2001, http://www.3gpp.org. [14] M. Zorzi and M. Rossi, “RAMON Project,” November 2001, Technical Report. [15] C.F. Chiasserini and R.R. Rao, “Stochastic battery discharge in portable communication devices,” in Proc. of 15th Annual Battery Conference, Long Beach, CA, January 2000. [16] Sung Park, A. Savvides, and M.B. Srivastava, “Battery capacity measurement and analysis using lithium coin cell battery,” in Proc. of the International Symposium on Low Power Electronics and Design, 2001, pp. 382–387. [17] T.F. Fuller, M. Doyle, and J.S. Newman, “Relaxation phenomena in lithium-ion-insertion cells,” J. Electrochem. Soc., vol. 141, no. 4, pp. 982– 990, April 1994. [18] R.M. LaFollette, “Design and performance of high specific power, pulsed discharge, bipolar lead acid batteries,” in Proc. of the 10th Annual Battery Conference on Applications and Advances, Long Beach, January 1995, pp. 43–47. [19] B. Nelson, R. Rinehart, and S. Varley, “Ultrafast pulse discharge and recharge capabilities of thin-metal film battery technology,” in Proc. of the 11th IEEE International Pulsed Power Conference, Baltimore, June 1997, pp. 636–641. [20] J.-F. Huard, A.A. Lazar, K.-S. Lim, and G.S. Tselikis, “Realizing the MPEG-4 multimedia delivery framework,” IEEE Network, vol. 12, no. 4, pp. 35–45, July-August 1998. [21] P.T. Brady, “A model for generating on-off speech patterns in two-way conversations,” Bell Systems Technical Journal, vol. 48, pp. 2445–2472, September 1969. [22] R. Fantacci and S. Nannicini, “Multiple access protocol for integration of variable bit rate multimedia traffic in UMTS-IMT-2000 based on wideband CDMA,” IEEE Journal on Selected Areas on Communications, vol. 18, no. 8, pp. 1441–1454, August 2000. [23] Third Generation Partnership Project (3GPP), “Common test environments for user equipment (UE). conformance testing,” September 2001, http://www.3gpp.org. [24] P. Fasano, “IP QoS for voice: The diffserv approach,” in Proc. of ISIT’99, June 1999.

Carla-Fabiana Chiasserini received her Laurea degree in electrical engineering from University of Florence, Italy, in 1996, and her Ph.D. degree from Politecnico di Torino, Italy, in 1999. Since then she has been with the department of Electrical Engineering at Politecnico di Torino, where she is currently an Assistant Professor. She was at the University of California at San Diego as a visiting researcher in 1999 and 2000, and as a visiting professor in 2001. Her research interests include architectures, protocols and performance analysis of wireless networks.

Michela Meo received the Dott. Ing. degree in Electronic Engineering in 1993, and the Ph.D. degree in Electronic and Telecommunication Engineering in 1997, both from Politecnico di Torino, in Italy. Since then she has been working in the Telecommunication Networks Research Group of the Politecnico di Torino, where she is currently an Assistant Professor. Her research interests are in the field of performance evaluation of communication networks with a particular focus on wireless systems.