Introduction
Recently, broadband and high-quality wireless communication systems have been standardized in the European Telecommunications Standards Institute (ETSI) Broadband Radio Access Network (BRAN) [1], IEEE 802.16 [2], and other standards that support multimedia services. In order to realize broadband wireless communications, the use of millimeter- or centimeter-wave bands is being explored. Also, in order to realize high-quality communications, the use of automatic repeat request (ARQ) as well as forward error correction (FEC) is being considered. When the millimeter- or centimeter-wave band is used, securing a line of sight is a basic need. However, if there are some obstructions, consecutive frame errors occur. Thus, ARQ is especially needed. Although several ARQ schemes have been proposed [3–6] and introduced in actual systems, the ARQ scheme must be suited to the radio channel property and quality of services. Moreover, hardware size must be considered for scheme selection. Therefore, we need an ARQ scheme that is flexible against these system requirements.
The purpose of this article is to propose a new combined ARQ with selective repeat (SR) and go back N (GBN), and confirm its flexibility. We implemented the combined ARQ in hardware and put it in the medium access control (MAC) layer to execute broadband communications in the radio channel. We evaluated throughput, cell loss rate, and cell transfer delay on a pseudo Rayleigh fading channel. In addition, we also evaluated ARQ performance on an actual 40 GHz band radio channel.
The proposed combined ARQ scheme with SR and GBN is described. The experimental condition for evaluating the proposed ARQ is shown. We present experimental results and some discussion, and lastly we conclude the article with experimental results of our proposed ARQ.
Proposed Combined ARQ with SR and GBN
In our proposed ARQ scheme, there are two ARQ modes: SR-ARQ and GBN-ARQ mode. The combined ARQ operates as SR-ARQ mode in initial state and is then shifted to GBN-ARQ mode if the number of consecutive error frames is greater than or equal to GBN threshold n regardless of completion of SR retransmission. It goes back to SR-ARQ mode after not less than one out of some GBN retransmission frames is successfully received in a receiver.
The operation of this combined ARQ is as follows. The receiver executes an error check by FEC decoding, and then discards the data in the error frame if an error frame is found. In the next step, the receiver reads the number of correct ARQ frames as a sequence number. If the data in the error frame is discarded, the discontinuity of sequence numbers is found. Then the number of consecutive error frames e and GBN threshold n is compared. If e < n, SR-ARQ mode is selected: if e ≥ n, GBN-ARQ mode is selected. Figure 1 illustrates this operation with GBN threshold n = 3. When the receiver receives frame 4 after reception of frame 1, frame errors are detected in frames 2 and 3. Then SR-ARQ mode is selected for each error frame because e = 2 (< n). On the other hand, when the receiver receives frame 13 after frame 9, frame errors are detected in frames 10, 11, and 12. In this case, GBN-ARQ mode is selected because e = 3 (≥ n). When the receiver requires GBN retransmission, it sends the information of one request frame including frame number 10 to the transmitter, which is the first frame number of the consecutive frame error group. After the receiver receives not less than one retransmission frame in GBN-ARQ mode, it goes back to SR-ARQ mode.
Moreover, the receiver informs the transmitter of the frame number periodically. This number indicates the latest frame correctly received without discontinuity of sequence number.
After the receiver sends the retransmission request information, a timer in the receiver starts. If the receiver receives the data in the retransmission frame successfully, the timer stops. However, if the receiver cannot receive the data in the retransmission frame until the timer reaches the prescribed value, the receiver requires retransmission of the same error frame (RSEF) again. If the iteration number of RSEF operations reaches the maximum number r, current RSEF operation is stopped. If there are some remaining frames waiting to send request information for their RSEF, the frame whose waiting time is the longest has priority over other frames. Therefore, RSEF operations of other frames are postponed until current RSEF operation is succeeded or stopped. This operation is the same as a conventional stop-and-wait scheme, but data transmission in frames, except frames waiting for RSEF, is done independent of RSEF. This means data transmission in new frames and first-time data retransmission in each error frame are executed independent of RSEF.
Experimental Conditions
Layer 1 Format
We evaluate the proposed ARQ scheme on an actual 40 GHz band radio channel using a trial platform developed in the Advanced Communications Technologies and Services (ACTS) System for Advanced Mobile Broadband Applications (SAMBA) project [7, 8]. The multiple access scheme is time-division multiple access/frequency-division duplex (TDMA/FDD).
The air transmission bit rate is 64 Mb/s; the layer 1 radio frame has a fixed time period of 1.713 ms and is divided into 80 TDMA slots. The unit of retransmission is one slot, in which two asynchronous transfer mode (ATM) cells are stored. The Reed-Solomon code is also used for FEC. Each slot has code redundancy (CR) and control information (CI) fields. CR is for redundancy bits of FEC, and CI is for the control signal. In our experiment the CI field is used for transmission of information about ARQ status.
Only one slot in a frame is used as a control channel; another slot is used as a receiving acknowledgment channel to indicate the latest frame correctly received without discontinuity of sequence numbers, described earlier. The remaining 78 slots can be used for data transmission. Thus, this system has a capacity of up to 38.6 Mb/s of ATM cell transmission.
Main Parameters
In our lab experiment, a fading simulator is inserted only in the forward data link; the feedback link is assumed to be an error-free channel. Here, a single-path Rayleigh channel is simulated by the fading simulator. We call this channel a pseudo Rayleigh fading channel in this article. Therefore, the information of the receiving acknowledgment frame and retransmission request frame from the receiver is sent to the transmitter successfully all the time.
Table 1 shows the main parameters used in our experiment. Average frame error rate (FER) is defined as the probability that one radio slot is successfully received at the receiver using FEC. For the measurement of the throughput, ATM cell generation rate is set to 91,000 cells/s, which corresponds to link capacity of 38.6 Mb/s. For the measurement of cell loss rate and cell transfer delay, ATM cell generation rate is set to 20,000 cells/s. This rate seems fast enough to transmit MPEG-2 data.
Experimental Results
We measured the performance of the combined ARQ scheme on both an actual 40 GHz band radio channel and the pseudo Rayleigh fading channel in a laboratory environment. In the pseudo Rayleigh fading channel, Doppler frequency fD is set to 100 Hz or 4 kHz. In the following, throughput, cell loss rate, and cell transfer delay are discussed.
Throughput Performance
Figure 2 shows the throughput performance vs. FER on an actual 40 GHz band radio channel. In this environment, most of the frame errors seem to be single occurrences. The maximum number r of RSEF is always set to 7 in measuring throughput performance. In Fig. 2, the theoretical curves of SR-ARQ and GBN-ARQ are depicted with broken lines, where random frame error is assumed.
When n = 1, the ARQ always operates in GBN-ARQ mode. Therefore, the plots with n = 1 are close to the theoretical curve of GBN-ARQ. The plots with n = 32 are mostly on the theoretical curve of SR-ARQ because ARQ does not operate in GBN-ARQ mode. The plots with n = 2 are also mostly on the theoretical curve of SR-ARQ because two consecutive frame errors rarely occur.
Figure 3 shows the throughput performance on the pseudo Rayleigh fading channel. Black plots show the performance on condition fD = 4 kHz. Since it is observed that most of the frame errors randomly occur in this case, the performance is very similar to those on the actual 40 GHz band radio channel with n = 1 and 32, respectively. If n = 2, however, the throughput performance is degraded extremely in the range of FER > 5 
10–2. This is because two or more consecutive frame errors become dominant, which causes a need for frequent GBN-ARQ mode operation.
White plots in Fig. 3 show the performance on condition fD = 100 Hz. From Fig. 3, we can see that the throughput performance with fD = 100 Hz is better than that with fD = 4 kHz in case of n = 1. If the level fluctuation of a radio channel by fading is too slow, as in fD = 100 Hz, the fade duration in this condition is longer than a layer 1 radio frame duration, so consecutive frame errors are observed in the receiver. In this case, the number of GBN retransmission requests with n = 1 is smaller than that under random error conditions. For instance, if the frame error occurs randomly, the number of GBN retransmission request with n = 1 becomes approximately the same as the number of frame errors. However, if the frame errors occur consecutively, the number of GBN retransmission requests becomes smaller than the number of the frame error because the GBN retransmission is required only once after the consecutive error frames are recognized. Therefore, the number of frames that are discarded despite correct receipt is smaller, and the throughput performance is better than that under random error conditions.
In this experiment, buffer size for data queuing in the transmitter is so large that overflow of new data does not occur even if the combined ARQ executes many SR retransmissions. However, if n = 32 is employed on condition that the buffer size is small for hardware size reduction, overflow of new data may occur. Therefore, there is a possibility that throughput with a small GBN threshold is higher than with a large GBN threshold. We will take the optimum buffer size into consideration in designing the combined ARQ scheme.
As shown in Figs. 2 and 3, the throughput performance with n = 2 is as good as that with n = 32 on both random and consecutive error channels. Since buffer size has an impact on hardware size, the result that the performance difference is small between n = 2 and n = 32 is desirable for implementation.
Cell Loss Rate Performance
Figure 4 shows cell loss rate (CLR) performance vs. GBN threshold n on the pseudo Rayleigh fading channel.
Black plots show the performance on condition fD = 4 kHz. Here, the maximum number r of RSEF is set to one, the average FER is 4.28 10–2, and the ATM cell generation rate is set to 20,000 cells/s. GBN-ARQ mode always operates if n = 1; therefore, if RSEF operation fails, the number of lost cells becomes large. As shown in Fig. 4, if n ≥ 2, when fD = 4 kHz, CLR becomes low. However, it does not become zero because r is set to one, so RSEF operation is easily stopped. If r ≥ 3, no cell loss is observed during the experiment even if n is set to one on the pseudo Rayleigh fading channel with fD = 4 kHz. The reason is as follows. Since the frame error randomly occurs in this environment, the probability of successful retransmission is very high if retransmission is tried not once but a few times. Moreover, the link capacity is so large against the cell generation rate that a lot of channels can be secured for retransmission.
No cell loss is observed on the actual 40 GHz band radio channel with r = 7 either. The reason is the same as in the case above.
White plots in Fig. 4 show the CLR for fD = 100 Hz. Average FER is 3.28 10–2. Zero CLR cannot be obtained because of long fade durations in the radio channel.
We can see from Fig. 4 that the larger the maximum number r of RSEF, the lower the CLR becomes; and the larger the GBN threshold n, the lower the CLR becomes. The reason is as follows. With large r, the transmitter can execute RSEF operation many times (i.e., up to r + 1 times) on condition that the link capacity is set large enough to execute many retransmissions. Therefore, the probability that the retransmission data does not reach the receiver becomes very low. With large n, the CLR becomes low. Since many retransmissions are required in GBN-ARQ mode and some of them do not reach the receiver, the probability that the data transmission fails becomes large. Therefore, if large n is employed, the number of GBN retransmissions can be reduced, and the CLR becomes small.
From the results above, it is better to use large r and n to get good CLR performance.
Cell Transfer Delay Performance
Figure 5 shows the average cell transfer delay (CTD) performance for various GBN thresholds n on the pseudo Rayleigh fading channel.
Black plots show the performance when fD = 4 kHz. The CTD for n = 1 is larger than that for other n because many retransmissions are required in GBN-ARQ mode, and some do not reach the receiver. CTD performance with r = 3 or 7 is nearly the same as that with r = 1. With frame errors randomly occurring, the first retransmission for each error frame mostly succeeds. Therefore, it seems adequate to choose the maximum number r of RSEF to be one.
White plots in Fig. 5 show CTD performance when fD = 100 Hz. As we can see, the larger the GBN threshold n, the larger the CTD becomes. At all points of GBN threshold, CTD with r = 7 becomes larger than with r = 3. The reason is as follows. For large n, many SR retransmissions are required, so the frames waiting for RSEF are frequently generated on consecutive error channels. This behavior causes transfer delay because RSEF is executed one by one like a stop-and-wait scheme. On the other hand, for large r, the combined ARQ executes RSEF many times. This operation causes the data readout operation from the queuing buffer to be postponed until RSEF succeeds or stops, even if other frames are correctly received.
From the results shown above, CTD becomes small if small n and r are chosen.
Cell duration violation (CDV) is another issue concerning delay performance. It specifies a violation of a cell arrival interval. A maximum interval of cell arrival is also measured with n = 16 and r = 7 in our experiment. With fD = 4 kHz and average FER 3.28 10–2, the maximum interval of cell arrival is 1.7 ms. With fD = 100 Hz and average FER 4.28 10–2, it is 6.8 ms.
An interval of data readout from the queuing buffer in the receiver is equal to a layer 1 radio frame duration. If some error frames are found in each data readout operation, the operation is stopped at the last correct frame, which is one frame before the error frame. If the data retransmission of the error frame succeeds by the next data readout timing, the data readout operation is started again. Since the required interval of RSEF in our experiment is 16 slots, as shown in Table 1, RSEF can be executed five times during one layer 1 radio frame duration (80 slots). Therefore, on a random error channel such as fD =4 kHz, most retransmissions succeed within one layer 1 radio frame. However, on a slow fading channel such as fD = 100 Hz, often the data in the retransmission frame cannot reach the receiver correctly during one layer 1 radio frame, because the fade duration of the radio channel is longer than the frame duration.
It is enough to execute RSEF operation five times to obtain error-free transmission on random error channel. On the other hand, if error-free transmission is required on a slow fading channel, the maximum number r of RSEF should be larger than 20 (= 6.8 ms ÷ 1.7 ms 5). In this case, the time duration of RSEF is longer than the fade duration of the radio channel (= 6.8 ms). If error-free transmission is not needed under slow fading conditions, the maximum number r of RSEF can be smaller than 20. The smaller the maximum number r of RSEF, the smaller the average CTD becomes.
If the main parameters of the ARQ scheme are to be decided in a certain mobile communication system, system requirements such as the fade duration of the radio channel should be taken into consideration.
Conclusion
We propose a combined ARQ scheme with SR and GBN for broadband wireless communications systems and discuss its performance regarding throughput, cell loss rate, and average cell transfer delay on a pseudo Rayleigh fading channel. Moreover, a performance evaluation on an actual 40 GHz band radio channel has also been executed.
Concerning throughput performance, there is no significant difference between GBN threshold n = 2 and n = 32. It is desirable for hardware size.
If the combined ARQ executes retransmission of as many frames as possible in SR-ARQ mode, the cell loss rate becomes lower and the average cell transfer delay larger. Therefore, if error-free transmission has priority over transfer delay, this combined ARQ should be operated in SR-ARQ mode as much as possible. If small transfer delay has priority over error-free transmission, this combined ARQ should be shifted to GBN-ARQ mode at an early stage.
The time interval of ATM cell arrival was also measured. Since the number of consecutive frame errors closely depends on fade duration, the ARQ parameters should be included in considering the detailed behavior of radio channel.
Finally, we can design an optimum ARQ scheme by using the proposed combined ARQ in accordance with various system requirements such as throughput, cell loss rate, cell transfer delay, radio channel environment, and hardware size.
Acknowledgment
This article is based on our previously published material from WAS 2000 organized by DELSON GROUP.
References
[1] ETSI, "Requirements and Architectures for Wireless Broadband Access," TR 101 031 V2.2.1, Jan. 1999.
[2] IEEE 802.16, "Preliminary Draft Working Document for 802.16 Broadband Wireless Access System Requirements," IEEE 802.16s0-99/5, Oct. 1999.
[3] S. Lin, D. J. Costello Jr., and M. J. Miller, "Automatic-Repeat Request Error Control Schemes," IEEE Commun. Mag., vol. 22, no. 12, Dec. 1984, pp. 5–17.
[4] S. S. Chakraborty, E. Yli-Juuti, and M Liinaharja, "An ARQ Scheme with Packet Combining," IEEE Commun. Lett., vol. 2, no. 7, Jul. 1998, pp. 200–2.
[5] U. Lambrette, L. Brühl, and H. Meyr, "ARQ Protocol Performance for a Wireless High Data Rate Link," IEEE 47th VTC Proc., vol. 3, May 1997, pp. 1538–42.
[6] M. Yoshioka, A. Ohta, and M. Umehira, "A Novel PRIME ARQ for Wireless ATM," IEICE Technical Report, RCS97-197, Jan. 1998, in Japanese, pp. 37–44.
[7] M. Dinis et al., "The SAMBA Trial Platform in the Field," 4th ACTS Mobile Commun. Summit, Sorrento, Italy, June 1999, pp. 1013–18.
[8] T. Fujino et al., "Design of Baseband Signal Processing Unit in SAMBA Trial Platform for Mobile Broadband Applications," 3rd ACTS Mobile Commun. Summit, Rhodes, Greece, June 1998, pp. 854–59.
Biographies
Noriyuki Fukui received his B.E. degree in electronics and communication engineering from Waseda University, Tokyo, Japan, in 1990. He has been with Mitsubishi Electric Corporation since 1990, where he has been engaged in the research and development of wireless communication systems. His research interests include radio frame synchronization control, radio access control and error control technologies.
Akihiro Shibuya received his B.E. degree in electrical engineering from Tohoku University, Miyagi, Japan, in 1983. He has been with Mitsubishi Electric Corporation since 1983, where he has been engaged in the research and development of broadband wireless systems. His research interests are in the area of low layers of wireless networks, random access protocols, and error control technologies.
Keishi Murakami [M'90] received his B.E. and M.E. degrees in electronics and communication engineering from Waseda University, Tokyo, Japan, in 1974 and 1976, respectively. He has been with Mitsubishi Electric Corporation since 1976, where he has been engaged in research and development on digital mobile communication systems and digital satellite communication systems. He is currently manager of the Wireless Communication Systems Department in the Information Technology R&D Center of Mitsubishi Electric Corporation. His current research interests include digital modulation/demodulation, adaptive equalization, wireless communication signal processing, and wireless communication technologies.
