IEEE COMMUNICATIONS LETTERS

1

A MAC/PHY Cross-Layer Design for Efficient ARQ Protocols Dongmyoung Kim, Student Member, IEEE, Youngkyu Choi, Student Member, IEEE, Sunggeun Jin, Student Member, IEEE, Kwanghun Han, Student Member, IEEE, and Sunghyun Choi, Senior Member, IEEE

Abstract—In many wireless data systems, automatic repeat request (ARQ) and hybrid automatic repeat request (HARQ) work independently. In this paper, we introduce a new ARQ feedback scheme called Cross-Layer ARQ (CL-ARQ), where ARQ closely interoperates with HARQ by making use of the acknowledgement of HARQ internally. By doing so, CL-ARQ not only removes the overhead for ARQ acknowledgement messages, but also improves the average delivery delay and TCP throughput. Especially, when the entities in charge of HARQ and ARQ are located far apart geographically, CL-ARQ remarkably outperforms the legacy scheme. Index Terms—Automatic repeat request (ARQ), Hybrid ARQ (HARQ), Cross-layer approach

I. I NTRODUCTION Automatic Repeat reQuest (ARQ) is widely used in wireless systems to guarantee the reliable delivery of data packets. Besides the pure ARQ, many of recent wireless communication systems [1], [2] adopt Hybrid Automatic Repeat reQuest (HARQ) in order to combat wireless channel fluctuation [4]. Since HARQ scheme strongly depends on the physical layer techniques, HARQ often runs at the lower layer than the layer where ARQ operates. While both HARQ and ARQ basically pursue the reliable delivery of information, they complement each other rather than compete with. This can be justified as follows: 1) In many cases, HARQ only provides limited reliability because the maximum number of HARQ retransmissions is limited to a small value. 2) ARQ can provide the link level reliability such as inorder delivery while HARQ only considers the reliability of each channel-coded transmission data unit. However, the role of the acknowledgement for ARQ is somewhat overlapped with that of HARQ acknowledgement because successful transmissions of HARQ units are prerequisite to the successful delivery of ARQ unit. Motivated by this fact, we propose a new ARQ protocol, Cross-Layer ARQ (CLARQ), where an HARQ acknowledgment is utilized as an ARQ acknowledgement implicitly. By adopting CL-ARQ, we can reduce the total delivery delay and save the resource required for ARQ acknowledgement transmissions. Furthermore, TCP throughput can be increased due to the reduced delay. Dongmyoung Kim, Youngkyu Choi, Kwanghun Han, and Sunghyun Choi are with the School of Electrical Engineering and INMC, Seoul National University, Korea. Sunggeun Jin is with ETRI, Korea. This work was in part supported by the IT R&D program of MKE/IITA (2008-F-007-01, Intelligent Wireless Communication Systems in 3 Dimensional Environment) and the ITRC support program of MKE/IITA (IITA2008-C1090-0801-0013).

II. S YSTEM M ODEL Some wireless systems, where both HARQ and ARQ are employed, implement these two retransmission schemes at the different layers while the others at the same layer. For example, High-Speed Downlink Packet Access (HSDPA) locates HARQ and ARQ at Medium Access Control (MAC) and Radio Link Control (RLC) layer, respectively. On the other hand, IEEE 802.16e runs both HARQ and ARQ at MAC layer [1], [3]. Note that we consider the former type of systems in this letter. For the convenience of discussion, we will refer to the entity in charge of ARQ as MAC and the entity in charge of HARQ as PHY, respectively. We consider a time-slotted system and assume that multiplechannel interlaced HARQ is used. Under the operation of mchannel interlaced HARQ, the HARQ retransmission opportunity is given m slots later from the previous transmission, and other packets can be transmitted independently over other channels to fully utilize the wireless medium. Typically, the number of maximum retransmissions in HARQ is limited to a finite number, Kmax . If Kmax consecutive HARQ transmissions fail, the packet is discarded from the sender’s PHY, and then the recovery of the packet is up to the MAC ARQ. We consider Selective Repeat ARQ (SR-ARQ) as a retransmission scheme at the MAC, and assume that the retransmitted packets have the priority over new packets in channel access while the packets retransmitted by HARQ has the highest priority. Also, we assume that retransmissions by ARQ are repeated until the delivery is successful. III. P ROPOSED S CHEME In this section, we introduce CL-ARQ which is a new ARQ feedback scheme. In CL-ARQ, the acknowledgement of ARQ is not returned to the sender explicitly. Instead, the sender’s MAC infers the status of delivery by seeing the HARQ ACK/NAK message received by the sender’s PHY. For this purpose, the sender maintains a matching table which specifies the mapping between MAC Protocol Data Unit (MPDU) and PHY Protocol Data Unit (PPDU), and the relationship is used to translate HARQ ACK/NAK into ARQ ACK/NAK message. If HARQ transmission fails consecutively for Kmax times, an ARQ NAK message is generated by the sender’s PHY and sent to the sender’s MAC layer internally. Similarly, an ARQ ACK message is generated whenever an HARQ ACK arrives per a successful transmission. Note that CL-ARQ can be used along with any kind of ARQ protocol such as Go-Back-N or Selective Repeat ARQ because CL-ARQ just specifies the

IEEE COMMUNICATIONS LETTERS

MAC sender

2

ds MAC NAK (CL-ARQ)

m slots (Kmax= 2)

Retrans. (CL-ARQ)

MAC NAK (Legacy)

Retrans. (Legacy)

PHY sender HARQ ACK/NAK

PHY receiver

HARQ NAK (=MAC NAK for CL-ARQ)

error

dm

If success, send up pkt

Fail notify (Legacy)

MAC NAK (Legacy)

MAC receiver

dr

The timing diagram for transmission delays depending on ARQ schemes when Kmax = 2.

way how to send and receive the acknowledgement messages. If NAK-based ARQ scheme (e.g., RLC) is employed as MAC ARQ, we can translate only HARQ NAK into an internal feedback message. The operation of CL-ARQ is illustrated in Fig. 1, where the propagation delay between MAC and PHY at the sender is denoted by ds , and the propagation delay between MAC and PHY at the receiver by dr . Further, dm represents the wireless propagation delay. For the purpose of comparison, we consider a legacy scheme as follows: Whenever a packet is successfully received by the receiver’s MAC, an ACK for the packet is returned immediately. Similarly, NAK message is returned immediately after the receiver’s MAC realizes a delivery failure. The receiver can detect the failure by observing Kmax consecutive HARQ failures. This scheme provides the fastest way to return NAK messages among all the legacy ARQ schemes. The operation of this legacy scheme is also illustrated in Fig. 1. CL-ARQ has two different features compared with the legacy scheme. First, the receiver does not transmit ARQ ACK/NAK messages into the air. Accordingly, the corresponding amount of air resource required by the feedback channel can be reduced from the feedback channel. The second feature is that CL-ARQ can reduce the waiting time for ACK/NAK message, since the message is not generated by the receiver’s MAC, but directly by the sender’s PHY. This property is especially effective when it takes significant time to exchange packets between receiver’s MAC and PHY. For example, it has been known that the one way delay between BS and RLC can be several tens of millisecond in WCDMA systems. In this case, the MAC to MAC delay (ds + dm + dr ) is possibly much larger than the PHY to PHY delay (dm ) since the propagation delay incurred by wired line connecting PHY and MAC (either ds or dr ) is very large. IV. P ERFORMANCE E VALUATION We use ns-2 simulator [6] to evaluate the performance of CL-ARQ. In the simulations, we choose QPSK with 3/4 convolutional channel coding as modulation and coding scheme, and Chase Combining (CC) HARQ [4] is assumed. Furthermore, we adopt Jakes fading model [7] to describe the fast fading channel. We fix the set of reference parameters as follows: MAC packet size L = 1000 bits; transmission rate R =

250

Average delivery delay (msec)

Fig. 1.

CL-ARQ (5dB) CL-ARQ (7dB) Legacy (5dB) Legacy (7dB)

200

150

100

50

0 10

Fig. 2.

20

30 40 d_r (msec)

50

60

The impact of dr on the average delivery delay.

10 Mbps; Doppler frequency fd = 50 Hz; maximum number of the consecutive HARQ (re)transmissions Kmax = 3; number of HARQ channels m = 10; ds = 1 msec; dm = 1 msec. We also assume that there are only one transmitter and one receiver in the system for simplicity. A. Delivery Delay Here, we analyze the delivery delay of ARQ, which is defined as the time elapsed from a packet’s first transmission attempt to its delivery to the upper layer at the receiver. A delivery delay consists of transmission and resequencing delays. The transmission delay is defined as the time elapsed from a packet’s first transmission attempt to its successful reception at the receiver. On the other hands, the resequencing delay is defined as the time duration how long the packet waits in the resequence buffer at the receiver before delivered in sequence towards the upper layer. We evaluate the average delivery delay for different values of dr , i.e., wireline delay between MAC and PHY (base station side) in the uplink transmission. We consider two average SNR values, namely, 5 dB and 7 dB, which cause roughly 2% and 5% of transmission errors at the MAC layer, respectively. Fig. 2 compares the performance of CL-ARQ and legacy scheme with respect to the average delivery delay. We observe that the proposed CL-ARQ always yields lower

IEEE COMMUNICATIONS LETTERS

B. Performance of the TCP Traffic We now examine the TCP performance. TCP is the most popular transport layer protocol in the Internet. The size of the TCP segment is fixed at 1460 bytes, and dr is fixed at 50 msec as assumed in [5]. For TCP simulations, we consider the scenario that the file transfer is conducted during 30 seconds using File Transfer Protocol (FTP) on the top of a TCP connection. We obtain the averaged results after running each simulation scenario 8 times. First, we present TCP performance as the size of receive window varies. The error recovery mechanism of TCP basically belongs to sliding window-based ARQ. In sliding window-based ARQ, all the packets within a window can be (re)transmitted regardless of the transmission result of other packets. Whenever the first packet in the window receives ACK, the window moves forward to transmit the next packet. If the ACK for the first packet does not arrive, the packets, which are waiting to enter the window, cannot be transmitted. This phenomenon raises the issue that the link capacity cannot be fully utilized even if the medium is idle. We refer to this as ‘window stall problem’. If the size of window is K, the maximal achievable throughput is limited to K/RT T due to this problem. TCP uses the minimum value of congestion window and receive window size as transmit window size, and the TCP receive window size eventually determines the actual window size when the end-to-end bandwidth is sufficiently large. The TCP throughput versus receive window size is shown in Fig. 3. In the result, we observe that the window stall problem limits the TCP throughput for both ARQ schemes. However, CL-ARQ significantly outperforms the legacy ARQ scheme in the entire region. As explained before, the window size and RTT determine the TCP throughput in the environment where window stall problem exists. Since CL-ARQ can significantly reduce RTT, the TCP throughput can be improved. Then, we analyze the impact of the average SNR on the TCP throughput. In the simulations, we consider the two different values as TCP receive window size, i.e., 17.52 KB and 64 KB, because they are the recommended and maximum receive window sizes used by Windows XP, respectively. Fig. 4 shows that TCP throughput can be greatly enhanced by CL-ARQ in some environments. Furthermore, the throughput is less influenced by the channel condition when CL-ARQ is utilized. However, the performance gap between CL-ARQ and legacy scheme decreases as the average SNR value increases, and the gap finally goes to zero in the very high SNR region where retransmission never occurs.

6 CL-ARQ (5dB) CL-ARQ (7dB) Legacy (5dB) Legacy (7dB)

Throughput (Mbps)

5

4

3

2

1

0 0

Fig. 3.

10

20

30 40 50 Receive window size (KB)

60

70

80

TCP throughput performance with various receive window sizes. 7 CL-ARQ (17.52KB) CL-ARQ (64KB) Legacy (17.52KB) Legacy (64KB)

6

Throughput (Mbps)

delivery delay. It is noteworthy that the gain comes mainly from the improvement of the resequencing delay. Furthermore, the average delay is less sensitive to the channel condition in CL-ARQ. The delivery delay of CL-ARQ in 5 dB and 7 dB are quite similar because CL-ARQ allows fast retransmissions, while the legacy scheme imposes large delay in 5 dB region. Considering that many types of multimedia traffic requires the QoS requirement for delay, the improvement of delivery delay provided by CL-ARQ should be quite beneficial for multimedia applications.

3

5 4 3 2 1 0 3

Fig. 4.

4

5

6 7 Average SNR (dB)

8

9

10

TCP throughput performance for various average SNR values.

V. C ONCLUSION In this letter, we proposed a new ARQ scheme which uses the ACK/NAK of HARQ as the implicit acknowledgement for ARQ. By employing the CL-ARQ, we can save the airresource required for sending ARQ acknowledgement and reduce the delivery delay by allowing ARQ to retransmit faster. In addition, CL-ARQ can improve the TCP performance by reducing the impact of the window stall problem, which frequently happens in the real systems. R EFERENCES [1] IEEE 802.16e-2005 and 802.16-2004/Cor1-2005, Amendment to Part 16: Physical and Medium Access Control Layers for Combined Fixed and Mobile Operation in Licensed Bands and Corrigendum 1, Feb. 2006. [2] 3GPP TR 25.848 v4.0.0, Third Generation Partnership Project: Technical Specification Group Radio Access Network; Physical layer aspects of UTRA High Speed Downlink Packet Acces (Release 4), Mar. 2001. [3] 3GPP TS 25.322 v3.18.0, Third Generation Partnership Project: Technical Specification Group Radio Access Network; Radio Link Control (RLC) Specification (Release 1999), 3GPP, Jun. 2004. [4] D. Costello, J. Hagenauer, H. Imai, and S. Wicker, “Applications of error-control coding,” IEEE Trans. Inform. Theory, vol. 44, pp. 2531– 2560, Oct. 1998. [5] H. Lin and S. K. Das, “Performance Study of Link Layer and MAC Protocols to Support TCP in 3G CDMA Systems,” IEEE Trans. Mobile Computing, vol. 4, no. 5, pp. 489–501, Sept. 2005. [6] The Network Simulator – ns-2, [Online]. Available: http://www.isi.edu/nsnam/ns/. [7] W. C. Jakes, Microwave Mobile Communications. Wiley, 1974.