LETTER Performance Evaluation of Data Link Protocol with Adaptive Frame Length in Satellite Networks

Transcription

1 IEICE TRANS. COMMUN., VOL.E87 B, NO.1 JANUARY LETTER Performance Evaluation of Data Link Protocol with Adaptive Frame Length in Satellite Networks Eung-In KIM, Student Member, Jung-Ryun LEE, Nonmember, and Dong-Ho CHO, Member SUMMARY We propose a new data link protocol with an adaptive frame length control scheme for satellite networks. The wireless communication channel in satellite networks is subject to errors that occur with time variance. The frame length of the data link layer is another important factor that affects throughput performance in dynamic channel environments. If the frame length could be chosen adaptively in response to changes in the dynamically varying satellite channel, maximum throughput could be achieved under both noisy and non-noisy error conditions. So, we propose a frame length control scheme that acts adaptively to counter errors that occur with time variance. We model the satellite channel as a two-state markov block interference(bi) model. The estimation of the channel error status is based on the short-term bit error rate and the duty cycle of noisy bursts. Numerical and computer simulation results show that the proposed scheme can achieve high throughput for both dense and diffuse burst noise channels. key words: data link protocol, adaptive frame length, satellite channel, ARQ, markov model 1. Introduction Satellite based systems will play a very important role in the next generation of mobile communication systems. In a satellite network, mobile stations experience multipath fading, severe shadowing and weather impairments, which cause burst errors to occur frequently. Therefore, in order to provide reliable transmission of packet data in such an environment, an efficient data link protocol is required, even though this could cause slight service delays [1], [2], [3]. Error recovery schemes to cope with satellite channel noise have been investigated extensively. Conventional error recovery methods are based on automatic repeat request(arq) retransmission schemes, such as Stop-and-Wait(SAW), GO-Back-N(GBN), and Selective-Repeat(SR) [3], [4]. Other approaches are related to the hybrid ARQ(HARQ), in which automatic repeat request scheme is combined with forward error control(fec) [5], [6], [7]. In the ARQ and HARQ schemes, a fixed frame length is usually used, which is designed with the worst channel condition and the overhead required per frame in mind. The data link Manuscript received September 4, Manuscript revised October 4, The authors are with the Communication and Information Systems Lab., Department of Electrical Engineering and Computer Science, Korea Advanced Institute of Science and Technology(KAIST). protocol designer chooses a frame length that is small enough to transmit information when the error status is at its worst [9], [10]. So, when channel conditions are good, the previously fixed frame length would be relatively small, because larger frames could be transmitted without error. In other words, in a clean channel, a longer frame has some advantages because more information can be transmitted. By contrast, in burst error channels, a longer frame is more easily contaminated by fading than a shorter one, resulting in the inefficient use of channel capacity. In this case, the shorter one becomes the candidate. However, the smallest frame length does not show the best performance because of the fixed overhead required in the data link protocol. The frame length of the data link layer is another important factor that affects the throughput performance of satellite packet transmission. If the frame length could be chosen adaptively in response to the dynamically varying satellite channel, maximum throughput could be achieved in both noisy and nonnoisy error conditions. In order to maximize the longterm throughput performance, the short-term performance must be optimized over the time-varying satellite channel. This short-term best performance could be achieved if the optimal frame length could be chosen almost instantaneously. Further, the optimum frame length could be obtained by proper estimation of the satellite channel. So, if the short-term channel status of the instantaneous burst error channel could be estimated, good throughput performance would be achieved continuously. However, the estimation of error channel is difficult because there are various kinds of degradation due to fading, shadowing, and propagation loss. The bit error rate(ber) and the duty cycle of noisy bursts during the short-term interval result from channel error. At the receiver, if the bit error rate and the duty cycle of noisy bursts in current channel are estimated from cyclic redundancy check bits of received frames, the channel status will be estimated indirectly. In this letter, an adaptive frame length control scheme for the data link layer is proposed that will reduce the burst error effect and improve performance over the satellite networks. We use a two-state markov block interference model as a channel model of the satellite networks [7], [11], [14]. This model does not include all factors of the channel error, but is sufficient

2 2 IEICE TRANS. COMMUN., VOL.E87 B, NO.1 JANUARY 2004 GOOD P01 1-P01 1-P10 Fig. 1 P10 BAD A two-state markov model to show the behavior of the error status. In addition, the estimation of the bit error rate(ber) and the duty cycle of noisy bursts during a short time interval is used for deciding the optimum frame length. The remainder of this letter is organized as follows. In section 2, the bursty channel model is discussed. In section 3, throughput analysis of a frame length control scheme is introduced and computer simulation results and discussion are presented in Section 4. Conclusions are presented in Section Channel Model We model the satellite channel as a two-state markov process [12], [13]. This traditional Gilbert error modelling approach is not identical to a real satellite error channel, but it has the advantages of simplicity and manageability. Lutz and Schodorf showed that this markov model could be used as a model of satellite error [11], [12]. In a two-state markov model, the states are classified into GOOD and BAD states, each with a BER (good state with B q, bad state with B n, and B n > B q ). The BAD state means a channel situation in which it is difficult to receive the packet frames properly. B n and B q values depend on the channel environments, including fading, shadowing and scattering. We assume that the state transitions occur with probabilities p 01 and p 10, respectively, whenever each bit is transmitted, as shown in Fig. 1. For the markov channel model, the average burst length transmitted while in the BAD state is given by N b = 1 p 10 (1) Also, the average BER is described as follows B av = p 01B q + p 10 B n p 01 + p 10 (2) Moreover, the duty cycle of the noisy bursts, or the probability of being in the noisy state is given by p 01 P dc = (3) p 01 + p 10 Four parameters, such as N b, B av, P dc, and B n /B q, decide the two-state channel model. N b and P dc characterize the burstiness of the channel, with p 01 and p 10 describing the time variation of the channel behavior. To reduce the complexity, we can use the approach proposed by Lugand et al [7]. For given B av and P dc, B q is described as B q = B av P dc (4) From (2) and (4), we can obtain the bit error rate of the BAD state as B n = B av P dc (1 P dc )B av (5) for P dc > B av ( 1 2 B av)( B av) 2B av (6) In particular, when P dc = p 01 = 1 p 10, this channel model becomes the two-state block interference(bi) channel model proposed by McEliece and Stark [14]. The BI model is completely determined by P dc and B av. So, we assume that the satellite channel is modelled by a two-state markov BI model. In this model, the burstiness is determined by the proper value of P dc. The higher the value of P dc, the less burstiness, because the values of p 01 and (1 p 10 ) increase, even though the long term average bit error rate remains the same. 3. Throughput Analysis of Frame Length Control Scheme The frame length of the data link layer is usually decided when the service connection is established, considering the error status of the wireless satellite channel. However, during the service period, the status of the wireless satellite channel varies dramatically. So, optimum throughput is not obtained with fixed frame length. In order to reduce the number of retransmissions and conserve channel resources, an adaptive frame length scheme is required. Channel errors are caused by various factors, such as fading, shadowing, scattering, and path loss. It is very difficult to predict the cause of bit corruption. So, in this proposed scheme, we use a short-term average bit error rate and the duty cycle of noisy bursts as criteria of frame length control. During the short time interval, the average BER and duty cycle are calculated at the receiver. Then, the receiver feeds back the short term BER and the duty cycle to the transmitter. The transmitter will send the frame with the optimum length by considering the feedback information. 3.1 Frame Error Rate From the markov model shown in Fig. 1, we can obtain each markov state s time portion in the equilibrium

3 LETTER 3 Fig. 2 Frame error rate vs. bit error rate when P dc = 0.25 status, π as π = πp [ ] [ ] [ ] 1 p π1 π 2 = π1 π 01 p 01 2 p 10 1 p 10 (7) From (7), the time portion of the GOOD state(π 1 ) is p 10, while p 01 is the time portion of the BAD state(π 2 ). If a frame with l bits is transmitted over the channel, the frame error probabilities of the two states are 1 (1 B q ) l and 1 (1 B n ) l, respectively. So, the frame error rate(fer) of the channel is described as follows. f(l, P dc, B av ) = π 1 {1 (1 B q ) l } + π 2 {1 (1 B n ) l } = P dc {1 (1 B av P dc ) l } + (1 P dc ){1 (1 B av P dc + (1 P dc )B av ) l } (8) As shown in (8), the frame error rate depends on the average bit error rate(b av ), the duty cycle of noisy bursts(p dc ) and the frame length transmitted (l). Fig. 2 shows an example of frame error rate versus bit error rate when P dc is With a fixed noisy duty cycle, the average BER varies from 10 7 to Then, the frame error rate is calculated from (8) for different frame lengths(1440 bits, 2880bits, and 8640 bits). The frame error rate has a value between 0 and 0.2, as shown in Fig. 2. The frame error rate increases when the bit error rate increases, and under a channel error with the same bit error rate, the frame error rate is low when the frame length is small. From Fig. 3, we can see the relationship among noisy burst, frame length, and frame error rate when the bit error rate is 10 6, too. It is clear that the frame error rate increases when the duty cycle of noisy bursts decreases, since the low Fig. 3 Frame error rate vs. duty cycle of noisy bursts when B av = 10 6 duty cycle value means that the channel status is more bursty. This is caused by the fact that the average 1 burst length (N b ) is equal to 1 P dc (1). For the same bit error rate, the frame error rate varies according to the duty cycle of noisy bursts. 3.2 Throughput Analysis In order to formulate the relationship between the throughput and the channel model parameters (B av, P dc ), we use the following definitions. Let the total amount of information to transmit be L bits and the frame length be l bits, while the header length of each frame is h bits. The transmitter sends the data with M bps data transmission rate. The propagation delay of the satellite channel is assumed to be t pd second. Then, the number of frames to transmit is L l and each frame has l + h bits long. Let K be the number of missing frames when L l frames are transmitted over the channel. Then, the probability density function of K and the expectation of K are described as follows, when the frame error rate is f(l + h, B av, P dc ). P [K = k] = ( L l k ) (f(l + h, B av, P dc ) k (1 f(l + h, B av, P dc )) L l k ) (9) E[K] = L l f(l + h, B av, P dc ) (10) The expectation of K means the average number of frames contaminated during the transmission over the channel. So, we need to retransmit k frames that were not received properly. And we know that k frames should be retransmitted if k frames among retransmitted K frames are also destroyed during the retransmission. If we let the total elapsed time caused by all

4 4 IEICE TRANS. COMMUN., VOL.E87 B, NO.1 JANUARY 2004 retransmissions be T rt, T rt is given by T rt = L l f(l + h, B av, P dc ) l + h M + L l f(l + h, B av, P dc ) 2 l + h M + L l f(l + h, B av, P dc ) 3 l + h M +... L(l + h) f(l + h, B av, P dc ) = Ml 1 f(l + h, B av, P dc ) (11) And we let T fix be the absolute time to transmit L l frames. Then, T fix = t pd + L l+h l M. Hence, the total transmission time (T tx ) of L data bits when the frame length is l, including retransmission, is given by T tx = T fix + T rt L(l + h) 1 = t pd + ( Ml 1 f(l + h, B av, P dc ) ) (12) If we define the throughput as the throughput (η) is formulated as η = L T tx = t pd + L data bits successfully received total elasped time, L(l+h) Ml(1 f(l+h,b av,p dc )) (13) From (13), we know that the throughput depends mainly on the frame error rate (f(l + h, B av, P dc )) and the frame length(l). On the other hand, the frame error rate depends on three parameters: frame length, average bit error rate, and the duty cycle of noisy bursts, as shown in the previous subsection. The optimum frame length can be analytically calculated under the satellite error channel determined by B av and P dc, since we can obtain the frame length(l) which maximizes the throughput performance. If the average BER and duty cycle can be estimated accurately from the data bits previously received during the proper period at the receiver, we can estimate the optimum frame length for good throughput performance. Fig. 4 shows the throughput performance versus the average bit error rate when a noisy burst has a fixed duty cycle(p dc = 0.25). This figure is a graphical representation of the relationship between throughput and bit error rate from (13). We assume the data transmission rate(m) is 144 Kbps and the noisy duty cycle is The throughput is calculated for typical frame lengths, such as 1440 bits(= 10ms), 2880 bits, and 8460 bits, when a frame header length is 720 bits. From Fig. 4, in the case of higher BER, we know that a shorter frame length shows better throughput performance. By contrast, a longer frame length shows good performance when the channel error is low. However, in a mid-level error environment in which the bit error rate is between 10 5 and 10 3, the throughput performance curves are crossed. In this case, the frame length that shows maximum throughput is variable. That is, the optimum frame length depends on the bit error rate when the Fig. 4 Throughput vs. bit error rate when P dc = 0.25 and M = 144Kbps Fig. 5 Throughput vs. duty cycle of noisy bursts when B av = 10 4 and M = 144Kbps duty cycle is fixed. We can see the effect of the duty cycle on the throughput in Fig. 5. Like the bit error rate case, we assume that M = 144kbps and that the duty cycle varies from 0 to 1. For the same bit error rate, different throughput performance is obtained according to the duty cycle of noisy bursts. A low value of the duty cycle results in low throughput and a high duty cycle value results in high throughput performance. The crossed throughput curve between 0.3 and 0.4 results from the frame header of 720 bits. Similarly, the optimum frame length to show good throughput performance depends on the duty cycle under the the given average bit error rate. Fig. 6 shows the throughput versus frame length when the duty cycle is fixed. It is necessary to note the throughput curve when the bit er-

5 LETTER 5 Table 1 Simulation Parameters Data packets 10 minutes Error model 2-state markov BI Variable frame length 10ms,20ms,...,60ms ARQ Scheme Selective repeat Error detection CRC Data transmission rate 144 Kbps Propagation delay 250 ms Redundancy(overheads) 5 ms No. of retransmissions 5 Fig. 6 Throughput vs. frame length when P dc = 0.25 and M = 144Kbps ror rate is The throughput neither increases nor decreases when the frame length increases from 1000 bits to bits. The maximum throughput is obtained with a frame length of around 1500 bits. This result shows that an optimum frame length exists for specific BER and duty cycle values. From the above results, we know the optimum frame length can be chosen over the entire range of the channel error. If the length of the transmitted frame could be controlled adaptively according to the shortterm bit error rate and duty cycle of noisy bursts, we could achieve better performance, compared with using a fixed frame length. 4. Results and Discussion 4.1 Simulation Environment In this simulation, the following assumptions are used. 1) We assume the packet transmission is made at the forward channel. 2) The reverse channel is assumed to be error-free and used only to transmit feedback information, including some acknowledgement information(acks or NAKs), a short term average bit error rate and the duty cycle of noisy bursts. 3) The shortterm average bit error rate and the duty cycle of noisy bursts are assumed to be always estimated successfully based on CRC. 4) The sliding window size and buffer size are assumed to be large enough not to degrade the performance. The sender transmits the frame successively for a given time(10 minutes) and the receiver stores and checks the received frames. Then, the short term average bit error rate and the duty cycle of noisy bursts are calculated during a given short time(about 600 milliseconds). The channel error status is provided by two-state markov model parameters. If the average bit error rate B av and the duty cycle P dc are given, the probabilities (p 01, p 10 ) and BERs of each state(b q, B n ) are determined, as explained in the previous section. We varied the long term average BER (B av ) from 10 6 to 10 2 in order to cover the entire error range of satellite networks. We assume that the burstiness of the satellite channel is diffuse(large duty cycle and low intensity), or dense(low duty cycle and high intensity). These terms were first introduced by Massey [15]. A value of P dc = 0.25 means a more diffuse burst channel than does a value of P dc = By contrast, a channel is assumed to be dense when P dc = However, the long-term average BER is almost the same for both cases. Firstly, fixed length frames from 10ms to 60ms with 5ms header are transmitted through dense and diffuse channels with different average BERs. In addition, frames with adaptively controlled length are delivered under the same channel conditions in order to compare previous schemes with the proposed scheme. The optimum frame length at the transmitter is decided from a table consisting of the frame length, bit error rate and duty cycle of noisy bursts. The parameters of the table are collected from the previous simulation for the entire channel error variation. Table 1 shows the simulation parameters used in this simulation. The data transmission rate from the transmitter is 144Kbps. A selective repeat ARQ scheme is used. The propagation delay of satellite networks is assumed to be 250ms and the number of retransmissions is limited to five. 4.2 Results and Discussion Fig. 7 shows the results of both numerical and simulation results when P dc is The comparison is made for two frame lengths (F L ), such as 10ms(= 1440bits) and 60ms long. Similar results are obtained from analysis and simulation. However, there are also some throughput gaps. On the whole, the throughput performance of computer simulation is better than the numerical results. This phenomenon is due to the fact that we restricted the number of retransmissions to five. However, for higher BERs, the numerical results are better. This is because in numerical analysis, we do not consider the time to feed back control messages, such as ACK, NACK, and channel status. In this simulation, the control messages are more frequently fed back to

6 6 IEICE TRANS. COMMUN., VOL.E87 B, NO.1 JANUARY 2004 Fig. 7 Comparison between numerical results and simulation results when P dc = 0.25 Fig. 8 Throughput performance vs. bit error rate when the channel has diffuse noisy bursts(p dc = 0.25) the transmitter when the channel error is high. So, the cause of performance degradation for high BERs is the time taken for information feedback. Fig. 8 shows the performance of the proposed and conventional schemes when P dc = When the frame length is fixed, the performance varies according to the long-term bit error rate, whose range is from 10 6 to With a frame length of 60ms, optimum throughput is achieved when BER is 10 6, but the worst occurs when BER is By contrast, a frame length of 10ms is the optimum when BER is In the middle range of BER between 10 5 and 10 3, the optimum frame length varies. Throughout the entire channel error variation, the adaptive frame control scheme operates well, as shown by the throughput curve of adaptive frame length in Fig. 8. When the error rate is low, the new proposed scheme performs at 99.8% of the optimum. When the BER is 10 3, it has 99.2% of the throughput that could be acquired in case of 10ms frame length, which is an optimal frame length for a BER of Fig. 9 shows that the proposed scheme could be applied to the dense burst error channel, too. Also, from Fig. 8 and Fig. 9, we can see that the adaptively controlled frame length scheme shows good throughput performance. In order to achieve optimum throughput over all ranges of bit error rate, it is necessary to estimate exactly the current channel error status. However, it is almost impossible to estimate exactly the satellite channel error because the channel status changes dynamically and the cause of degradation is indistinct. We just use the short term average BER and the duty cycle of noisy bursts to estimate the satellite channel, so this inaccurate estimation is assumed to be the cause of the Fig. 9 Throughput performance vs. bit error rate when the channel has dense noisy burst(p dc = 0.05) slight decline in performance from the optimum. There are a number of parameters in addition to the frame length that influence throughput performance, for example, sliding window size, buffer size, and timer duration. We can adaptively change the window size and timer duration, but in the case of the selective repeat ARQ scheme, the performance could be improved only a little when adaptive sliding window size and timer duration are used. However, if we use the Go-Back-N ARQ, better performance can be achieved by using adaptive window size and timer duration.

7 LETTER 7 5. Conclusions In this letter, we have studied a data link layer protocol with adaptive frame length. We have proposed a scheme that adaptively controls frame length based on the error status of a wireless satellite channel. We modelled the satellite channel as a two-state markov block interference model. The error channel is estimated by using the short-term average bit error rate and the duty cycle of noisy bursts. We analyzed the proposed frame length control scheme and ran simulations for both dense and diffuse burst error environments. We found that better performance could be achieved by using the newly proposed scheme. In other words, better performance can be achieved by changing the frame length in response to the channel error status, compared with the scheme using a fixed frame length. [1] S. Lin, D. J. Costello Jr., Error Control Coding: Fundamentals and applications, Englewood Cliffs, Addison-Wesley, [2] S. Lin, D. J. Costello Jr., M. J. Miller, Automatic-Repeat Request Error-Control Schemes, IEEE Communications Magazine, pp. 5-17, Dec [3] Jing Zhu, Sumit Roy, An Adaptive Two-copy Delayed SR- ARQ for Satellite Channels with shadowing, IEEE Vehicular Technology Conference, pp , [4] L. Casone, G. Ciccarese, M. D. Blasi, L. Patrono, and G. Tomasicchio, An Efficient ARQ Protocol for a Mobile Geostationary Satellite Channel, IEEE Global Telecommunications Conference, pp , [5] W. Li, V. K. Dubey, C. L. Law, An adaptive hybrid-arq scheme combating burst-errors caused by power control lag in Ka-band LEO satellite systems, IEEE Military Communicatins Conference, pp , [6] Jeffrey B. Schodorf, Mark A. Gouker, Performance Evaluation of a Hybrid ARQ Protocol Implementation for EHF SATCOM on the Move Systems, IEEE Military Communicatins Conference, pp , [7] L. R. Lugand, D. J. Costello Jr., R. D. Deng, Parity Retransmission Hybrid ARQ Using Rate 1/2 Convolutional Codes on a Nonstationary Channel, IEEE Trans. on Com., Vol.37, No.7, pp , July [8] S. Kallel, Analysis of memory and incremental redundancy ARQ schemes over a nonstationary channel, IEEE Trans. on Com., Vol.40, No.9, pp , September [9] 3GPP2, Data Service Options for Spread Spectrum Systems:Radio Link Protocol Type 3 [10] 3GPP TS V5.0.0, Radio Link Control(RLC) Protocol specification [11] E. Lutz, D. Cygen, M. Dippold, F. Dolainsky, and W. Papke, The Land Mobile Satellite Communication Channel - Recording, Statistics, and Channel Model, IEEE Trans. on Vehicular Tech. Vol.40, No.2, pp , May [12] Jeffrey B. Schodorf, Error Control for Ka-Band Mobile Satellite Communication Systems, IEEE Vehicular Technology Conference, pp , [13] A. Konrad, Ben Y. Zhao, Anthony D. Joseph, Reiner Ludwig, A Markov-Based Channel Model Algorithm for Wireless Networks, U.C Berkeley Technical Report, UCB/CSD , May [14] R. J. McEliece, W. W. Stark, Channels with block interference, IEEE Trans. Inform. Theory, Vo.IT-30, No.1, pp.44-53, Jan [15] J. L. Massey, Coding techniques for digital communications, notes for a tutorial session, IEEE Int. Conf. Commun., Seattle, Jun References