A Gateway Node with Duty-Cycled Radio and Processing...

5

A Gateway Node with Duty-Cycled Radioand Processing Subsystems for WirelessSensor Networks

ZHONG-YI JIN, CURT SCHURGERS, and RAJESH K. GUPTA

University of California, San Diego

Wireless sensor nodes are increasingly being tasked with computation and communication inten-

sive functions while still subject to constraints related to energy availability. On these embedded

platforms, once all low power design techniques have been explored, duty-cycling the various sub-

systems remains the primary option to meet the energy and power constraints. This requires the

ability to provide spurts of high MIPS and high bandwidth connections. However, due to the large

overheads associated with duty-cycling the computation and communication subsystems, existing

high performance sensor platforms are not efficient in supporting such an option. In this article, we

present the design and optimizations taken in a wireless gateway node (WGN) that bridges data

from wireless sensor networks to Wi-Fi networks in an on-demand basis. We discuss our strategies

to reduce duty-cycling related costs by partitioning the system and by reducing the amount of time

required to activate or deactivate the high-powered components. We compare the design choices

and performance parameters with those made in the Intel Stargate platform to show the effec-

tiveness of duty-cycling on our platform. We have built a working prototype, and the experimental

results with two different power management schemes show significant reductions in latency and

average power consumption compared to the Stargate. The WGN running our power-gating scheme

performs about six times better in terms of average system power consumption than the Stargaterunning the suspend-system scheme for large working-periods where the active power dominates.

For short working-periods where the transition (enable/disable) power becomes dominant, we per-

form up to seven times better. The comparative performance of our system is even greater when

the sleep power dominates.

Categories and Subject Descriptors: C.3 [Special-Purpose and Application-Based Systems]:

Real-time and embedded systems; C.4 [Performance of Systems]: Design studies

General Terms: Design, Experimentation, Measurement, Performance

Additional Key Words and Phrases: Embedded systems, gateway, power savings, sensor nodes

ACM Reference Format:Jin, Z., Schurgers, C., and Gupta, R. K. 2009. A gateway node with duty-cycled radio

and processing subsystems for wireless sensor networks, ACM Trans. Des. Autom. Electron.

Syst., 14, 1, Article 5 (January 2009), 17 pages, DOI = 10.1145/1455229.1455234 http://doi.

acm.org/10.1145/1455229.1455234

Authors’ address: Z. Jin, University of California, San Diego, 9500 Gilman Drive, La Jolla, CA

92093; email: [email protected].

Permission to make digital or hard copies of part or all of this work for personal or classroom use is

granted without fee provided that copies are not made or distributed for profit or direct commercial

advantage and that copies show this notice on the first page or initial screen of a display along

with the full citation. Copyrights for components of this work owned by others than ACM must be

honored. Abstracting with credit is permitted. To copy otherwise, to republish, to post on servers,

to redistribute to lists, or to use any component of this work in other works requires prior specific

permission and/or a fee. Permissions may be requested from Publications Dept., ACM, Inc., 2 Penn

Plaza, Suite 701, New York, NY 10121-0701 USA, fax +1 (212) 869-0481, or [email protected]© 2009 ACM 1084-4309/2009/01-ART5 $5.00 DOI 10.1145/1455229.1455234 http://doi.acm.org/

10.1145/1455229.1455234

ACM Transactions on Design Automation of Electronic Systems, Vol. 14, No. 1, Article 5, Pub. date: January 2009.

5:2 • Z. Jin et al.

1. INTRODUCTION

A wireless sensor network (WSN) consists of a collection of wireless sensornodes, which are small embedded devices with onboard sensors and wirelessradios. Without wires, sensor nodes either rely on limited energy supply frombatteries, or harvested energy from intermittent sources such as solar or wind.To ensure long lifetimes demanded by the application and deployment scenar-ios, the sensor nodes have to be very energy efficient. Popular sensor nodes suchas the Berkeley Mote [Polastre et al. 2004] address this issue using low powerhardware as well as aggressive power management techniques. However, theirdesign choices also make these nodes useful only to applications requiring lim-ited processing power, short communication range and low network bandwidth.

A common WSN architectural solution to these constraints is to deploy withinthe sensor network a small number of high performance nodes equipped withhigh powered components like fast processors or high bandwidth radios. Ashigh performance sensor nodes are usually placed in the same environmentas regular sensor nodes, they also rely on limited energy sources. Therefore,energy-efficiency is critical for the high performance nodes to last as long asthe rest of the sensor nodes.

There are two observations that we can explore to improve the energy-efficiency of high performance nodes. First, as a general fact, using compo-nents with high peak power consumption does not necessarily imply highenergy consumption. Studies have shown that when a sufficient amount ofdata need to be processed or transmitted, high performance processors or ra-dios that consume more power than their sensor node counterparts may com-plete the same amount of work faster and therefore end up using less energy[Jejurikar and Gupta 2004; McIntire et al. 2006]. Secondly, in the specific caseof sensor networks, those high powered components are not required to be ac-tive all the time as sensor networks usually do not generate large amounts ofdata until certain triggering events are detected. In other words, the node or itscomponents needs to be active only a fraction of the time to achieve applicationgoals. Therefore, duty-cycling based power management techniques, such asselectively enabling or disabling components, are important in reducing energyconsumption for high performance nodes.

A platform needs to meet two requirements to support efficient duty-cycling.One is that it needs to consume very little (or no) power when there are noongoing activities. While general purpose high performance nodes such as theStargate [Crossbow 2004] provide a good balance of performance and powerconsumption, they are not designed to support efficient power managementvia duty-cycling. For example, the lowest power consumption of a Stargate is16.2mW in its inactive suspended state [Margi et al. 2006]. In contrast, a typi-cal sensor node such as the Telos Mote uses only about 10μW in a similar state[Polastre et al. 2005]. This high standby power consumption significantly limitsthe effectiveness of duty cycling, making the Stargate less energy efficient forvery low duty cycle sensor network applications such as environmental mon-itoring. The other requirement is that a platform needs to be able to activateor deactivate various subsystems with very little energy and latency overhead


A Gateway Node with Duty-Cycled Radio and Processing Subsystems • 5:3

Fig. 1. Wireless sensor networks with 802.11b based gateway nodes.

according to runtime demands. Having low latency is critical because it notonly affects the energy cost of duty-cycling but also determines the feasibilityof applying duty-cycling as tasks may have deadlines. Existing high perfor-mance nodes built around 32-bit embedded processors and embedded versionsof traditional desktop Operating Systems (OSs), both old [Crossbow 2004] andnew [Banerjee et al. 2005; McIntire et al. 2006], generally take a long time totransit in and out of the suspend or power off states, limiting the performanceand potential applications of duty-cycling [Margi et al. 2006; Banerjee et al.2005; McIntire et al. 2006].

In this article, we describe the design, implementation and evaluation ofa wireless gateway node (WGN) that enables efficient power managementthrough duty-cycling. As shown in Figure 1, gateway nodes are often intrin-sic parts of sensor networks and are required to bridge data between sensornetworks and servers/devices in other networks. Without gateway nodes, datacollected by a sensor network are only local to that sensor network and there-fore no remote monitoring or interactions can be achieved. The low and burstytraffic load of sensor networks makes the WGN an ideal application of our lowpower design. Specifically, we focus on using Wi-Fi (802.11b) radios to interfacesensor nodes with devices in other networks because Wi-Fi radios offer higherbandwidth and consume less energy per bit than most existing sensor noderadios [Raghunathan et al. 2004], making them a useful air interface for manysensor network applications.

We build a working prototype of the WGN with commercial off-the-shelf com-ponents and evaluate the design experimentally using two power managementschemes. Our results show significant reductions in latency and average powerconsumption compared to the Stargate configured as a packet forwarding gate-way node. Given the Stargate and similar platforms are designed for generalpurposes and do not support fast wakeup from power off state, instead of a com-prehensive comparison of the strengths and weaknesses, our results provide ameasure of potential reduction in power consumption due to architectural re-design for duty-cycling and for the gateway application.

2. RELATED WORK

To energy efficient design, the importance of separating real-time monitoringfunctions that have to be optimized for low power from functions invoked withlight duty-cycles is first unveiled in the development of the WINS nodes [Pottieand Kaiser 2000]. The WINS node enables continuous sensing over an extend



Fig. 2. WGN block diagram.

period of time by partitioning the system into a low powered event-detectionsubsystem and a high powered processing subsystem which can be enabled ordisabled on demand.

Our work is based on [Jin et al. 2007] and is directly comparable to the emerg-ing Microservers [Rahimi et al. 2005; Banerjee et al. 2005]. Numerous effortshave been undertaken to reduce the average power consumption of these de-vices. Triage [Banerjee et al. 2005] extends the lifetimes of their Microserversby trading latency for energy reduction. Its tiered architecture consists of aslightly modified Stargate computer and a MicaZ mote, which is used to poweron the Stargate only when sufficient amount of data are being batched for pro-cessing. Due to the large latency in powering on/off the Stargate, their platformis not usable for our gateway application in terms of delay and power consump-tion. The LEAP platform [McIntire et al. 2006] faces similar issues, as stated intheir future work section. The Cyclops [Rahimi et al. 2005] platform integratesa CPLD (Complex Programmable Logic Device) with a sensor node for theirimage sensing application. It is not suitable for our purposes as the CPLD isnot ideal for general purpose applications. The PASTA platform [Schott et al.2005] consists of a family of modules that can be stacked together. By embed-ding a power microcontroller clocked at 32kHz into each module, their archi-tecture enables independent power control of modules and module subsystems.Their processor module (PXA module) also uses an Intel PXA255 processor.Since their demonstrated mode of operation is to keep the processor modulein sleep state during periods of inactivity to save power, no experiments andlatency numbers are reported for activating this module from power-off state.Unlike PASTA, our work also explores the use of power-off state to further re-duce standby power consumption by minimizing wakeup energy and latencyoverhead.

3. DESIGN APPROACH

Figure 2 shows a high-level block diagram of our design. To minimize standbyenergy consumption, we exploit the low power operation of a sensor nodeprocessor and use it for subsystem scheduling and power management. A sec-ond, more powerful application processor is used to provide on demand process-ing capabilities to subcomponents such as the Wi-Fi radio, which are physicallyconnected to the application processor. Power to each individual subcomponent



Fig. 3. WGN architecture.

is either controlled directly by the sensor node processor or indirectly by theapplication processor through the sensor node processor. We use a serial inter-face for inter-processor communication because it is supported in various formsby most of the existing sensor node processors.

Unlike the PASTA platform [Schott et al. 2005] where multiple microcon-trollers are used to regulate power to the modules, we have the sensor nodeprocessor acting as the master device for power management. Our approachsimplifies the design while also taking advantage of the fact that the sensornode processor is almost always on for sensing or networking purposes. Thisenhances the efficiency of running the power management itself.

Besides supporting the low-power sleep mode as described in the PASTApaper, we also want to be able to put the application processor in and out ofpower-off mode to further reduce standby power consumption. The tradeoffsof using these two modes for energy reduction are studied in [Jejurikar andGupta 2004]. To reduce the energy and latency overhead in activating the ap-plication processor from power-off mode, we found it is critical to minimize bothhardware and software startup time. On the hardware side, commercial micro-controllers for embedded devices are usually designed to support fast startup.On the software side, it is important to minimize the amount of the code thatneeds to be loaded and executed every time during boot up. This includes bothOS and application specific code.

4. PLATFORM IMPLEMENTATION

The hardware architecture of our WGN is illustrated in Figure 3. As explainedin the introduction, our main contributions are on the level of the architec-tural design approach (see also section 3). To illustrate these ideas and per-form experiments, we have to make specific design choices for our test bedplatform. Our approach, however, is not restricted to these specific choicesalone.

We use the Ubicom IP2022 processor [Ubicom 2003] as our application pro-cessor. With a maximum clock speed of 120MHz (approximately 120MIPS),this 8-bit processor is less powerful than the 400MHz (480MIPS, Dhrystone2.1) PXA255 on the Stargate. However, it is significantly faster than the mi-crocontrollers on most existing sensor nodes and provides sufficient processingcapability to our gateway application which is not computation intensive. Withits integrated flash memory and RAM, this processor does not need complex



external hardware support and thus does not incur extra energy and latencyoverhead. We also find the IP2022 a convenient choice as it is used in the DPACmodule [Dpactech 2006] with a Wi-Fi radio.

Because of a clear separation of the master (sensor node processor) and therest of the system on our platform, we can select any existing sensor nodes suchas the Berkeley Mote [Polastre et al. 2004] for our power management purposes.Our design only requires that the sensor node processor and the applicationprocessor can communicate and wake each other up as necessary. This is easierto implement than those that require external logics or chip-set supports as inPC platforms. Commonly used serial interfaces such as the UART, I2C, or SPIare sufficient to meet these requirements. We select the SPI since it supportsduplex data transfers at a rate of more than 11Mbps, sufficient to sustain thethroughput of the Wi-Fi radios.

For evaluation purposes, we use for power management a home grown sensornode equipped with a PIC18F452 microcontroller that consumes about 8mW(3.3V) in full power mode (10MIPS). The hardware SPI on the PIC is still avail-able. The power to the DPAC module is managed by the PIC through a MOSFETswitch. Alternatively, we can use an I2C power switch to control additional com-ponents. We use Ubicom IPOS, a lightweight OS for the IP2022. The IPOS pro-vides the very basic OS functions and gets compiled with the target application.Our entire gateway application is about 60 Kbytes in uncompressed form.

5. POWER MANAGEMENT SCHEMES

We experiment with two different power management schemes for our gatewayapplication to evaluate the platform as well as to understand the power ver-sus latency tradeoffs. We refer to these two schemes broadly as power-gatingand power-saving. In the former, the emphasis is on subsystem shutdown forboth communication and processing, while the latter seeks to exploit variousslowdown modes.

5.1 Power-Gating Scheme

Our power-gating scheme saves energy by turning off the power to the Wi-Firadio and the application processor. As no packets can be sent or received whilein the power off state, a successful use of the power-gating scheme requiresparticipating gateway nodes and devices in other networks to coordinate theiractive periods while minimizing the total amount of energy required for suchsynchronizations [van Dam and Langendoen 2003; Polastre et al. 2004; Ye et al.2002]. We measure the overheads of such protocols when used with our WGN.The overheads are quantified either as time in seconds or energy in Joulescalculated by integrating power over time.

5.2 Power-Saving Scheme

Our power-saving scheme reduces energy consumption by putting the radioin the 802.11b power saving mode [IEEE 1999] while keeping the applicationprocessor in various low power modes. Since most Wi-Fi radios natively supportthe 802.11b power saving mode, it is significantly faster to put the radio in and



out of the power saving mode than to suspend and resume the entire radio ineach duty cycle. Although the speedup is hardware dependent, it is reported inone case that it is almost 86 times faster to resume from power saving modethan from suspended mode [Agarwal et al. 2005]. As shown in Figure 4 andFigure 6, the speedup is approximately 9 times on our Intersil Prism3.0 Wi-Fichip based WGN, without taking into consideration the re-association time.

One challenge in supporting such a scheme is to reduce the wakeup latencyand power consumption of the Wi-Fi radio running in the 802.11b power savingmode. When in the 802.11b power saving mode, a radio wakes up periodicallyto check for any incoming packets buffered at the AP (Access Point) and thesleep duration is determined by the listen interval. A listen interval is estab-lished during the association process and is fixed until new association attemptsare made. Since the reassociation process involves multiple packet exchangesbetween the station and the AP, it is expensive to change the listen intervalfrequently. However, a fixed listen interval is not ideal for most sensor networkapplications, since events occur randomly. Long listen intervals introduce largecommunication delays while short listen intervals waste energy if there areno events or data to send or receive. Thus the choice of the listen interval isa matter of design tradeoff between energy savings and latency incurred. In-stead of listening at a fixed interval, we develop an event-driven approach totransition in and out of the power saving mode as explained subsequently inits implementation.

Our strategy is based on the 802.11b standard [IEEE 1999] that after astation exits power saving mode, the AP should send all buffered packets tothat station as if they just arrived. Therefore, if the listen interval is set to anarbitrary large value (up to 216 beacon intervals), one can eliminate its effectsand dynamically control the packet receiving time by forcing the station outof the 802.11b power saving mode. Although a successful packet exchange isrequired between the station and the AP to enable or disable the 802.11b powersaving mode, this can by done by simply changing the power management bitin the frame control field of any outgoing packets. A potential benefit of thisstrategy is that with some buffering, sending and receiving can be performedwithin the same active periods and therefore reduces the total amount of timethe radio and the application processor need to be awake. We experiment withthis approach by sending and receiving 1024 bytes of data in an UDP packetevery 6 seconds for a period of 30 minutes using the same method describedin section 6.4 and observe no packet loss. We also verify this technique on aLinux laptop with a Netgear WG511 Wireless PC card (V1.0) and latest hostAPdriver (V0.5.1) and observe similar results. Note that to avoid overrunning thebuffer of the AP, a small packet should be sent to instruct the receiving devicerunning our scheme to wakeup more frequently before transmitting a largeamount data.

6. EXPERIMENTAL RESULTS

In this section, we present our experimental results of evaluating our WGNwith the two power management schemes discussed in section 5.



Fig. 4. Switch from off (gated) mode to active mode.

6.1 Measurements and Chart Interpretation

Power consumption is measured as the product of voltage and current. OurWGN is directly powered by a DC power supply of 3.3V. To measure current, a re-sistor of 1Ohm is connected in series with the DC power supply. The voltage dropacross the resistor is sampled using a National Instrument DAQPad-6020E(12Bit, 100KS/s) and stored in Lab-View spreadsheet format. For comparisonpurposes, hardware components such as the Wi-Fi radio and the processors aswell as software components like the OS, IP and UDP libraries are always beinginitialized to default states during initial power up. For simplicity, sensor dataare randomly generated by the PIC processor rather than from real sensors orother sensor nodes.

6.2 Power-Gating Overheads

In this experiment, we measure how long it takes to activate the Wi-Fi radioand the application processor from power-off state and the overall power con-sumption of the system in active state. We start the experiment by poweringup the whole system, including the radio and both processors. The activationprocess ends when the application processor is running at full speed, ready toexecute real application tasks. The exact time of completion can be observedon the chart as a sharp voltage drop induced for evaluation purposes by shut-ting the application processor off. Since the 802.11b association-time variessignificantly with different access points, authentication methods and networksettings, it is excluded from measurements in this experiment.

As shown in Figure 4, the WGN is powered up at around 4.47 seconds.It takes about 280ms (4.47s-4.75s) to switch from power-gating mode to ac-tive mode. This is about 43 times faster than powering up the Stargate from



Fig. 5. Power saving scheme.

power-off state based on results from Triage [Banerjee et al. 2005]. The reasonthe system still consumes about 82mA of current after shutting off the applica-tion processor is that the Wi-Fi radio draws power even after being put into offstate, probably due to hardware limitations. A similar case has been reportedin Pering et al. [2005] where a Netgear MA701 CF Wi-Fi card, which is alsobased on the Intersil chip, consumed 83mA (3.3V) of power in off mode. For ourpower-gating scheme, we work around this by shutting off power to both theapplication processor and the Wi-Fi radio.

6.3 Power-Saving Overheads

In this experiment, we evaluate the performance of our power-saving scheme.The experiment setup is similar to the one in Figure 1 but consists of only oneWGN, one access point (a part of the UCSD campus wireless network) and oneserver that is connected to the internet through wired ethernet. We performthe evaluation by first passing data from the sensor node to the applicationprocessor and then have the application processor send the data to the serverthrough the Wi-Fi radio.

Figure 5 shows the results of running the power-saving scheme code inFigure 7. The system is powered up at around 4.0s. The networking subcompo-nent (Wi-Fi radio and the application processor) enters into the sleep state inapproximately 12s. At about 16s (4 + 4 + 8), the PIC interrupts and sendsthe first 64 bytes of data to the IP2022. The rise of power consumption at thatpoint is the result of waking up the IP2022 and the radio from low power states.After waking up, the IP2022 instructs the radio to exit the 802.11b power sav-ing mode and sends the 64 byte of data in a UDP packet to the server. If thereare incoming packets buffered at the AP for this node, they will be delivered bythe AP upon receiving this packet.



Fig. 6. Zoomed in view of Figure 5.

Certain fixed delays are introduced to simplify synchronization efforts. Asshown in Figure 7, the PIC waits one second after the interrupt for the IP2022to fully wakeup. The one second timer of the IP2022 can keep the networkingsubcomponent from entering into low power states when a continuous streamof data is being sent or received. Taking these delays into considerations, thetime of the uprising edges shown in Figure 5 can be calculated as:

T0 = 16; T(n+1) = Tn + 1 + 8; (1)

where T(n) is the time to send the nth packet. The amount of time to send 64bytes of data in various formats from the PIC to the IP2022, from the IP2022to the Wi-Fi radio and from the Wi-Fi radio to the access point is too smallto be recognized under the displayed time scales. Figure 6, a zoomed in viewof Figure 5 indicates that it takes about 30ms (16.37s-16.4s) to wake up theIP2022 and fully power up the radio from the 802.11b power saving doze state.

The 30ms wakeup time is slower than the reported 4ms wakeup time of thePXA255 processor on the PASTA platform [Schott et al. 2005]. We believe thisis due to a counting difference as our measurement also includes processor pre-charge time [Schott et al. 2005] and the time to fully wakeup the Wi-Fi radio.The PXA module on the PASTA platform does not contain a Wi-Fi radio andconsumes 7-10mW in sleep state which is significantly less than the 0.495Wsleep power of our WGN. The high sleep-power of our power-saving schemeis caused by the abnormal power consumption of the Intersil Wi-Fi chip asexplained in section 6.2.

6.4 An Echo Server

Our last experiment measures the performance of our 802.11b based power-saving technique discussed in section 5.2 using a real world application. We im-plement on the application processor an echo server that listens on a predefined



Fig. 7. Pseudo code of our power-saving scheme.

port and echoes incoming UDP packets back to the sender. The WGN is con-nected wirelessly to the Internet via the UCSD campus wireless network. Acontinuous stream of data is sent from a desktop computer through the Internetto the WGN running the power-saving scheme code as shown in Figure 7. Roundtrip time is calculated as the difference between the sending time, which is en-coded as part of the probe packet being sent, and the receiving time recordedwhen the echoed probe is being received. As shown in Figure 8, after an ini-tial delay of six seconds, the wait time is eliminated completely because boththe radio and the application processor are kept alive for one second by eachpacket sent or received. As long as packets come in or out with an inter-packetdelay of less than one second, the WGN will not return back to the power-savingmode.

7. ANALYSIS

Average system power consumption is calculated1 based on the amount of en-ergy consumed in one working-period, which is defined as the period from thebeginning of one active period to the beginning of the next active period. Forexample, the period from 16s to 25s in Figure 5 is a working-period with a du-ration of nine seconds. A power managed working-period can be further dividedinto four sub-periods: a period to enable the system, a period of doing the real

1 Average S ystem Power = Total ener g y consumed in one working−periodDuration of the working−period .



Fig. 8. Round trip delay with a continuous stream of data.

Table I. Latencies and Power Consumption

Stargate Our WGN

Suspend Suspend Always Power Power

Wi-Fi system on gating saving

scheme scheme scheme scheme scheme

Enable Latency 0.485s 3.329s - 0.28s 0.03s

Enable Power 0.751w 0.155w - 0.545w 0.693w

Active Power 2.009w 2.009w 1.419w 1.419w 1.419w

Disable Latency 0.313s 0.757s - 0s 0.003s

Disable Power 1.62w 1.11w - 1.419w 1.32w

Sleep Power 0.751w 0.054w - 5.13mw 0.495w

work, a period to disable the system and a sleep period. A duty-cycle is calcu-lated as the percentage of the time that a system does real work over an entireworking-period. It does not include time spent in enabling or disabling the sys-tem. Note that we can maintain a fixed duty-cycle by proportionally changingthe working time and the duration of the working-period.

Table I lists the durations of these periods as well as the corresponding powerconsumptions in these periods for both the Stargate and our system runningdifferent power management schemes. We choose to compare our platform withthe Stargate because it is one of a few gateway nodes commonly used in sen-sor networks. Other gateway nodes, such as those based on the Soekris board[Hartung et al. 2006], share similar architecture as the Stargate. The compar-ison is made under the condition that the Stargate and our platform are usedas gateway nodes to perform simple packet forwarding functions for a wirelesssensor network. These functions are not computation intensive.

For our WGN, the enable-power of the power-gating scheme is less than thatof the power-saving scheme because not all components are powered up at thesame time. As shown in Figure 4, there is actually a long delay before the WGNreaches full power. The reason for the high sleep-power of our power-savingscheme is explained in section 6.3. The Stargate has very high sleep-powerbecause its PXA255 processor is in charge of power management and can notbe powered off completely.

In the remainder of this section, we compare the performance of our powermanagement schemes using the WGN with the performance of the followingtwo commonly used schemes on the Stargate:



Fig. 9. Average system power consumption under various working-periods and at a fixed 1% duty

cycle.

(1) Suspend-Wi-Fi Scheme: suspend the Wi-Fi radio only.

(2) Suspend-System Scheme: suspend both the Wi-Fi radio and the Stargatecomputer.

The Stargate data are based on measurements from [Margi et al. 2006]. Wecombine the latencies that are reported separately for the Wi-Fi radio and thePXA255 and compute the average power consumption. Similar to our approach,the authors in Margi et al. [2006] measure data without sensors attached. Ac-cordingly, we use the “Processor Core Idle” data for the Suspend-Wi-Fi schemeand the “Proc./Radio Core Sleep” data for the Suspend-System scheme. The ac-tive power consumption in the active period is based on 50% TX and 50% RX.For our own schemes, the power consumed by the entire WGN is reported. Theload on the 1Ohm resistor is included to simulate a real sensor. Our always-onscheme simply keeps the system in maximum power all the time and is used toserve as a baseline. The sleep-power of our power-gating scheme is the same asthe sleep-power of the attached sensor node. Since it is too small to be measuredaccurately, we simply plug the standby power consumption of the Mica2 motefrom [Polastre et al. 2005] in the table.

Figure 9 shows the average system power consumption of the five schemesrunning under various working-periods and at a fixed 1% duty cycle, whichis very common for sensor networks. With a fixed duty cycle, the time spentdoing real work increases in proportion to the duration of the working-period,and therefore the average amount of real work per unit time remains constant.Large duty-cycle latencies mean less sleep time. The WGN running our power-gating scheme performs about six times better than the Stargate running thesuspend-system scheme for large working-periods where the active power dom-inates. For short working-periods where the transition (enable/disable) power



Fig. 10. Lifetimes under various working-periods and at a fixed 1% duty cycle (2200 mAh at 3

volts).

becomes dominant, we perform up to seven times better. This is partially dueto the small transition latencies that result from applying the 802.11b basedpower saving techniques described in section 5.2.

Figure 10 shows the lifetimes of our WGN and the Stargate running the fiveschemes under various working-periods and at a fixed 1% duty cycle. They arecomputed based on a power supply of 2200mAh at three volts from a pair of AAbatteries. The WGN could last longer with smaller duty cycles because of theextremely low sleep-power.

Although it is possible with some hardware modifications to eliminate thesleep power of the Stargate processor by powering it off and to reduce the active-power of the Stargate processor by dynamic voltage scaling (DVS) [Banerjeeet al. 2005; McIntire et al. 2006], our WGN would still perform better becauseof lower duty-cycle latencies. This is a direct result of the new architecturewe propose. To demonstrate its benefits, while normalizing away the specifichardware choices, we consider the Stargate schemes assuming the same hard-ware power numbers as our own schemes. In other words, we substitute thehardware power numbers of the Stargate suspend-Wi-Fi scheme and suspend-system-scheme with the ones from our power-saving and power-gating schemes,respectively. The results are shown in Figure 11, which corresponds to Figure 9,but is normalized in terms of hardware numbers and illustrates the gains specif-ically due to our architecture.

We notice in Figure 11 that when the active-power becomes dominant (largeworking periods), the power consumed by the corresponding schemes of thetwo systems starts to converge. This is because we assume the correspond-ing schemes have the same hardware power numbers. While it seems that theenable/disable latencies are not important in terms of average system power



Fig. 11. Average system power consumption based on normalized power under various working-

periods and at a fixed 1% duty-cycle.

Fig. 12. System response time.

consumption for applications that are dominated either by active or sleep power,they are critical in determining the responsiveness of a system. A large latencyin activating a subsystem from low-power or power-off mode would be pro-hibitive for many sensor network applications to employ duty-cycling. Smallerlatencies also provide additional space for applications to trade latencies for en-ergy savings. Figure 12 shows the system response time under different powermanagement schemes. When running power-gating scheme, our WGN is about12 times better than the Stargate running the suspend-system scheme. TheWGN running the power-saving scheme is about 16 times better than the WGNrunning the suspend-Wi-Fi scheme.

8. CONCLUSION

In this article, we present the design and optimizations of a low power WGNfor low duty cycle sensor network applications. By introducing a dual-processor



hardware architecture and a choice of appropriate duty-cycling of the processingand radio subsystems, we successfully reduce the standby power consumptionwhile also providing support for spurts of high MIPS and high bandwidth con-nections. We improve the performance of duty-cycling with respect to energyand latencies on the WGN by reducing software and networking protocol re-lated overheads and through careful system integration. The result is a gatewaynode that supports efficient power management through duty-cycling.

We also demonstrate the performance gains of our architecture by normal-izing away hardware specific power numbers. This suggests our architecture isnot constrained to any specific hardware such as the IP2022 processor. As newhardware comes out, the set of guidelines that we follow in choosing the righthardware and software components for our WGN can be applied to build newplatforms. We believe our architecture is very useful for building other types ofhigh performance nodes or the emerging Microservers.

ACKNOWLEDGMENTS

We want to thank Dr. Douglas A. Palmer of Calit2 for invaluable help andencouragement.

REFERENCES

AGARWAL, Y., SCHURGERS, C., AND GUPTA, R. 2005. Dynamic power management using on demand

paging for networked embedded systems. In Proceedings of the Asia South Pacific Design Au-tomation Conference (ASP-DAC), 755–759.

BANERJEE, N., SORBER, J., CORNER, M. D., ROLLINS, S., AND GANESAN, D. 2005. Triage: a power-aware

software architecture for tiered microservers. Tech. Rep. 05-22, University of Massachusetts-

Amherst. April.

CROSSBOW. 2004. Stargate. http://www.xbow.com.

DPACTECH. 2006. DPAC. www.dpactech.com.

HARTUNG, C., HAN, R., SEIELSTAD, C., AND HOLBROOK, S. 2006. Firewxnet: a multi-tiered portable

wireless system for monitoring weather conditions in wildland fire environments. In Proceedingsof the 4th International Conference on Mobile Systems, Applications and Services. ACM Press,

New York, NY, 28–41.

IEEE. 1999. 802.11b Spec 1999 edition. http://grouper.ieee.org/groups/802/11.

JEJURIKAR, R. AND GUPTA, R. 2004. Dynamic voltage scaling for systemwide energy minimization

in real-time embedded systems. In Proceedings of the International Symposium on Low PowerElectronics and Design. ACM Press, New York, NY, 78–81.

JIN, Z., SCHURGERS, C., AND GUPTA, R. 2007. An embedded platform with duty-cycled radio and

processing subsystems for wireless sensor networks. In SAMOS, S. Vassiliadis, M. Berekovic,

and T. D. Hamalainen, Eds. Lecture Notes in Computer Science, vol. 4599. Springer, 421–

430.

MARGI, C. B., PETKOV, V., OBRACZKA, K., AND MANDUCHI, R. 2006. Characterizing energy consump-

tion in a visual sensor network testbed. In Proceedings of the 2nd International IEEE/Create-NetConference on Testbeds and Research Infrastructures for the Development of Networks and Com-munities.

MCINTIRE, D., HO, K., YIP, B., SINGH, A., WU, W., AND KAISER, W. J. 2006. The low power energy aware

processing (leap)embedded networked sensor system. In Proceedings of the 5th InternationalConference on Information Processing in Sensor Networks. ACM Press, 449–457.

PERING, T., RAGHUNATHAN, V., AND WANT, R. 2005. Exploiting radio hierarchies for power-efficient

wireless device discovery and connection setup. In Proceedings of the 18th International Con-ference on VLSI Design, held jointly with 4th International Conference on Embedded SystemsDesign. IEEE Computer Society, 774–779.



POLASTRE, J., HILL, J., AND CULLER, D. 2004. Versatile low power media access for wireless sensor

networks. In Proceedings of the 2nd International Conference on Embedded Networked SensorSystems. ACM Press, New York, NY, 95–107.

POLASTRE, J., SZEWCZYK, R., AND CULLER, D. 2005. Telos: enabling ultra-low power wireless re-

search. In Proceedings of the 4th International Conference on Information Processing in SensorNetworks.

POTTIE, G. J. AND KAISER, W. J. 2000. Wireless integrated network sensors. Comm. ACM 43, 5,

51–58.

RAGHUNATHAN, V., PERING, T., WANT, R., NGUYEN, A., AND JENSEN, P. 2004. Experience with a low

power wireless mobile computing platform. In Proceedings of the International Symposium onLow Power Electronics and Design. ACM Press, 363–368.

RAHIMI, M., BAER, R., IROEZI, O. I., GARCIA, J. C., WARRIOR, J., ESTRIN, D., AND SRIVASTAVA, M. 2005.

Cyclops: in situ image sensing and interpretation in wireless sensor networks. In Proceedingsof the 3rd International Conference on Embedded Networked Sensor Systems. ACM Press, New

York, NY, 192–204.

SCHOTT, B., BAJURA, M., CZARNASKI, J., FLIDR, J., THO, T., AND WANG, L. 2005. A modular power-

aware microsensor with > 1000x dynamic power range. In Proceedings of the 4th InternationalSymposium on Information Processing in Sensor Networks. IEEE Press, 66.

UBICOM. 2003. IP2022. http://www.ubicom.com.

VAN DAM, T. AND LANGENDOEN, K. 2003. An adaptive energy-efficient mac protocol for wireless

sensor networks. In Proceedings of the 1st International Conference on Embedded NetworkedSensor Systems. ACM Press, New York, NY, 171–180.

YE, W., HEIDEMANN, J., AND ESTRIN, D. 2002. An energy-efficient mac protocol for wireless sen-

sor networks. In Proceedings of the 21st Annual Joint Conference of the IEEE Computer andCommunications Societies, 1567–1576.

Received September 2007; revised April 2008; accepted June 2008


A Gateway Node with Duty-Cycled Radio and Processing...

Documents

Transcript of A Gateway Node with Duty-Cycled Radio and Processing...