Data Link Control Protocols for a Wireless ATM Network

4 downloads 0 Views 311KB Size Report
On the other side the transmitter keeps sending cells as long as the Sliding Window is not full. When the window becomes full the transmitter sets a timeout timer.
Data Link Control Protocols for a Wireless ATM Network Dimitris Skyrianoglou and Lazaros Merakos Communication Networks Laboratory, Department of Informatics, University of Athens E-mail: { dimiski | merakos }@di.uoa.gr, ABSTRACT Two ARQ schemes, the Selective Repeat Protocol (SRP) and the Partial selective Repeat superIMposEd on gbn (PRIME) ARQ protocol are examined as alternatives to Go-Back-N ARQ, the ARQ scheme used in WAND (Wireless ATM Network Demonstrator, ACTS project AC085). The work is focused on the performance evaluation and performance comparison of these two schemes as well as on the comparison of these schemes with GBN ARQ. 1. INTRODUCTION Wireless Asynchronous Transfer Mode (Wireless ATM) has presently evolved as a promising technology for extending the ATM capabilities to the Wireless environment so as to provide mobile users with new services beyond voice and data. To this end, ACTS AC085 project Magic WAND (Wireless ATM Network Demonstrator) [1],[2] aimed at specifying an indoor Wireless ATM system with QoS guarantees for Mobile Users. A typical WAND system consists of several Access Points (APs) connected to a classic ATM switch which connects WAND with the fixed network. The switch is controlled by a Control Station attached to the switch. Each Mobile Terminal (MT) is associated with an AP and communicates through it with the core ATM Network. The main obstacle in extending ATM to a wireless environment is the nature of the wireless environment itself. While ATM assumes reliable, full duplex, point-to-point communication, on wireless links the transmission is broadcast, errorprone and there is no dedicated direction. These problems call for the utilisation of a DLC (Data Link Control) Layer between the ATM and the Physical Layers. The major function of DLC is performing error and retransmission control, thus supporting the operation of the MAC Layer, which controls the access to the radio link, and improving the transmission quality. In Magic WAND the MAC layer is built around the concept of MAC Time Frame. MAC Time Frame consists of four variable length periods. The Frame

Header (FH) period, used for transmitting the frame structure to MTs, the Downlink and Uplink Reservation Period, used by the AP and the MTs respectively for transmitting reservation based downlink and uplink traffic, and the Contention Period, used for transmitting unexpected reservation requests. Each of these period is further divided in time slots. The duration of a time slot equals the time needed for the transmission of an ATM cell payload (48 bytes) plus the MAC radio header (6 bytes) Within each period, data is transmitted in the form of MAC Protocol Data Units (MPDUs).Each MPDU consists of an MPDU-Header and an MPDU-Body which, in turn, consists of a variable number of ATM cells. [3] 2. DATA LINK CONTROL (DLC) In order to cope with transmission errors, Magic WAND system applies Go-Back-N (GBN) ARQ Protocol in DLC Layer. This decision was basically based upon the obvious simplicity of GBN Protocol. This is a definite advantage of GBN but, as simulation studies indicate, it appears that GBN introduces a heavy load of retransmissions due to the fact that even correctly transmitted cells might have to be retransmitted several times before they are delivered to the ATM Layer. Moreover, the high number of retransmissions cause a significant increase to the mean delay of all connections.[2][4] Trying to overcome these drawbacks, two alternatives ARQ schemes for DLC Layer will be discussed in this paper. The first alternative is the application of Selective Repeat Protocol (SRP) ARQ, while the second alternative discussed is the application of PRIME (Partial selective Repeat superIMposEd on gbn) ARQ in WAND DLC Layer. 2.1 Selective Repeat Protocol (SRP) ARQ In order to avoid the multiple retransmission problem, introduced by the use of GBN, Selective Repeat (SRP) Protocol would the natural choice. In SRP each cell is assigned a Sequence Number (SN). SNs are assigned modulo M (with M being a positive integer), so as to avoid excessively large SNs. The protocol uses a Sliding Window of size N.

This window specifies the SNs of the cells which are allowed to be sent by the transmitter without receiving an acknowledgement from the opposite side. N is an important parameter of the system. [5],[6] In SRP, the receiver side uses also a Sliding Window of size N and accepts all the cells that are correctly received and have a valid SN (i.e. within the window limits). If these cells are in the right order they are delivered to the upper layer (ATM) and the window is slided accordingly, otherwise they are stored in the buffer until all the cells preceding the stored ones have been correctly received. In WAND system, cells arrive at DLC Layer in groups within MPDUs. After receiving an MPDU the receiver has to extract the correctly received cells and then prepare the feedback for the transmitter. Feedback in SRP can be of various forms but for the system under study the feedback is assumed to have the form of explicit negative acknowledgements (i.e. the receiver informs the transmitter about the cells that have NOT been received yet). This feedback is piggybacked on the opposite direction traffic (in WAND system all connections are full duplex). In case there is no traffic in the opposite direction at that moment, DLC generates a dummy cell (assigning to it an invalid SN). On the other side the transmitter keeps sending cells as long as the Sliding Window is not full. When the window becomes full the transmitter sets a timeout timer. If no feedback is received until the expiration of the timer, the transmitter resumes the transmission of the cells in the window. When feedback is received the transmitter re-sends the negative acknowledged cells, erases the acknowledged cells and slides the window accordingly. 2.2 PRIME (Partial selective Repeat superIMposEd on gbn) ARQ The main drawback of SRP protocol is its increased complexity. Another major problem is the fact that it is not always feasible (especially in the case of large N) to piggyback the required feedback within a cell header or an MPDU header. For these reasons, a hybrid solution combining the advantages of SRP with the simplicity of GBN would be desirable. PRIME ARQ protocol, first proposed for AWACS LLC Layer [7], tries to combine the effectiveness of SRP with a simpler retransmission mechanism requiring a smaller amount of feedback than SRP.

In PRIME ARQ the transmitter side operates in the same fashion as in SRP. The main differences lie at the receiver side. The receiver keeps a list of NMSN Managed Sequence Numbers (MSN). Upon arrival of a cell the receiver checks whether the SN of the cell matches with one of the MSNs. If this is the case, the cell is accepted and the MSN list is updated accordingly, otherwise the cell is rejected. The control flow chart of PRIME ARQ at the receiver (assuming NMSN=3) is shown in figure 1. Yes Cell Arrival

SN=MSN 1

No

Yes

SN=MSN 2 SN := GetCellSN( )

No

Yes

SN=MSN 3

MSN1:=MSN2

MSN2:=MSN3

MSN3:=MSN3+1

No

Cell Rejected

Cell Accepted

Figure 1. Control Flow of PRIME ARQ at the Receiver After receiving a whole MPDU and deciding upon acceptance or rejection of the cells within it the receiver prepares the feedback to be sent to the Transmitter. This feedback has again the form of negative acknowledgements but instead of sending NAKS for all erroneously received cells, it sends only a limited number of NTSN Transmitted Sequence Numbers. Generally NTSN = NMSN, but is is possible to set NTSN < NMSN without violating the consistency of the protocol. This is a major advantage for PRIME ARQ since it eliminates the need for extensive feedback for the transmitter. In general the relation that holds for Window Size N, NMSN, and NTSN in PRIME ARQ is: N  NMSN  NTSN It turns out that (assuming NTSN = NMSN,), PRIME ARQ performs the same as SRP ARQ when the number of erroneous cells per frame are less than NMSN. If the number of erroneous cells are more than NMSN then the cells after the NMSN erroneous cells are retransmitted sequentially in the fashion of GBN ARQ. 3. PERFORMANCE EVALUATION In order to evaluate the performance of the proposed DLC schemes, a detailed simulation model of WAND system was developed with the use of

system performance considerably and it is worth considered as a valid alternative for WAND DLC. Mean Delay for SRP and GBN (Input Load 54%, Timeout = 22msec) 8 Mean Delay (in sec)

OPNET simulation tool [8]. The behaviour of the fading radio channel was simulated, using the finite radio channel model developed for the WAND system [9]. According to this model, the channel is modelled as a Semi-Markov process where each state corresponds to a certain bit or cell error rate (BER or CER). Finally, the traffic models used for the simulation were based on independent discretetime batch Markov arrival processes (D-BMAP), proposed in [10].

SRP GBN

6 4 2 0

The simulation parameters used during the evaluation of the proposed schemes are presented in tables 1 and 2.

Connections Characteristics Mean Rate = 1024 kbps Deviation () = 512 kbps wireless hop CDT = infinite Table 2: Connections Characteristics The performance measures used for the evaluation and the comparison of the proposed schemes are the following:  Mean Delay. This is the mean transfer delay experienced by each ATM cell on the wireless hop, until its delivery to the ATM layer.  Mean Queue Size. This is the mean number of cells in the transmitter buffer.  Mean number of Transmissions. This is the average number of attempts required for a successful transmission of a cell. Figure 2 presents the comparison of SRP and GBN in terms of Mean Delay. The measurements regarding GBN were obtained from [2]. As it can be observed the mean transfer delay for SRP is substantially lower than GBN. In some cases SRP performs almost 10 times better than GBN in terms of cell transfer delay. This observation proves that, despite its increased complexity, SRP improves the

12

16 20 24 Window Size N

28

32

Figure 2. Comparison of SRP and GBN Figures 3, 4 and 5, illustrates the performance of SRP in terms of Mean Delay, Mean Queue Size and Mean Number of Transmissions for various values of window size and timeout duration. The results indicate that both of these parameters affect the system performance. As it can be observed, in figures 3 and 4, by increasing the window size, both the mean delay and mean queue size values decrease. This is expected since larger window means that the transmitter can transmit more cells without waiting for feedback, thus the cells wait less in the transmitter buffer. On the other hand increasing the timeout duration leads to greater delays and greater queue size since, in this way, cells have to remain for more time in the transmitter’s buffer, if no feedback is received. Mean Delay in SRP(Load 45%) Mean Delay (ms)

Channel Characteristics Channel capacity = 20 Mbps Slot duration = 2210-6 sec (time required for the transmission of an ATM cell) MT-to-AP distance = 33m SNR at the receiver = 30 dB Average CER = 3.3 * 10-3 Average BER = 1 * 10-1 Table 1: Channel Characteristics

8

1000 800

Timeout Duration = 1000 slots (22msec) Timeout Duration = 1500 slots (33msec)

600 400 200 0 8

12 16 20 24 28 32 Window Size N

Timeout Duration = 2000 slots (44msec)

Figure 3. Mean transfer delay in SRP The mean number of transmission attempts for each successfully delivered cell in SRP is shown in figure 5. As it can be observed, by increasing the window size, the mean number of attempts also increases. This is again explained by the fact that with large values of window size, more cells get transmitted, without first receiving feedback, and this increases the probability of a successful transmission. On the contrary, large values of timeout duration delay

8

12

16

20

24

28

32

Window Size N

2100

1800

1500

50 0 2100

1

100

1900

1.02

Timeout Duration = 2000 slots (44msec)

1700

1.04

150

1500

1.06

200

1300

Timeout Duration = 1500 slots (33msec)

1.08

250

30

1.1

300 Mean Queue Size

Mean Number of Transmissions

Timeout Duration = 1000 slots (22msec)

1200

Mean Queue Size at the transmitter in SRP for various values of Timeout duration (N=32, Input Load 45%)

Mean Number of transmissions per delivered cell in SRP (Input Load 45%) 1.12

900

Figure 6. Mean transfer delay in SRP for various values of Timeout duration

Figure 4. Mean queue size at the transmitter in SRP

1.14

600

Timout Duration

1100

12 16 20 24 28 32 Window Size N

900

8

0

Timeout Duration = 2000 slots (44msec)

700

0

20

500

500

40

300

1000

60

300

1500

80

90

Timeout Duration = 1500 slots (33msec)

100

100

2000

120

30

Mean Queue Size

2500

Mean Delay in SRP for various values of Timout duration (N=32, Input Load 45%)

70

Timeout Duration = 1000 slots (22msec)

Mean Queue Size in SRP (Input Load 45%)

Timeout duration also lead to greater delays and queue size since, in this case, the time interval allowed for the receiver to provide the feedback is not large enough, so the transmitter may unnecessarily retransmit some cells that have already been delivered successfully, instead of transmitting new cells.

Mean Delay (in msec)

result in sparse retransmissions of the not acknowledged cells, thus the number of transmission attempts is reduced. Another explanation of this effect can be derived from the behaviour of the channel model. When in a state with high CER and BER, the model tends to transit to a better (in terms of BER and CER) state. Thus, by increasing the timeout duration, the probability of a successful transmission (since the channel would be in a better state) is also increased, so the mean number of transmission attempts decreases.

Timeout Duration

Figure 5.Mean number of transmission attempts for SRP

Figure 7. Mean queue size at the transmitter in SRP for various values of Timeout duration

In figures 6 and 7, the mean transfer delay and the mean queue size attained by SRP for various values of Timeout duration are presented. As it can be seen, the two curves are quite similar and the mean queue size is almost proportional to the mean transfer delay. It appears that for window size N=32 and input load 45% the optimum timeout duration is around 100 slots (=2.2 msec). Higher values for Timeout duration lead to greater delays and queue size since the cells have to wait longer in the buffer before they are retransmitted. Shorter values of the

The figures 8, 9 and 10 present the comparison of SRP and PRIME ARQ. Figure 8 illustrates the comparative results regarding mean transfer delay. It can be seen, that although all variations of PRIME ARQ perform worse than SRP, the difference is not substantial. It seems that the limited feedback does not affect the efficiency of the PRIME ARQ considerably even in the case where feedback is really constrained (case of NTSN = 4 in the diagram).

Comparison of Mean Delay for SRP/PRIME ARQ (Timeout 33msec) Mean Delay (ms)

SRP, N=16 230 PRIME ARQ, N(MSN)=16, N(TSN) = 12

180

PRIME ARQ, N(MSN) = 16, N(TSN) = 8

130 80 11%

22% 33% 45% Input Load

56%

PRIME ARQ, N(MSN) = 16, N(TSN) = 4

Figure 8. Mean Delay for SRP and PRIME ARQ

Mean Queue Size

Comparison of Mean Queue Size for SRP/PRIME ARQ(Timeout 33msec)

650 600 550 500 450 400 350 300 250 200 11%

SRP, N=16

22% 33% 45% Input Load

56%

PRIME ARQ, N(MSN)=16, N(TSN) = 12 PRIME ARQ, N(MSN) = 16, N(TSN) = 8 PRIME ARQ, N(MSN) = 16, N(TSN) = 4

Figure 9. Mean Queue Size for SRP and PRIME ARQ

Comparison of Mean Number of transmissions per delivered cell for SRP/ PRIME ARQ (Timeout 33msec)

1.12

SRP, N=16 1.11 Mean Number of Transmissions

The same observation applies for figures 9 and 10 where SRP and PRIME ARQ are compared in terms of mean queue size at the transmitter and mean number of transmission attempts. Again the performance of PRIME ARQ although worse, does not differ substantially from this of SRP. These observations lead to the conclusion that PRIME ARQ combines well the positive characteristics of SRP and GBN and should be considered as a very good alternative for WAND system.

1.1

PRIME ARQ, N(MSN)=16, N(TSN) = 12

1.09 1.08

PRIME ARQ, N(MSN) = 16, N(TSN) = 8

1.07 1.06 11%

22% 33% 45% Input Load

56%

PRIME ARQ, N(MSN) = 16, N(TSN) = 4

Figure 10. Mean Number of Transmission Attempts for SRP and PRIME ARQ 4. CONCLUSIONS Two alternative ARQ schemes, SRP and PRIME ARQ, for WAND WDLC layer were presented. Both alternatives were evaluated and compared through simulation studies which indicate that SRP and PRIME ARQ should be considered as valid alternatives for WAND system. SRP, although more complicated, appears to perform much better than GBN, the initial choice for WAND WDLC layer, while PRIME ARQ seems to combine very effectively the advantages of GBN and SRP and attain performance comparable to the performance of SRP. REFERENCES [1] J.Mikkonen et al., "The Magic WAND Functional Overview", IEEE JSAC, Vol.16, No.6, August 1998. [2] Deliverable 3D4 of Magic WAND (AC085),“Wireless ATM MAC Results”, available at: http://www.tik.ee.ethz.ch/~wand/ [3] N.Passas, L.Merakos, D.Skyrianoglou, F.Bauchot, G.Marmigere and S.Decrauzat, “MAC Protocol and Traffic Scheduling for Wireless ATM Networks”, ACM Mobile Networks and Applications Journal, special issue on Wireless LANs, 1998. [4] J.Meierhofer, “Data Link Control for Indoor Wireless ATM Networks”, in Proc. Wireless ’98, Calgary, Canada, July 1998. [5] D.Bertsekas and R.Gallager, "Data Networks", Englewood Cliffs, NJ: Prentice Hall, 2nd ed., 1992. [6] S.Lin, D.J.Costello and M.J.Miller, “Automatic Repeat Request error-control Schemes”, IEEE

Communication Magazine, vol.22, pp. 5-17, December 1984. [7] Ohta et al., "PRIME-ARQ: A Novel ARQ Scheme for High-speed Wireless ATM Design, implementation and performance evaluation-", in Proc. of IEEE VTC '98, vol. II, pp. 1128-1134, May 1998. [8] OPNET Modeler, MIL 3 Inc. 3400 International Drive NW, Washington, DC 20008, 1993.

[9] J.Meierhofer, U.P.Bernhard, and T.Hunziker, “Finite State Radio Channel Model for Indoor Wireless ATM Networks”, in Proc. of IEEE ICT ’98, vol. IV, pp.39-43, Chalkidiki, Greece, June 1998. [10] Blondia and O. Casals, “Performance Analysis of Statistical Multiplexing of VBR Sources”, in Proc. INFOCOM ‘92, pp. 828-838, 1992.