Implementation Aspects of Reliable Transport Protocols in Wireless Sensor Networks

Transcription

1 Implementation Aspects of Reliable Transport Protocols in Wireless Sensor Networks Tuan Le, A. Y. Dong, R. Liu, Sanjay Jha and Z. Rosberg CSIRO ICT Centre, Australia School of Computer Science and Engineering, UNSW, Australia, Abstract Our previous study comparing analytically a comprehensive set of reliable data streaming protocols has shown that an hybrid protocol comprising stop-and-wait hop-by-hop ARQ with implicit ACK (SW HBH iack ARQ) and event-tosink reliable transport (ESRT) is the best energy-efficient generic transport protocol for wireless sensor networks. The implementation of this hybrid protocol requires better understanding of when to switch between ESRT and SW HBH iack ARQ, how to set the retransmission timeout and how to adapt the protocol parameters to the link qualities. Theses questions are studied by field experiments and simulation and the findings are reported in this paper. I. INTRODUCTION Generally, the reliability requirement of data streaming applications in wireless sensor network (WSN) is not absolute but rather statistical. That is, the reliability is determined by an ensemble of sensed data delivered at the sink node rather than the reliability of each sensed data. Such statistical reliability objective, defined in [7], guarantees adequate sensed data delivery while preserving energy otherwise used for absolute reliability. Consequently, the lifespan of unattended WSN is increased. We have analyzed and compared the energy-efficiency of a comprehensive set of statistically reliable transport protocols [7] from which an hybrid protocol comprising stop-and-wait (SW) hop-by-hop (HBH) ARQ with implicit ACK (iack) and event-to-sink reliable transport (ESRT) emerges as a very attractive generic transport protocol for wireless sensor networks (WSN). The protocol is simple to implement, does not require large buffers and is very energy efficient. However, several implementation aspects of the hybrid protocol could not be addressed by our analysis and requires field experiments and detailed simulations. Some of these aspects addressed in this paper are: when to switch between ESRT and SW HBH iack ARQ, how to set the retransmission timeout and how to adapts the protocol parameters to the link qualities. Our theoretical model in [7] assumes that link qualities are available and that packets are forwarded without any queueing delay. In practice, link qualities should be estimated and their variation in time and location may have a major effect on the implementation. Additionally, with implicit ACK, the time to overhear an ACK depends on the forward queue length. Theses questions are studied by field experiments and simulation and the findings are reported in this paper. Related studies are described in Section II. The field experiments and their finding about link qualities are given in Sections III and IV. The simulation study on setting the retransmission timeout is given Section V. II. RELATED WORKS A. Event-to-Sink Reliable Transport A statistically reliable transport protocol called event-to-sink reliable transport (ESRT) is presented in [1] for sensed data streaming applications not requiring absolute reliability. This protocol achieves statistically reliable delivery by controlling the amount of sensed data transmitted at the sources, avoiding acknowledgments (ACKs) and retransmissions all together. B. ARQ Protocols Classical ARQ protocols with absolute reliability can be divided into three basic schemes: send-and-wait (SW), Go- Back-N (GBN) and selective repeat (SR) [3] [4] [5] [6] [11]. In [7], the following statistical reliable variants of these protocols limiting the number of retransmission so as to achieve a confidence level have defined and analyzed. 1) SW HBH ACK ARQ: With SW HBH ACK ARQ, reliability is assured hop-by-hop. When a receiver receives a packet from a transmitter, it sends an ACK to the transmitter. The transmitter retransmits a packet if a preset timer timeouts before an ACK is received. Due to the relatively high error rates of WSN links, SW HBH ACK ARQ [8] appears as an attractive scheme. Considering the energy used for ACK transmissions, using an implicit ACK (iack) rather than an explicit ACK, is expected to be even more attractive. 2) SW HBH iack ARQ: Native SW HBH ACK uses explicit ACK messages which consumes energy. With wireless links, the transmitter can overhear the forwarding transmission and interpret it as an implicit ACK. Obviously, when a packet reaches the sink, there will be no further forwarding.

2 Thus, the packet cannot be implicitly acknowledged. The sink node therefore needs to send an explicit acknowledgment. Retransmissions are as with explicit ACKs. That is, the transmitter retransmits the packet if a timeout fires. This ARQ version, referred to as SW HBH iack ARQ, has been proposed in [9]. If the qualities of next upstream and next downstream links are highly correlated, the energy saving with SW HBH iack ARQ is dramatic since ACKs are almost free. A potential issue may rise when using an improper retransmission timeout. Unlike with explicit ACK where the ACK is sent immediately upon packet reception, an iack is sent only when the packet is being forward. While the expected timeout with explicit ACK is fixed, the expected timeout with iack depends on the forward queue length and therefore variable. Variable expected timeout may introduce redundant retransmissions or excessive packet delay. Another potential issue may rise in a scenario where node i 1 sends a packet which is received successfully at node i and has been forwarded to node i + 1 but node i 1 did not overhear the iack. In such case, node i 1 is retransmitting the packet and node i must re-forward it just for the sake of iack. The result is packet duplication requiring a packet sequence number management which is more complicated than with explicit ACK. 3) GBN and SR ARQ with Bulk ACKs: Unlike SW, GBN and SR ARQ use continuous transmission to improve packet delay. To track losses, both protocols use sequence numbers (SN) to label the sensed data packets. With SR ARQ, each transmitter transmits all available data constantly without waiting for any ACK. The receiver, uses negative acknowledgements (NACKs) indicating the SN of the missing packets to request retransmissions. When a NACK arrives, the source retransmits only the missing packets. With GBN ARQ, each transmitter transmits continuously all available data using a sliding window of a certain size. To save energy, the receiver sends a single ACK/NACK packet for every batch of K packets received successfully. It has been shown in [7] that the HBH versions of GBN and SR ARQ outperform the end-to-end versions. Additionally, among all SW protocols, SR, GBN and ESRT, SR HBH ARQ is the most energy efficient protocol. However, it may not suit most current WSNs due to its buffer limitations, e.g., MICA2 which has only 4 KB [12]. 4) Hybrid ARQ: Traditional ARQ protocols bound the maximum number of retransmissions by a fixed number. Our previous works [7] has revealed that energy-wise, an hybrid protocol combining ESRT and SW HBH iack ARQ in a manner depending on the path length and the link quality outperforms all other protocols mentioned above. In light of its optimality, addressing the issues of SW HBH iack ARQ iack mentioned above become more acute. III. WSN LINK QUALITIES - AN EXPERIMENTAL STUDY In this section, we report and discuss the findings on WSN link qualities revealed by our experiments using SW HBH iack ARQ in MICAz motes. A. High Sensitivity to Mote Location Fig. 1. Experiment Topology 1 In our experiment, we deployed 11 equally spaced MICAz motes running B-MAC at distances of 10 meters apart as shown in Figure 1. To reduce random measurement errors, we placed two motes at each distance. To avoid the ground reflection on the radio transmission, each mote was elevated one meter above the ground. Mote 1 was implemented as the source sending a total of 1000 packets at a rate of two packets per second. The other 10 motes were implemented as receivers logging the sequence number and the RSSI (received signal strength indication) of packets received from mote 1. Table I shows the results of our first experiment. The results show that the measured RSSIs fall off as distances increase. However, the packet reception rates do not fall off monotonically with distance. Mote 4 (at 20 meters) received only 64 packets, while mote 5, at the same distance, received 945 packets although their RSSI were about the same. Similarly results are observed with motes 8 and 9 at distance 40 meters. At 50 meters, mote 11 did not receive any packet, so the RSSI is nil. We are currently experimenting with mechanism that is able to extract RSSI on corrupted packets. The observation from this experiment is that the RSSI is generally consistent with distance. However, the path loss ratio could be quite different even though the RSSI is the same. The inconsistence could happen because of many factors, such as mote malfunctioning, impairments (shadowing, fast fading) and location-sensitive interference of other wireless devices in our experimental field. As the results from motes 4 and 9 are particularly suspicious. To eliminate mote malfunction as the reason of inconsistence, we repeated the first experiment after swapping some of the motes as depicted in the Figure 2. It is noted that it was raining during the repeated experiment and we stopped it after sending 320 packets. The results of the second experiment shown in Table II indicate that mote 4, at a distance of 10 meters, received all the packets. Mote 9, now at a distance of 30 meters, also received

3 TABLE I MICAZ MOTES EXPERIMENT RESULT 1 Distance 10m 20m 30m 40m 50m Received Packets (M2) 1000 (M3) 1000 (M4) 64 (M5) 965 (M6) 985 (M7) 995 (M8) 953 (M9) 32 (M10) 212 (M11) Null RSSI(db) (M2) -81 (M3) -80 (M4) -92 (M5) -88 (M6) -89 (M7) -89 (M8) -91 (M9) -93 (M10) -93 (M11) Null Optimal Received Packets Error rate 0 5.5% 1% 4.7% 78.8% Optimal RSSI TABLE II MICAZ MOTES EXPERIMENT RESULT 2 Distance 10m 20m 30m 40m 50m Received Packets (M4) 320 (M3) 320 (M10) Null (M11) 288 (M8) 288 (M9) 320 (M6) 96 (M7) 32 (M10) Null (M11) Null RSSI(db) (M4) -85 (M3) -81 (M10) Null (M11) -91 (M8) -90 (M9) -89 (M6) -91 (M7) -91 (M10) Null (M11) Null Optimal Received Packets Error rate 0 10% 5% 20% 100% Optimal RSSI Null Fig. 2. Experiment Topology 2 all the packets. These results eliminated the suspicion of mote malfunction. However at distances of 20 and 40 meters, we again observed bad reception with different motes this time. At distance of 50 meters, both motes did not receive any packet. Observations: The RSSI is not the only decisive factor to determine radio link quality. In our experiments, we observed that two motes with similar RSSI measures can have significantly different reception rates. Further experiment results eliminated mote malfunctioning factor. The reception quality is very sensitive to the specific mote location. B. Temporal Link Quality Both above experiments suggest that within meters the error rate increases dramatically. In order to identify the source of this link quality variation, we conducted the third experiments. In this experiment, the receiver motes were placed two meters apart at distances of 40 to 48 meters from the transmitter motes. In order to make sure the result veracity, we still placed two motes at the same distance. The results are shown in Table III. We observed that all the motes (except mote 4) received most of the transmitted packets. Mote 4 at distance of 42 meters received none, while its co-located mote (mote 5) received 965 out of These results are totally different from the measurements in our previous experiments. We suspect besides being location-dependent; reception quality could also be time-dependent. Therefore, we conducted the following experiment at different times on the same day. The experiment topology is depicted in Figure 3. The first mote was placed 20 meters from the source and all others at 5 meters apart. Since we had only 11 motes, a single mote is placed at each distance so as to experiment with more locations. The results of the experiment are shown in Table IV. The measured results were quite unusual: only the motes at distances of 20 and 65 meters received packets; all other motes received nil. Consequently, we repeated the experiment at different time, whose results are given in Table V. Unexpectedly, the findings have changed again. At distances of 25, 35, 40, 50, and 65 meters, no packets were received. Meanwhile motes at distances of 20, 45, 55 and 60 meters had mediocre reception rates. The mote at distance of 30 meters is the only mote with reasonable reception. An amazing fact was that all the motes had similar measured RSSI. Observations: The link quality is not only locationdependent but also time varying. Once again link quality is shown to be inconsistent with RSSI which contradicts to the literature. We suspect the irregular measurement we observed was the results of interferences from WLAN which operate at the same IMS frequency band of 2.4 GHz. Therefore, to account for inconsistent reception in time and distance (i.e., incompatible to classical channel models), a transport protocol should be adaptive and capable of tuning itself to sporadic link quality changes. Moreover, the adaption should be based on realtime channel estimation. This conclusion is applied in our experiments with various transport protocols described in the next section. IV. EXPERIMENTS WITH MICA2 In order to test the hypothesis of WiFi interference, a set of experiments were conducted using MICA2 motes operating at 916 MHz, away from WLAN frequency bands. The logic

4 TABLE III MICAZ MOTES EXPERIMENT RESULT 3 Distance 40m 42m 44m 46m 48m Received Packets (M2) 1000 (M3)1000 (M4) Null (M5) 965 (M6) 1000 (M7) 1000 (M8) 990 (M9) 990 (M10) 1000 (M11) 970 RSSI(db) (M2) -86 (M3) -86 (M4) Null (M5) -85 (M6) -88 (M7) -88 (M8) -90 (M9) -91 (M10) -89 (M11) -90 Optimal Received Packets Error rate % 1.5% Optimal RSSI TABLE IV MICAZ MOTES EXPERIMENT RESULT 4 Distance 20 m 25 m 30 m 35 m 40 m 45 m 50 m 55 m 60 m 65 m Received Packets 323 Null Null Null Null Null Null Null Null 909 RSSI(db) -92 Null Null Null Null Null Null Null Null -91 TABLE V MICAZ MOTES EXPERIMENT RESULT 5 Distance 20 m 25 m 30 m 35 m 40 m 45 m 50 m 55 m 60 m 65 m Received Packets 570 Null Null RSSI(db) -92 Null Null Fig. 3. Experiment Topology behind this experimental design is that: If the MICA2 results have the same irregular behaviour as MICAz, then there are some inherent problems with the radio. If we could get a stable measurement results on MICA2 in this WiFi free band, this would be strong evidence that MICAz measurement results were tampered by WiFi interference. The experiments were conducted in the open yard at CSIRO Marsfield site with five MICA2 motes. The motes were placed 20 meters apart from each other in a line and were elevated to one meter above the ground. A. Reception measurements The transmission power was set to 0vdBm and we first measured the loss rate at each hop by programming each mote to broadcast 1000 beacon packets to its neighbors, After one mote finished its schedule of 1000 packets, the next mote started broadcasting. Only one mote transmits data at one time, so there is no collision problem in this experiment. Each experiment was repeated three times. The number of received packets and error rate of the three experiments are given in Tables VI-VII. The average of the three experiments is used as the input parameter for the adaptive transport protocols in the following experiments. The average loss rate for each hop at each direction is shown in Figure 4. Figure 5 is the visualization of the environment packet loss ratio of each hop from which we observed the following. Loss rates at equal length links have high variation. Loss rates are not reciprocal. Fig. 4. Experiment Setup TABLE VI NUMBER OF RECEIVED PACKETS Direction Exp1 Exp2 Exp 3 average Loss rates are very sensitive to mote location - slight displacement may cause dramatic changes. Loss rates vary in time but not dramatically (unless TABLE VII PERCENTAGE OF ERROR RATE Direction Exp1 Exp2 Exp 3 average % 71.1% 55.5% 60.7% % 5.4% 2.7% 4.2% % 6.2% 7.5% 8% % 43.4% 32.7% 33.4% % 31.4% 13% 22.8% % 99.9% 97.2% 98.6% % 0.6% 0.6% 0.7% % 0.9% 0.8% 0.8%

5 WLANs transmit at the same band). Variation is mainly at high loss rates ( 20%), where the variation could be larger than 20%. Observation: MICA2 results are much more stable compared to MICAz results. This implies that MICAz results were heavily interfered by WLAN operating at the same 2.4 GHz band. We are currently studying this interference effects in more details, and designing efficient interference avoidance algorithms to tackle this problem. Fig. 5. Packet Loss Ratio V. RETRANSMISSION TIMEOUT - A SIMULATION STUDY In this section, we report and discuss our finding on the impact of setting the retransmission timeout in Implicit Sendand-Wait Hop-By-Hop Acknowledgment (SW HBH iack). The main approach taken in implicit ACK protocol is based on the nature of wireless broadcasting. With wireless links, the transmitter can overhear the forwarding transmission and interpret it as an Implicit Acknowledgment. For example, when a node i forwards the packet of its downstream node j (a child of node i in routing tree) to its upstream node k (the parent of i in routing tree), j can overhear this forwarding and consider as an acknowledgment. Obviously, when a packet reaches the sink, there will be no further forwarding. The sink node therefore, needs to send an explicit acknowledgment (eack). Recently, SW HBH iack transport protocol emerges as an attractive transport protocol for WSNs [7] since it is simple to implement, and is energy-efficient. However, the retransmission timeout has not been investigated yet. Large retransmission timeout values increase transmission delay, while premature retransmission timeout values cause redundant packet transmissions, and thus, not energy-efficiency. In this section, we conduct an extensive simulation studies on the impact of retransmission timeout setting on performance of the SW HBH iack protocol. A. Simulation setup The simulated topology comprises 10 equally spaced nodes, as depicted in Figure 6, where h = 9. For every node 0 i 8, the probabilities that a transmission from node i to node i + 1 and from node i + 1 to node i are received successfully are denoted by 1 p i and 1 q i, respectively. The simulations were executed in NS-2 with a modified version of MAC layer allowing overhearing for SW HBH implicit ACK purpose. To prevent the redundant retransmissions, the maximum number of retransmissions for a data packet at each node is set according to [7]. We use the destination-sequenced distance-vector (DSDV) routing protocol. Every node, except the sink, generates packets of size 40 bytes at a rate 0.5 per second and transmit them toward the sink via the shortest path. To prevent RTS/CTS message exchanges, we disabled the RTS/CTS option resulting in the hidden terminal problem [2]. We measured the following performance metrics: the number of transmissions, the average packet delay, and the average goodput. Let RTT denote the estimated single hop round trip time. RTT is taken as twice the estimated time for a packet to travel from a node to its subsequent hop. In the simulation we compare three cases of retransmission timeouts under various single link packet losses: 1xRTT, 5xRTT and 10xRTT. For simplicity, we set p i = q i = p. Each performance value in the figures is an average of 10 simulations. The averages are plotted along with their 95% confident intervals. B. Simulation results Number of Tranmissions Fig. 6. Topology Fig. 7. Number of transmissions, p = 0.01 Figures 7-8 show the total number of transmissions vs the path length for the various timeouts in the two extreme link error rate cases of, i.e. HBH loss rate p = 0.01 and

6 Number of Tranmissions Average packet delay Fig. 8. Number of transmissions, p = 0.60 Fig. 10. Average packet delay, p = Average packet delay Goodput Fig. 9. Average packet delay, p = 0.01 Fig. 11. Average Goodput, p = 0.01 p = The results suggest that a longer timeout is more energy-efficient than a short timeout. With low error rates (e.g., p = 0.01), a timeout of 10xRTT can save up to 14% in the number of transmission compared to 1xRTT (Figure 7). With high error rates (e.g., p = 0.60), the saving is up to 20% (Figure 8). The reason is quit clear, a timeout of 10xRTT saves unnecessary retransmissions. It is interesting to note that the number of transmissions with 10xRTT is just slightly higher than that with 5xRTT. Thus, a timeout of 5xRTT is sufficient to recover most of the transmission errors. The drawback of a long timeout is its effect on packet delay shown in Figures When the link error rate is p = 0.60, the average packet delay with 10xRTT is slightly higher than that with 5xRTT, and approximately 10 22% higher than that with 1xRTT (Figure 10). When the link error rate is p = 0.01, the average packet delay with 10xRTT is approximately 35 40% higher than that with 5xRTT, and approximately 35 55% higher than with 1xRTT (Figure 9). Finally, Figures show the average goodput vs path length. Although the goodput fluctuates, the differences for the various timeouts are relatively small (about ± 3%). Generally, the goodput with 1xRTT is slightly better than the others. Conclusion: Our simulation results suggest that there is a trade-off between retransmission timer setting and packet delay. A large timeout value of retransmission timers tends to be energy-efficiency, but increases packet delivery delay, and vice versa. Additionally, if packet delay is not an issue, the goodput results suggest that using a timeout of 1xRTT is best.

7 Goodput Fig. 12. Average Goodput, p = 0.60 ACKNOWLEDGMENT The authors would like to thank Carol Wilson from CSIRO for her advice on radio propagation and interference. REFERENCES [1] O. Akan and I. Akyildiz, Event-to-Sink Reliable Transport in Wireless Sensor Networks, IEEE/ACM Transactions on Networking, vol. 13, no. 5, pp , Oct [2] V. Bharghavan, A. Demers, S. Shenker, and L. Zhang, MACAW: A Media-Access Protocol for Packet Radio, Proc. ACM SIGCOMM, London, England, Aug [3] D. Bertsekas and R. Gallager, Data Networks, Englewood Cliffs, NJ: Prentice-Hall, [4] H. Burton and D. Sullivan, Error and Error Control, Proceeding of IEEE, vol. 60, pp , Nov [5] S. Lin, D. Costello and M. Miller, Automatic-repeat-request error-control schemes, IEEE Communications Magazine, vol. 22, n. 12, pp. 5 17, Dec [6] Z. Rosberg and M. Side, Selective-Repeat ARQ: The Joint Distribution of the Transmitter and the Receiver Resequencing Buffer Occupancies, IEEE Trans. on Commun., vol. 38, no. 9, pp , Sept [7] Z. Rosberg, R. Liu, L. D. Tuan, S. Jha, A. Y. Dong and J. Zic, Energy Efficient Statistically Reliable Hybrid Transport Protocol for Sensed Data Streaming, CSIRO ICT Centre Pub. no. 07/213, June Available at: [8] R. Stann and J. Heidemann, RMST: reliable data transport in sensor networks, Proceedings of the First IEEE International Workshop on Sensor Network Protocols and Applications, Anchorage, Alaska, pp , May [9] A. Woo and D. Cellur, A Transmission Control Scheme for Media Access in Sensor Networks, Proceeding of ACM MobiComm 01, July [10] Avoiding RF Interference Between WiFi and Zigbee, Product pdf files/wireless pdf/zigbeeandwifiinterference.pdf [11] M. Yoshimoto, T. Takine, Y. Takahashi and T. Hasegawa, Waiting time and queue length distributions for Go-Back-N and selective-repeat ARQ protocols, IEEE Transactions on Communications, vol. 41, no. 11, pp , Nov [12] MICA2,