Sensor Node Lifetime: An Experimental Study - CiteSeerX

Sensor Node Lifetime: An Experimental Study Hoang Anh Nguyen, Anna F¨orster, Daniele Puccinelli, Silvia Giordano Networking Laboratory University of Applied Sciences of Southern Switzerland CH-6928 Manno, Switzerland {hoang.nguyen, anna.foerster, daniele.puccinelli, silvia.giordano}@supsi.ch

Abstract—Node lifetime is a key performance metric in wireless sensor network (WSN) research. Simplistic assumptions and na¨ıve lifetime estimation techniques invariably prove to be extremely unreliable in practice, to the point that premature battery depletion notoriously affects real-world deployments. In this paper we adopt an experimental approach and employ various types of real-world batteries to determine the actual lifespan of a sensor node under common operating conditions. We present a rich set of results from an extensive experimental campaign based on the widely used TelosB platform running TinyOS. We have measured the actual node lifetime using various brands of commercial batteries as a function of different combinations of application parameters. Some of our observations match previously published results that are often neglected, while others underscore less known properties of low-power radios. Keywords-wireless sensor networks, lifetime, experiments, TinyOS, TelosB

I. I NTRODUCTION The lifetime of a sensor node begins when it first boots and ends when it is no longer able to communicate or perform its other basic tasks. Network lifetime can be defined in many different ways [1], but it ultimately depends on the lifetime of the individual nodes. While everyone in the sensor network community agrees that sensor node lifetime is a key metric, preliminary studies both in simulation and on real hardware use calculations based on datasheet for the expected network lifetime. Such predictions are usually based on basic assumptions about duty cycle and power consumption of sleep/receive/transmit states of the onboard radio and fail to take into account the non-linear behavior of the batteries, their large performance spread, or the power consumption of processing, flash writing, sensing, etc. For example, the energy cost of writing to Flash on sensor nodes like the popular TelosB platform is often ignored or considered as less significant than the energy cost of a radio operation. In this paper we present the results of an experimental study of the lifetime of real nodes with various application parameters. We focus on duty-cycled radio operation and Flash usage across different battery brands. Specifically, we focus on high duty-cycles. Our data provides a deep insight into the battery consumption of real world sensor nodes and batteries. For example, we measured that power consumption

of flash writing might be comparable to radio transmission in some application scenarios and that different battery brands can affect the total node lifetime as much as 25%. The paper continues with a description of related efforts in Section II. Section III explains our experimental methodology, and Section IV presents our experimental results. II. R ELATED WORK The extension of battery lifetime is an omnipresent theme in sensor network research and permeates all levels of protocol design. Indeed, premature battery depletion notoriously affects real-world deployments [2], which are designed with the goal of long-time untethered operation. In spite of its importance, it is typically estimated by extrapolating from energy consumption measurements or datasheet figures. While it is not uncommon for sensor network lifetime studies to assume a linear battery discharge, the presence of non-linearities was already underscored a decade ago in [3], where it is shown that the discharge rate, the duty-cycling, and the DC-to-DC converter affect the usable battery capacity. While the DC-to-DC converter disappeared from later sensing node hardware, the other factors pinpointed in [3] continued and will continue to affect the behavior of sensing nodes. The impact of various parameters on battery performance was studied experimentally in [4] using a testbed of MICA2DOT motes using measurements on real batteries along with other techniques. The focus of [4] is the impact of the electrochemical phenomena of batteries on sensor network design parameters; it is shown that the rate capacity effect (a stronger discharge current reduces the battery capacity), the charge recovery effect (if the battery is occasionally allowed to idle, its capacity increases), and thermal affects (operating at higher temperatures increases the battery capacity but reduces the battery cycle life) all have non-trivial implications on higher-level decisions, such as power control and adaptive sampling. The importance of a realistic battery model that accounts for non-linearities and its impact on simulation have also been studied in [5]. Recently, the non-linear phenomena that occur inside batteries have been the subject of two notable studies. In [6], the battery recovery effect is investigated in detail, while [7] adapts a pre-existing model for the estimation of the remaining battery lifetime of a node battery that accounts

for both the recovery and the rate capacity effect so that it can be implemented on a low-end mote-class device. Despite the overriding importance of battery lifetime, there are only a few experimental sensor network studies that employ real batteries and report discharge curves [8] [9]. One of the reasons for the lack of experimental data obtained using real batteries is the relatively long time it takes to drain them. A novel hardware measurement methodology based on using special capacitors with a very high capacity, the so-called Goldcaps, was developed in [10]: a Goldcap is charged and the time a node can live off of it is measured, so that the energy consumption of sensor network protocols can be realistically evaluated within a few hours. Because it enables deterministic repeatable experiments, this method was also employed by some later studies such as [11] and [12]; it is certainly very valuable, but it abstracts away from the real-life phenomena underscored in [3] and [4], which seriously affect real-world deployments where nodes do run on real-world batteries. At the same time, over the past few years, experimental sensor network research has leveraged the existence of hugely helpful remote-access testbeds such as Harvard’s MoteLab [13] and TU Berlin’s TWIST [14], where nodes are purposefully wall-powered. Working on such testbeds allows researchers to isolate out the effect of batteries, but also forces them to estimate node lifetime based on techniques such as online software energy estimation [15], which on one hand give a reasonable ballpark estimate, but on the other hand abstract away from the non-linearities of real-batteries. The focus of the present study is indeed to determine the extent to which the vagaries of real-world batteries affect the operation of common sensing node hardware as well as the common assumptions and expectations on the behavior of batteries. III. E XPERIMENTAL M ETHODOLOGY We completed a rich set of experiments using moteclass hardware powered by commercial off-the-shelf AA batteries. The main goal of our study is to identify the impact of typical WSN application scenarios on the battery lifetime. Generally speaking, our methodology consists of programming sensor nodes with an application and letting them operate until their energy reserve is depleted. During its lifetime, each node logs data on its on-board flash memory, which can be retrieved and parsed after the completion of each experiment. We use TelosB motes programmed using TinyOS1 . The nominal power consumption data of TelosB nodes is provided in Table I and was taken from the TelosB data sheet and the CC2420 data sheet. The transmission power in our experiments is set to 0 dBm, the maximum setting of the CC2420 radio. 1 www.tinyos.net

Component Module RF Transceiver

Mode

Current draw

Active Sleep Receive Transmit (at 0 dBm) Sleep

1.8 mA 5.1 µA 19.7 mA 17.4 mA 0.01 mA

Data from www.memsic.com and www.ti.com Table I N OMINAL POWER CONSUMPTION OF T ELOS B NODES .

The general application scenario is depicted in Figure 3. It is based on the most widely used WSN application, where the node gathers environmental data at regular intervals and transmits them to a base station, sleeping between any two data gatherings (duty-cycled data collection). In our case, we have 4 parameters that control the exact behavior of the node. D represents the duty cycle of the node, S is the number of packets the node sends to the base station at every cycle, while R is the number of packets it receives from the base station during every cycle. Finally, C is the frequency of logging data on the on-board flash memory, meaning that the node logs data every C-th cycle. By controlling these parameters we are able to capture the impact of all radio modi (sleeping, receiving and transmitting), as well as the impact of using the on-board flash memory. We conducted a total of 6 different experiments, whose parameters are summarized in Table II. Additionally, we have used various brands of AA batteries in order to capture the impact of nominal capacity of different battery models. The batteries we have used are listed in Table III with their nominal and measured capacities. The measured capacities are taken from an online battery capacity survey2 . Each of the experiments in Table II is conducted with various brands of batteries (using three nodes for each brand). All batteries are new and unused, but no details about their history are known (e.g. how long they have been stored in the shop). IV. E XPERIMENTAL R ESULTS Our experiments serve to underscore many interesting properties of real sensor nodes, some of which defy common expectations. In this section we will begin with the less surprising results, obtained in the absence of communication, and then move on to examine the results obtained with active radio communication. The reported node lifetime is the time between the power-up and the last logging (in the absence of communication) or the reception of the last packet (with active radio communication). We refer to nodes that operate in the absence of communication as silent nodes, and to the 2 http://www.laurinieminen.com

S.1 S.2 S.3 S.4 C.1 C.3

Parameters D [%] S R duty cycle sent received Silent experiments 100 0 0 75 0 0 75 0 0 25 0 0 Communication experiments 75 1 50 75 50 1

Node lifetime with 100 % duty cycle over initial capacity of batteries 100

C log

95

1 1 30 1

Lifetime [h]

Experiment

Energizer Ultra+

90 Panasonic Extreme Power

85 80 75

30 30

Panasonic Industrial

70 2500

2600

Table II E XPERIMENTAL PARAMETER SETTINGS FOR OUR STUDY

Node lifetime over duty cycle

(100-D)% from full cycle

180

time

160 Lifetime [h]

switch off radio

140 Energizer Ultra+ Ansmann Varta High Energy

120

5000 msec send 0 .. S packets to base station receive 0 .. R packets from base station

Figure 1.

3000

200

beginning of next cycle

beginning of cycle switch on radio

2700 2800 2900 Nominal capacity of batteries, [mAh]

Figure 2. Node lifetime measured for different values of the nominal battery capacity for various alkaline battery brands.

every C-th cycle, write 8 bytes to flash log D% from full cycle

Varta High Energy

Ansmann

100 80 100

75 Duty cycle in percent (100% == no sleep)

25

One cycle of the sensor node lifetime. Figure 3. node.

experiments that run in the absence of communication as silent experiments. A. Battery properties and brands Observation 1. There is a significant difference in node lifetime depending on the battery brand and nominal capacity (up to 25% in our experiments). The use of five different alkaline battery brands (listed in Table III) enables us to study the impact of different battery brands as well as different values of nominal capacity on the Brand name

Capacity [mAh] Measured Measured at 15Ω* at 3Ω* Energizer Ultra+ 2950 2439 1624 Panasonic Industrial 2200 2166 1264 Panasonic Extreme Power n/a 2559 1444 Varta High Energy 2850 2545 1682 Ansmann Alkaline 2400 n/a n/a *Data taken from http://www.laurinieminen.com Nominal

Table III P ROPERTIES OF BATTERIES , AS USED IN OUR LIFETIME STUDY. A LL BATTERIES ARE OF TYPE AA (LR6)

Node lifetime at different values of the duty cycle of a silent

lifetime and reliability of a sensor node. Figure 2 presents the total lifetime of the nodes with 100% duty cycle (no sleep, radio is always on, no communication) for different values of the nominal capacity of the batteries. As can be seen in Figure 2, the relationship between the nominal capacity of the battery and the lifetime of the node is roughly linear. Further information about measuring capacity of batteries can be found at http://www.laurinieminen.com. Figure 2 clearly shows that the choice of the battery is crucial. Many brands of batteries are sold without clear identification of their nominal capacity or with vague statements such as ”high capacity”. WSN practitioners need to be aware of the fact that the choice of the battery can increase (or decrease) the lifetime of their deployments by as much as 25%, according to our results and could be even much more in other scenarios. B. Radio Duty-cycling Observation 2. Lower duty cycles of the radio can increase significantly the node lifetime. However, the increase is not linear and is up to 250% lower than expected. As common sense dictates, radio duty-cycling can sig-

Comparison between real world and theoretical lifetime 500

real world datasheet expected alwaysïon projected

Lifetime [h]

400 300 200 100 25

Figure 4.

75 Duty cycle [%]

100

Comparison of node lifetime: measured against projected. Impact of flash and radio on node lifetime Flash seldom, RX

450

Lifetime [h]

400 350 300 250

Flash seldom, no comm

200 150

S.2

Figure 5. lifetime.

Flash seldom, TX

Flash often, no comm S.3

C.1

C.2

The impact of radio communication on the total sensor node

nificantly reduce the energy consumption and can therefore increase a node’s lifetime, but not as significantly as one would expect. Figure 3 presents the results of our silent experiments as a function of the duty cycle of the radio. Figure 4, on the other hand, compares the real world experimental values with the expected values. The datasheet expected curve represents a calculation of the expected node lifetime on the basis of the power consumption data from Table I. The always-on projected curve, on the other hand, uses the data of our always-on experiment as a baseline to project the network lifetime in the presence of radio dutycycling (assuming a linear lifetime increase). A thorough understanding of the reasons for the significant discrepancy between the measured values and the expected ones is beyond the scope of this paper, but we do wish to underscore how large the discrepancy is. C. On-board flash memory Observation 3. Flash memory energy consumption is very high and may be comparable to the power consumption of the radio. Our results point out that the energy consumption of the on-board flash memory (logging) is significant on the TelosB hardware platform and might exceed the energy consumption of the radio, depending on the exact application

scenario. The reason for this is the used NOR flash memory on the TelosB nodes, which consumes more power than other technologies, e.g. NAND. Details about power consumption of flash memories are given in [16]. Figure 5 compares the achieved values of node lifetime in the presence of flash and radio usage. The data is given as [mean, min, max] values across all the battery brands that we employed. As shown in Table II, S.2 and S.3 are silent experiments with 75% duty cycle, and the only difference between them lies in the frequency of write-toflash operations. Indeed, writing to the flash memory rarely instead of every cycle is able to prolong the network lifetime as much as 300%. Compare this result to experiment C.2 (same Figure 5). Surprisingly, it is more power efficient to use the flash only sparingly and send out 50 data packets per cycle than to use the flash every cycle and not send out any packets. Consequently, our experimental evidence shows that the radio is not the only significant energy-consumer on board and that the power consumption of the flash might be comparable to that of the radio. However, other researchers come to different conclusions with the same hardware platform [17]. We plan to investigate the reasons for these differences in the future. D. Radio communication Observation 4. Even if radio transmission uses less power than radio reception, it uses up more energy. Although we have just showed that it is not completely true that the radio is the component with the largest impact on the overall energy consumption of a sensor node, the radio does of course have a significant impact on the total node lifetime, as can be seen in Figure 5. Recall that experiments C.1 and C.2 have both 75% duty cycle and log data on the flash memory every 30th cycle. The only difference between them is that C.1 focuses on receiving packets and C.2 on transmitting packets. Since the power consumption of radio transmission is lower, we are expecting the C.2 experiments to obtain longer lifetimes, which is obviously not the case. As shown in Figure 6(e) of [18], with the CC2420, the current draw of radio transmission is lower than the one of radio reception, but the overall energy consumption of transmission is higher. E. Half-dead sensor nodes Observation 5. Towards the end of their lifetime, nodes are still able to transmit data to others, but are not able to receive anything. The duration of this half-dead status may account for as 27 % of the total lifetime. We have just noted that radio transmission has a lower current draw than radio reception, but a higher energy consumption. An interesting implication is that, towards the end of the battery lifetime, a node may still be able to transmit, but may not be able to receive. This behavior is

Base station data (wall powered) 120 0.9

ï40

110

ï50

0.6 0.5 0.4 0.3

90 80

ï60 ï70

70

0.2

ï80

60

0.1 120

RSSI [dBm]

100

0.7

LQI [50ï110]

Delivery rate [%]

0.8

135

150 165 180 195 Hours from experiment start

50 120

210

135


ï90 120

210

135


210

135 150 165 180 195 Hours from experiment start

210

Node data (battery powered) 1

120

0.9

110

0.5 0.4 0.3

RSSI [dBm]

0.6

90 80

ï60 ï70

70

0.2

ï80

60

0.1

Figure 6.

ï50

100

0.7

LQI [50ï110]

Delivery rate [%]

0.8

0 120

ï40

135


210

50 120

135


210

ï90 120

Delivery rate, RSSI and LQI values of packets at the wall-powered base station (top set) and the battery-powered node (bottom set).

clearly seen in Figure 6, where the reception of packets by the battery-powered node and by the wall-powered base station is depicted in terms of delivery rate, Received Signal Strength Indicator (RSSI), and Link Quality Indicator (LQI) values. The node (the lower set of graphs) is able to send and receive data packets for approximately 156 hours, after that the base station continues receiving packets from the node for a total of 199 hours. F. RSSI and LQI Observation 6. The RSSI and the LQI levels become unstable shortly before the depletion of a node’s energy reserve. Figure 6 shows also the RSSI and LQI values of the sensor node, as provided by the CC2420 radio interface in TinyOS. The RSSI and LQI values of the packets received at the wall-powered base station from the battery-powered node are relatively stable throughout the lifetime of the node but degrade significantly towards its death. We conjecture that this is due to the degradation of the transmit power of the node induced by battery depletion. Although it is reasonable to expect some sort of degradation, we did not expect its extent to be so significant. V. C ONCLUSION AND F UTURE W ORK Our rich set of experimental results obtained with TelosB motes powered by AA batteries has enabled us to point out several observations, which we can summarize as follows.

1) WSN practitioners should exercise caution in the choice of their batteries, especially at deployment time. Our results indicate that batteries with a higher nominal capacity do result in a significantly longer lifetime. 2) It is well-known that duty-cycling is essential for saving energy, but its effects are significantly less dramatic than what one would expect, and its benefits wane as the duty-cycled is lowered. Our future work will focus on experiments at very low duty-cycle to further investigate this observation. 3) Beware of overusing the flash: it does cost more than the radio in energy terms, as shown by this study, but often neglected. 4) Packet reception costs more than transmission instantaneously (e.g., in terms of power), but not over time (e.g., in terms of energy), as also shown in [18] and clearly demonstrated in this paper. 5) There exists an end-of-life stage in which it is easier for nodes to transmit than it is to receive. 6) In the end-of-life stage, the transmit power does not remain stable. There are important implications for experimental research. For instance, inferring conclusions on lifetime estimates based on wall-powered testbeds should be done with caution. Moreover, the choice of the batteries at deployment time should be taken seriously. Furthermore, the flash memory

should not be overused, and its energy footprint should be kept in mind. There is also a series of implications that are of interest to simulation-based studies. Our data and our set of observations clearly indicates that a realistic battery model is an essential ingredient of a successful simulation. The battery model and the radio model should not be treated as separate modules, and their mutual interaction should be considered. Moreover, a realistic simulation should also account for the peculiar end-of-life stage that we demonstrated in our experiments. Our immediate future plans include extending the presented experimental study with other types of batteries (lithium, zinc-carbon) and with other application parameters, such as very low duty cycles. R EFERENCES [1] I. Dietrich and F. Dressler, “On the Lifetime of Wireless Sensor Networks,” ACM Transactions on Sensor Networks (TOSN), vol. 5, no. 1, pp. 1–39, February 2009. [2] R. Szewczyk, J. Polastre, A. Mainwaring, and D. Culler, “Lessons from a Sensor Network Expedition,” in The 1st European Workshop on Wireless Sensor Networks (EWSN’04), Berlin, Germany, Jan. 2004. [3] S. Park, A. Savvides, and M. Srivastava, “Battery Capacity Measurement and Analysis using Lithium Coin Cell Battery,” in International Symposium on Low Power Electronics and Design (ISLPED’01), Huntington Beach, CA, USA, Aug 2001. [4] C. Park, K. Lahiri, and A. Ranghunathan, “Battery Discharge Characteristics of Wireless Sensor Nodes: An Experimental Analysis,” in The 2nd Annual IEEE Communications Society Conference on Sensor and Ad Hoc Communications and Networks (SECON’05), Santa Clara, CA, USA, Sep. 2005. [5] M. Spohn, S. Sausen, F. Salvadori, and M. Campos, “Simulation of blind flooding over wireless sensor networks based on a realistic battery model,” apr. 2008. [6] C. Chau, F. Qin, M. Wasab, and Y. Yang, “Harnessing Battery Recovery Effect in Wireless Sensor Networks: Experiments and Analysis,” IEEE Journal on Selected Areas in Communications, vol. 28, September 2010. [7] J. Rahm, N. Fourty, K. A. Agha, and A. van der Bossche, “A Recursive Battery Model for Nodes Lifetime Estimation in Wireless Sensor Networks,” in IEEE Wireless Communications and Networking Conference (WCNC’10), Sydney, Australia, Apr 2010. [8] M. Haenggi and D. Puccinelli, “Routing in Ad Hoc Networks: A Case for Long Hops,” IEEE Communications Magazine, vol. 43, pp. 93–101, October 2005. [9] K. Szlavecz, A. Terzis, S. Ozer, R. Musaloiu-E., J. Cogan, S. Small, R. Burns, J. Grey, and A. Szalay, “Life Under Your Feet: An End-to-End Soil Ecology Sensor Network,” Microsoft Technical Report, Tech. Rep. MSR-TR-2006-90, 2006.

[10] H. Ritter, J. Schiller, T. Voigt, A. Dunkels, and J. Alonso, “Experimental evaluation of lifetime bounds for wireless sensor networks,” in The 2nd European Workshop on Wireless Sensor Networks (EWSN’05), Istanbul, Turkey, Feb. 2005. [11] T. Staub, T. Bernoulli, M. Anwander, M. Waelchli, and T. Braun, “Lifetime Evaluation for MAC Protocols on Real Sensor Hardware,” in The 2nd Workshop on Real-World Wireless Sensor Networks (REALWSN’06), Uppsala, Sweden, Jun. 2006. [12] P. Hurni and T. Braun, “Calibrating Wireless Sensor Network Simulation Models with Real-World Experiments,” in The 8th International Conferences on Networking (Networking 2009), Aachen, Germany, May 2009. [13] G. Werner-Allen, P. Swieskowski, and M. Welsh, “MoteLab: a Wireless Sensor Network Testbed,” in 4th International Symposium on Information Processing in Sensor Networks (IPSN’05), Los Angeles, CA, Apr. 2005. [14] V. Handziski, A. Koepke, A. Willig, and A. Wolisz, “TWIST: A Scalable and Reconfigurable Testbed for Wireless Indoor Experiments with Sensor Networks,” in 2nd International Workshop on Multi-hop Ad Hoc Networks: from Theory to Reality (REALMAN’06), Florence, Italy, 2006. [15] A. Dunkels, F. Osterlind, N. Tsiftes, and Z. He, “Softwarebased on-line energy estimation for sensor nodes,” in Proceedings of the Fourth Workshop on Embedded Networked Sensors (Emnets IV), Cork, Ireland, Jun. 2007. [16] G. Mathur, P. Desnoyers, D. Ganesan, and P. Shenoy, “Ultralow power data storage for sensor networks,” in Proceedings of 5th International Conference on Information Processing in Sensor Networks (IPSN), Nashville, TN, USA, 2006, pp. 374 –381. [17] T. O’Donovan, N. Tsiftes, Z. He, T. Voigt, and C. J. Sreenan, “Detailed diagnosis of performance anomalies in sensornets,” in Proceedings of the 6th ACM Workshop on Hot Topics in Embedded Networked Sensors (HotEMNETS), Cork, Ireland, 2010. [18] J. W. Hui and D. E. Culler, “Ip is dead, long live ip for wireless sensor networks,” in Proceedings of the 6th ACM conference on Embedded network sensor systems (SenSys), Raleigh, NC, USA, 2008, pp. 15–28.