1

Birthday Paradox for Energy Conservation in Sensor Networks

Santashil PalChaudhuri and David B. Johnson Computer Science, Rice University Houston, Texas
Abstract Recent advances in technology have made low-power wireless sensor networks feasible. Energy saving is one of the key design issues in sensor networks. There has been some work in trying to save energy in sensor and ad hoc networks, using approaches like switching to sleep mode or transmission power control. Our work in this paper is a probabilistic protocol, which is complementary to all the previous work in this area. A wireless network interface in sleep mode expends an order of magnitude less power than in idle mode, but no packets can be sent or received while in sleep mode. We propose a novel approach for energy conservation using the well known Birthday Paradox, for scheduling transition from idle to sleep mode. We trade-off latency for a signicant reduction in the amount of expended energy. This strategy provides some tunable parameters for the system, which can be used to adjust the performance and energy characteristics as necessary by an application.

I. I NTRODUCTION Energy consumption in low-power wireless sensor nodes is an important issue in wireless sensor networks. The sources of power consumption are communication, sensing and computation, with communication often being the chief power consumer. There has been a lot of research on conserving power in these sensor nodes. Most of this research on power conservation has focused on putting the network interface to sleep mode, or controlling the transmission power of the network interface. A network receiver interface has four different modes: 1. Transmit Mode: In this mode, the node is transmitting a packet. The power consumption in this mode is the highest amongst the four modes. 2. Receive Mode: In this mode, the node is receiving a packet. The power consumption in this mode is slightly lesser than the Transmit Mode 3. Idle Mode: A receiver in idle mode can either receive or transmit. It burns power because it has to listen to the wireless medium all the time to determine whether or not there is a packet transmission going on. This mode takes slightly lesser power than Receive Mode. 4. Sleep Mode: Sleep mode has very low power consumption. The network interface cannot receive or transmit in this mode unless it is woken up into idle mode by an explicit instruction. It takes a nite time to transition from sleep mode to idle mode and vice versa.

There has been some work in the area of energy management in sensor and ad hoc networks that provides for transition between these different modes based on need. Most important among these are Geographical Adaptive Fidelity (GAF) [23] and Span [2] in the routing layer, and the work by Tseng et al. [19] and Ye et al. [24] in the MAC layer. Our work, while having the similar approach of switching between the different power modes, is orthogonal to the above approaches. It can be added on top of any of the above approaches to produce more energy savings. In this paper, we propose this probabilistic approach based on a dynamic adaptation of the Birthday Paradox, which was outlined earlier by authors in [14]. The concept of Birthday Paradox has been applied before in elds such as cryptography and hashing algorithms. We utilize this paradoxical theory to our benet for saving energy. We propose two algorithms for transition of a wireless network interface from idle mode to sleep mode, and vice versa. The basic algorithm is motivated by the Birthday Paradox. Each node randomly wakes up and goes to sleep. The percentage between idle and sleep mode is determined by the neighbor node density as well as the remaining battery power. Neighboring nodes can communicate only when both are awake. A Sleep Indication Map (SIM) is sent by every node to all of its neighbors after every interval. Each node utilizes this SIM to schedule trafc to its neighbors. The second algorithm is similar to GAF or Span. A node transitions to sleep mode when it is not actively involved in sending, receiving, or forwarding data. The node periodically goes to idle mode and to sleep mode, with a duty cycle depending on the number of neighbors it has. Finally, we model these algorithms in the ns-2 network simulator and study their performance. The remainder of this paper is organized as follows. In Section II, we describe the energy usage characteristics on which our design and evaluation is based. In Section III, we discuss related work. In Section IV, we propose two algorithms for scheduling transitions to different power modes, which are evaluated via simulations in Section V. Finally, we conclude in Section VI and discuss potential new research directions.

TABLE I IEEE 802.11 2 M BPS WAVELAN PC C ARD C HARACTERISTICS

Mode Sleep Idle Receive Transmit Power Supply

Specication 9mA n/a 280mA 330mA 5V

Measured 14mA 178mA 204mA 280mA 4.74V

II. BACKGROUND We base all our simulation experiments on a specic wireless network card for IEEE 802.11. Our assumption is that, the network interfaces will give similar characteristic across a broad range of these interfaces. A. Energy Consumption in a Wireless Network Interface In the evaluation of any algorithm for energy conservation, an estimate of energy consumption is necessary. In particular, the more closely a simulation reects specic hardware, the more accurate the estimate of energy consumed in the simulation experiments is. Table I, from Feeney [5], shows the actual current drawn by a wireless network interface card in the four possible modes. This table shows that receive and idle mode require similar power while transmit mode requires slightly greater power. Sleep mode requires more than an order of magnitude less power than idle mode. This implies that the network interface expends similar energy, whether it is just listening or actually receiving data. Hence, intelligently switching to sleep mode whenever possible will generally create signicant savings. The energy consumed by a wireless network interface when a node transmits, receives, or discards a packet can be described using a linear equation (1)

small step discontinuity, reecting the xed fragmentation overhead. This model also does not consider energy consumed in unsuccessful attempts to acquire the channel (media contention), or in messages lost due to collision, bit error, or loss of wireless connectivity. Such experiments were deemed difcult to obtain in controlled experimental measurements [5]. Although bandwidth metrics count the number of packets sent over the wireless media, energy consumption metrics must take into account the reaction of all the network interfaces within wireless transmission range of the transmitters. The operations of the wireless network interface can be divided into these three operations:

The incremental payload costs, and , are the same as those for broadcast trafc. Table II, from [5] gives the experimental data of all the cases detailed earlier. B. Power Mode Transition From the above discussion, the energy expended by a wireless network interface in each mode is known. What is also important is the amount of energy expended and

GE7B 3 qsq !FDCS(2' 9 875 3 qs )(F6SF2q  GE7B 3 qs $ "  GE7B 3 qs   V X W f d !FDCSF2q vw)(u !FDCSFAq h`Yg!AeYc 9 875 3 qs $ "  9 875 3 qs   V X W V  R @FF6SFAq v#(u @(46S(2q ba`YT!UTSQP

There is a xed cost ( ), associated with device mode changes and channel acquisition overhead. The incremental component ( ), is proportional to the size of the packet. The total cost of a packet is the sum of the costs incurred by the sending node and all receivers. Experimental results [5] conrm the accuracy of the linear model. Experiments were performed to determine values for the linear coefcients and for various operations of the network interface. We use this energy model and the associated values [5], [6] in the simulations in this paper. The following discussion is about the assumptions of this model. The model does not consider link layer fragmentation. It is expected that the linear model would continue to apply, with each instance of fragmentation introducing a

Point-to-point Trafc: For point-to-point trafc, the xed cost includes both the channel access cost and the MAC negotiation cost, where the channel access cost is the same as that in the broadcast case. In the IEEE 802.11 MAC protocol, the source sends an RTS (Request-to-Send) control message, identifying the destination. The destination responds with a CTS (Clear-to-Send) message. Upon receiving the CTS, the source sends the data and awaits an ACK from the destination. This control overhead is taken into account in the xed overhead to send a point-to-point packet. Dening the xed channel access cost for sending and receiving a packet as and respectively, and the incremental payload costs as and respectively, Equation 1 gives:

GE7B 3 qsq !FDCSF2t

GE7B 31 !FDC4

G!FDC4A%#!p!4i6Fh`Yg!AeYc E7B 31$ "   GE7B 31   V X W f d 9 875 3 $ "  9 875 3   V X W V  R @FF641 %#! )(F641 ba`YT!UTSQP

9 875 31 @(FI4H GE7B 31 !FDC4A$ 9 875 31 @(4642$ 9 875 3 qsq @(46S(2U

GE7B 3 qsq !FDCS(2i$

9 875 3 qsq )(F6SF2r$

0 0

Broadcast Trafc: The IEEE 802.11 broadcast trafc, the sender listens briey to the channel and sends a message if the channel is clear. Dening the xed channel access cost and , and the incremental payload costs and , Equation 1 gives:

$ $ "     %#! $    )('&

TABLE II IEEE 802.11 2 M BPS WAVELAN PC CARD L INEAR M ODEL P OWER C ONSUMPTION M EASUREMENTS

the problem of saving power which is complementary to the other approaches. Solutions addressing power savings in the routing and MAC layer include the following:

Energy

Active
Idle

Critical time

Active

Sleep

Ttrans

Time

Ttrans

Fig. 1. State Transition Diagram from Idle to Sleep

latency required to switch from one mode to another. While switching between active (transmit or receive) and idle mode does not take any signicant time or energy, transition to sleep mode is of particular interest. Figure 1 shows that there is a minimum critical time, only over which a state transition to sleep mode is useful. amount of time is used by a network interface to go from idle to sleep mode and vice versa. During this time interval, no packet can be transmitted or received by the network interface, rendering it useless during this time. There is a need to minimize this useless amount of time as much as possible. For example, in the ORiNOCO World PC Card from Agere Systems [18], the time is specied to be less than 75 sec. This implies that if 1% of the time is allowed for this transition, the sleep period has to be at least equal to 10 ms. No extra energy is required to make this transition, according to the specication of the ORiNOCO card. III. R ELATED W ORK A critical issue for sensor nodes is power. Though access to data has become ubiquitous, access to power is a major constraint. A lot of research has been done in the area of low power. Low power is particularly important in sensor and mobile wireless ad hoc networks. Various attempts at saving power have been proposed, from the application layer[9] to the physical layer [16]. While significant power can be saved in all of these layers, we tackle

0 0 0

)(u   )( )(   ! 58 @AB P

Operation broadcast send broadcast recv p2p send p2p send

( J/byte)

( J) 266 56 454 356

Transmission Power Control: In wireless communication, controlling transmission power has a large impact, ranging from power savings to transmission rate, interference, and error rates. Different work in this area has focused on topology control, determination of optimum transmission power, and increasing network throughput [20], [13], [1], [10], [17], [7], [21]. Power-Aware Routing: In this area, the chief concern is extending the lifetime of a network. This work takes into account the remaining battery level of a node in determining optimum routes, such that none of the nodes are over utilized. The paper by Li et al [10] is a recent example of work in this area. This research area does not explicitly try to reduce power usage, but tries instead to maintain a uniformity in the power usage across all nodes, so that the network lifetime is extended. Low Power Mode: In this area of research , nodes are kept in one of the available power modes [25], [15]. Each node tries to be in the lowest power mode possible and still maintain connectivity in the network. The work by Xu et al. [23], [22] propose a grid-based energy-saving routing protocol, but it uses GPS for location information. Chen et al. [2] proposed an algorithm where some hosts serve as coordinators, which are chosen according to their remaining battery power, and the number of neighbors they can connect to. Coordinators are responsible for relaying packets and are kept awake, while other nodes can enter sleep mode. The work by Ye et. al [24] focus on building a low-power MAC protocol called S-MAC. In our work, we can take full advantage of a low-power MAC layer other than IEEE 802.11. So, this work is complementary to our power saving scheme. Work by Tseng et al [19] is the closest to our work. It proposes a few randomized algorithms for changing to sleep mode. But they focus on a tradeoff between neighbor discovery time and energy consumption, and not on network protocol performance under these conditions. In this paper, we tackle the overall protocol performance showing the affect of reduction of power usage on latency. Their protocol is not adaptive depending on the neighbor density, and hence cannot take advantage of high node density. In our work, we adaptively vary the energy usage of a node, thereby using the minimal energy required.

58 @AB P

McGlynn et al [11] use certain probabilistic methods similar to ours, for neighbor discovery. Time synchronization among nodes in a network is a problem several papers have addressed [3], [12]. IV. A LGORITHMS
FOR

S LEEP -M ODE S CHEDULING

B. Birthday Sleep Algorithm The Birthday Sleep Algorithm is proposed in this paper for scheduling transition to sleep mode for saving energy without losing much in terms of performance. B.1 Probability Basis The inspiration behind this protocol is the Birthday Paradox. The Birthday Paradox is the probability that at least two people in a room have their birthday on the same day. The paradox is that with as few as 23 people in the room, the probability exceeds one half. This is called a paradox, as it is counter-intuitive that with as few as 23 people and 365 possible days, the probability that any pair of people have the same birthday is more than the probability that no pair of people have the same birthday. This idea has been applied for neighbor discovery by McGlynn et al. [11]. We use a similar idea for convergence between a receiver and a transmitter. Over a period of xed length slot periods, two wireless nodes independently and randomly select of these slots. In each of these slots chosen by a node, the node remains in idle mode listening for packets. In the remaining slots, it goes to sleep mode, thereby saving energy. The energy saved is approximately percent of the original energy consumed by the wireless network interface. So, the energy saved has a linear relation with the percentage of the idle slots, evaluated later in Section V. We calculate the probability that both the nodes are in idle mode in the same slot, for at least one slot out of the possible slots. Let the two nodes be A and B. The number of ways A can choose out of things is

A. Extension to the IEEE 802.11 Power Saving Protocol The IEEE 802.11 MAC protocol [8] has support for a power-saving (PS) mode. In an infrastructure-based network, there is an access point (AP) that communicates data packets with each node. Periodically, the AP transmits a beacon frame spaced by a xed beacon interval. Each beacon frame contains a Trafc Indication Map (TIM). The TIM frame contains identities of PS hosts for which the AP is holding buffered packets. A PS node monitors the TIM frames and remains awake for the rest of the interval if it is listed in the TIM; otherwise the node goes to sleep. In IEEE 802.11 ad hoc mode, PS nodes behave similar to infrastructure mode, except that all the nodes contend for sending a beacon, compared to only the AP in infrastructure mode. The Ad hoc Trafc Indication Map (ATIM) frame is sent by each node having buffered packets. The ATIM frame contains information about the destination for the buffered packets. The destination node receiving this ATIM frame stays awake for the rest of the period, and otherwise goes into sleep mode. The problem with the PS mode of IEEE 802.11 is that it was designed for a single-hop (fully connected) network. When applied to sensor or ad hoc networks, certain problems arise. Tseng et al. [19] have recently analyzed the problems:

 r W

Clock Synchronization: IEEE 802.11 PS mode assumes a completely connected network, and hence a beacon frame transmission can synchronize all the nodes. In a sensor network though, this sim-

 S

 U

All the nodes in a sensor network do not always participate in receiving, sending, and forwarding of data packets. Rather, many nodes do not need to be in the topology to maintain connectivity. If these nodes can be identied, then they can be put into sleep mode, rather than them staying in idle mode. Sleep mode uses an order of magnitude less energy than idle mode, so this would be very useful in terms of saving energy. The nodes that are probable candidates for putting into sleep mode are those that have not originated, forwarded, or received data for a certain xed interval. These nodes are then sent to the sleep mode. However, these nodes are present in the network and can work as forwarders if the need arises. Being in sleep mode prevents a node from being able to receive any packets. So, there is a need to use some technique to periodically wake these nodes to idle mode.

ple clock synchronization is not possible because of communication delays and mobility. Clock Synchronization among all nodes in a sensor network is a problem that has been tackled in recent years. GPS can be used to provide exact time synchronization between the nodes. In the absence of GPS, clock synchronization can be ensured by some protocol. The work by Elson et al. [3] is a recent example of research in the domain of time synchronization. Neighbor Discovery: A wireless node is aware of the existence of neighbors only if there is an ongoing transmission while the node is awake. As PS mode reduces the period of being awake, it diminishes its chances of having accurate neighbor information as well as the chance of itself being discovered by other nodes. In Section IV-B.3, we describe why very accurate prediction of the number of neighbors is not necessary.

(2)

The number of ways B can choose

out of

P(At least 1 common slot)

The total number of ways A and B can choose each is

The number of ways at least 1 out of by B are same as A is

The probability that A and B have at least one idle mode slot in common is

(7)

The number of ways of B choosing slots out of is

Bounded Delay Along with providing a probabilistic guarantee for a packet from a node to reach its neighbor as detailed above, we also want to provide probabilistic guarantees for the delay bound. If the total probability of there being a common slot is calculated, it might be the case that the common slot is the last of the slots. In that case, there would be a signicant amount of latency involved in trying to send a packet to a neighbor. So, the probability of the common slot being within the rst of

This function is plotted for in Figure 3(a) and in Figure 3(b). Although the total probability might have been higher for larger , the probability of a matched slot within slots is not helped by larger . In fact, smaller give a larger probability that that there is a matching slot within lower values of . We chose the value of to be 10 instead of 50 for this reason. For , , there is a 93% probability that the slots of the two nodes are together alive within rst 2 alive slots. B.2 The Birthday Algorithm In this section, we take advantage of the conclusion of Section IV-B.1, where we showed the effect of choosing

 `  W W

 

 

For , the probability is 99.9% for . So, the amount of saved energy is %. If , the probability is 99.9% for . Hence, the amount of saved energy is % . This suggests that making larger is going to be benecial for saving energy.

The probability of B choosing at least 1 slot the same as those chosen by A, within slots is

@   W W

The probability that the two nodes will have a common slot increases with the increase in the ratio . Energy is saved during the slots where the receiver is kept in sleep mode.

The probability of choosing those chosen by A is

 W

Figure 2 shows the plot of this probability function for increasing , for various values of . These results show the following properties:

slots all different than (11)

(8)

(10)

(12)

 

s r ` 4  4  W  W  W

4 W 4    W   W

  

4    W  W



 4   ug  W  W 

   

The number of ways same as A is

out of

slots chosen by B are (5) slots chosen

things is (3) things (4)

1 0.8 0.6 0.4 0.2 0 0 5 10 m n=10 n=20 n=30 n=40 n=50 15 20

 s 4 W

 r W

Fig. 2. Probability of a matching slot

(6)

positions is analyzed. If a specic delay bound is needed by an application, he can choose a specic value of to be assured delivery of the packet within the delay bound with a probabilistic guarantee. To show that this probabilistic delay bound can be guaranteed, the earlier analysis has to be extended. Node out of slots. Node B chooses slots, A chooses where , and each of the slots is different than the slots chosen by A. The number of ways of B choosing slots out of is (9)

 4 W
 %

e 


0 0  H

1
P(At least 1 common slot)

1
P(At least 1 common slot)

0.8 0.6 0.4 0.2 0 1 2 3 4 5 k 6 7 m=10 m=20 m=30 m=40 8 9 10

0.8 0.6 0.4 0.2 0 1 2 3 4 5 k 6 7 m=5 m=6 m=7 m=8 m=9 m=10 8 9 10

Fig. 3. Probability of a matching slot within rst

slots

SIM sent out SIMs received

SIM sent out SIMs received

Time
Fig. 4. The Birthday Algorithm Strategy

If the policy is to save energy on a per node basis, then is set to a constant value, for . If the policy is to maximize the lifetime of a network, then may be set to a function of the remaining energy of , for , or a function of the estimate of the neighbor count.

     

   

For all other nodes,

depends on the policy.

All nodes are awake for the rst slot out of slots. Depending on , each node randomly chooses out of the slots to be idle. Each node builds a Sleep Indication Map (SIM) containing bits, with the bit set to 1 or 0 depending on whether the node will be idle or asleep in the slot of that period. At the beginning of every slots taking time in total, the SIM is sent out by each node in a local broadcast. Hence, every node has knowledge of the sleep and idle slots for all of its neighbors. Figure 4 shows how this algorithm works. A node uses the knowledge gained from the SIM packets from its neighbors to build a Sleep Indication Table (SIT). An example SIT is shown in Figure 5, where each row of the SIT is the SIM received from one of that nodes neighbors. A node uses its SIT to schedule packets to any neighbor. The specic policies it uses for scheduling are as follows: 1. A node schedules a broadcast by nding the earliest slot with maximum number of neighbors in idle mode. Then that broadcast will have the maximum intended impact.

5 5 455 455 44 44

53 455 233 44 22

31 1 33 012 1 011 22 00 00

F G

) ) ()) ()) ( (( (( ((

' 'A &'' &'' @AA && && @@

A % & @AA @ $%% && @@ @@ $$

% # $%% $ "## $$ $$ ""

# # "## "## "" ""

9 9 ! 899 899 !! 88 88

88

! 67 67 67 DE DE D 8 !! 6767 6767 6767 DDEE DDEE DD

F G

the probability with which a node stays in sleep mode. Using the strategies of clock synchronization stated previously, the synchronization of the clocks of all nodes in a sensor network can be achieved. For a particular value of , where is the number of slots in a time interval , a value of is chosen. is the number of slots in which a node is in idle mode listening for packets. The performance in terms of delay latency will be lower compared to the case in which the node is always in idle mode but the gain is in terms of energy saved by remaining in sleep mode. We can also give a probabilistic guarantee on the delay bound, by choosing the as detailed in the last section. The lower the ratio of , the larger the energy saving with a consequent lower performance. Hence, this ratio is a tunable parameter depending on the need for saving energy versus better performance. Let denote the probability of staying awake for , where is the number of slots is awake and is the total number of slots. If a node is a sender or receiver of data or if it is an active forwarder of data, then the node always remains awake. Hence, , for where = Set is the set of all sources at the current time, is the set of all destinations at the current time, and is the set of all forwarders at the current time.

n = 10 m=6

n slots (T time)

n slots (T time)

CBC

S  

) X

S

(a)

= 50

(b)

= 10

Idle Mode Sleep Mode

    

Neighbor 1 Neighbor 2 Neighbor 3 Neighbor 4 Neighbor 5

There is no need for accurate count of the number of neighbors, since the algorithm provides a natural feedback mechanism. If the neighbor count is underestimated by a node, that node stays awake for a longer time, during which it hears from more neighbors, and subsequently reduces its probability of staying awake. C. The Adaptive Sleep Algorithm The Adaptive Sleep Algorithm is a protocol for energy conservation similar to GAF [23]. It is more simplistic as it does not have any geographic information available. The algorithm is explained in Figure 6. A interval called Sleep Interval (SI) is the period of time from when the transceiver of the node is idle. It is a parameter tunable by the application, based on the knowledge of the frequency of queries or periodic interest by any data sink

1. Estimate Neighbor Count: There are a number of published approaches for estimating the neighbor count of a node in a wireless network. In this work, a simple approach is used, where each node keeps a list of all the nodes from which it has received a packet whenever it is listening. Each entry in the list is deleted after a period of time , unless the node receives another packet from that node. This list is purged of all stale entries, which have not been updated in time. In this way any extra overhead of neighbor discovery messages is avoided. 2. Estimate of probability of staying awake: The idea is to keep the sum total idle time for all nodes in a neighborhood constant. At any point of time in any neighborhood, we try to make the probability of nding the same number of idle nodes as equal. In this way, the whole region is covered uniformly. Nodes in the high node density region can be asleep for more number of slots than nodes in a low density

This re-calculation of sired.

is done as frequently as de-

d s

Solving for

, (15)

d e

 a  d

In densely populated sensor networks, many nodes can be interchangeably used for routing purposes. In a dense network, any node might sleep for a longer duration, as the relative utility (for forwarding purposes) of all nodes decrease. We use an approach similar to that used by Xu et al. [22].

B.3 Dynamic Adaptation

2. A node schedules the delivery of a unicast packet by looking for the nearest idle slot of the intended next hop neighbor. 3. If the information for a next hop node is not present in the SIT, then the source node sends the packet immediately.

where is the exact theoretical neighbor count of a particular in the neighborhood, is the for staying awake, and is a probability of constant. But this optimal solution cannot be implemented in practice because nodes do not have a an exact idea of their neighborhood in a sensor or ad hoc network. Nodes can fail or move at any the time. To achieve an approximation of this optimal solution, the estimate of the neighbor count is used by estimate its each node as specied above. Let neighbor count to be . Then this node will calculate its from the following equation

X X

Fig. 5. Example of a Sleep Indication Table (SIT)

 

hg f

( GH ' X ( GH W F B B D % ' H XX EF AAB CD &( ( GHG WW FE BBA DC %' ' 6 X E AA C & & 9@GH 5 WWX FEF BBAB DCD %'( @GH 8 6 4 E AA C # 5 3WX EF AB CD %%&%&'( @9@ 778 56 34 112 6 4 2 99 7 5 43 1 V S R 1 0 T 3 2 ) 0 $ U 8 6 Q 2 ) ) @@ 778 56 44 112  & 99 87 5 33 P1 V 778 T R I12  0 $ U9@ 7 56 Q34 21 #%& VV ST RR 1IP 0 $$ UU 8 TS QQ P12I ) ##$%&# "@ 778 S6 4 IIP !UV9 7 ST5 QR3 I1 )0 "  P    !   IIP    UV ST QR I )0 #$ "!"!"      !     b       b !"     a  bb  a  ababa a
5 Neighbors

1 2

7  q P 9

c d

7  2q P 9

Y Y` `YY
3

Sleep Mode Idle Mode n slots 4 5 6 7 8 n = 10 m=6 9 10

Duty Cycle Duty Cycle SP

Energy
SI Idle Mode SP

Sleep Mode

Time

SI = Sleep Interval SP = Sleep Period

Fig. 6. Adaptive Sleep Mode Strategy

region. To that effect, the optimal solution would be (13)

(14)

c e

B. Results for the Birthday Sleep Algorithm Results In this experiment, the total number of slots was 10. The number of slots awake randomly, , was varied from 5 to 10. This implies an idle percentage of 50% to 100% of total time. Figure 7(a) shows the decrease in energy usage with decreasing values of . The graph is linear as expected, because the energy savings was linearly related to the number of slots asleep. The decrease in energy consumption occurs because a smaller values of means a larger time spent in sleep mode, thereby saving more energy. For 50% of the slots being idle, an energy saving of 46% is obtained This is a substantial improvement in terms of energy savings. Figure 7(b) shows the latency as a function of . The latency has a exponential increase with decrease in the value of . We plot the latency versus the energy savings given by our approach in Figure 8. Guaranteeing probabilistic delay bound as described earlier is shown in this graph. For a specic latency bound, the maximum energy saving that can be achieved solely by this algorithm in given by the x-axis in Figure 8. C. Results for the Adaptive Sleep Algorithm Results To evaluate the Adaptive Sleep Algorithm, the duty cycle was varied from 10% of the Sleep Period (SP) to 90% of the SP, at 10% intervals. Figure 9(a) shows the energy usage for different values of the duty cycle. There is a linear decrease in energy usage for decreasing duty cycle. This is because smaller duty cycles mean more time spent in sleep mode and therefore more energy savings. The percent energy savings is maximum when the duty cycle is as low as possible. A 50% duty cycle resulted in energy savings of 30%, while a 10% duty cycle resulted in energy savings

V. S IMULATION R ESULTS This section presents the results of the simulation of the Adaptive Sleep Algorithm and the Birthday Sleep Algorithm. To evaluate the protocol, we modeled it in the ns-2 network simulator [4], which is a popular simulator for use in the networking community. The energy model described in Section II was incorporated into the simulator for modeling the energy usage of the nodes. A. Simulation Parameters The simulation parameters were chosen to be similar to those used in the sensor and ad hoc networking research community, so that comparison of the simulation results is easier. The specic parameters were: A xed area of meters was used in the simulation. The experiments were run for 900 simulated seconds each. The communication used was 10 Constant Bit Rate (CBR) ows between 10 randomly chosen pairs of nodes. Each ow consisted of 4 packets per second, each packet consisting of 512 bytes. The bandwidth of the wireless medium was set to be 2 Mbps.

node. After SI time interval, the periodic sleep cycle starts. The periodic sleep cycle repeats at a period called the Sleep Period (SP). It also has a variable Duty Cycle (D), which is the fraction of the SP period that the node is in idle mode. For fraction of SP, the node remains in sleep mode. The Duty Cycle is changed according to the same principle as detailed in Section IV-B.3. An estimate of neighbor count is used to vary the duty cycle. The intuition is to keep the total number of nodes idle at any time to be constant across the region. So, in a high node density area, each node can have a lower duty cycle than the nodes in a low node density region. Anytime a sender in sleep mode needs to send data, it transitions to active mode and starts sending. When a node is in sleep mode, if a packet is transmitted by a neighbor node, it will not be able to receive it. Depending on the length of time that the node is in sleep mode, it might wake up in time to receive the subsequent rebroadcasts of the messages by the nodes neighbors. Hence, the SP needs to be small enough to enable a node to receive subsequent broadcasts after having missed an initial packet. This problem of re-broadcast is not present if the IEEE 802.11 PS MAC protocol is used. The upstream nodes will buffer the packets and only transmit it when the downstream node is alive. A detailed simulation analysis of the performance of this protocol in comparison with our Birthday Sleep Protocol is done in the next section.

0.023 "lat_energy" 0.022 0.021 0.02 0.019

Latency

0.018 0.017 0.016 0.015 0.014 0.013 0 5 10 15 20 25 30 Percentage Energy Savings 35 40 45 50

Fig. 8. Latency versus Percent Energy Saving in Birthday Sleep Algorithm

Each data point in the graphs are taken from an average of 3 runs, each run having a different communication pattern.

0 0 0

50 "4pkt_energy" 45 40 35
% Energy saving

0.023 "4pkt_lat" 0.022 0.021 0.02 0.019 0.018 0.017 0.016 0.015 0.014 0.013 5 6 7 8 No. of slots awake randomly 9 10 5 6 7 8 No. of slots awake randomly 9 10

25 20 15 10 5 0

(a) Percent Energy Saving Fig. 7. Birthday Sleep Algorithm

Latency(secs)

30

(b) Delay Latency

of over 50%. This is a signicant improvement in terms of energy savings. Figure 9(b) shows the latency as a function of duty cycle. The latency has a linear increase with decreasing duty cycle. The latency doubles for a duty cycle of 10% when compared to the node being always idle. The latency increases from 12 ms to about 25 ms for a duty cycle of 10%. The slope of the line is not exactly -45 degree, because energy is also expended in transmitting and receiving packets in addition to being in idle mode. VI. C ONCLUSION This paper makes the following contributions in the area of power mode scheduling in ad hoc networks: We also proposed an extension to IEEE 802.11 Power Saving (PS) mode for multi-hop wireless networks. Using it, we proposed and modeled a probabilistic algorithm for scheduling transitions between idle and sleep modes. We then simulated this algorithm to show signicant energy savings. The novelty of our approach lies in the fact that we have applied a well known technique from probability theory for energy savings in sensor networks. Our approach is orthogonal to many other approaches for energy saving in this area, and can provide additive benet. We also give a probabilistic delay bound guarantee, as a trade-off with energy savings R EFERENCES
[1] [2] Jae-Hwan Chang and Leandros Tassiulas. Energy Conserving Routing in Wireless Ad-hoc Networks. In Proceedings of INFOCOM 2000, pages 2231, Tel-Aviv, Israel, March 2000. Benjie Chen, Kyle Jamieson, Hari Balakrishnan, and Robert Morris. Span: An Energy-Efcient Coordination Algorithm for Topology Maintenance in Ad Hoc Wireless Networks. In Proceedings of the Seventh Annual International Conference on Mobile Computing and Networking (MobiCom 2001), pages 85 96, Rome, Italy, July 2001. Jeremy Elson, Lewis Girod, and Deborah Estrin. Fine-Grained Network Time Synchronization using Reference Broadcasts. In

[4]

[5]

[6]

[7]

[8]

[9]

[10]

[11]

[12]

[13]

[14]

[3]

Proceedings of the 5th Symposium on Operating Systems Design and Implementation, Boston, Massachusetts, December 2002. Kevin Fall and Kannan Varadhan, editors. ns Notes and Documentation. The VINT Project, UC Berkeley, LBL, USC/ISI, and Xerox PARC, November 1997. Available from http://wwwmash.cs.berkeley.edu/ns/. Laura Feeney and Martin Nilsson. Investigating the Energy Consumption of a Wireless Network Interface in an Ad Hoc Networking Environment. In Proceedings of INFOCOM 2001, volume 3, pages 15481557, Anchorage, Alaska, April 2001. Laura Marie Feeney. An Energy Consumption Model for Performance Analysis of Routing Protocols for Mobile Ad Hoc Networks. Mobile Networks and Applications, 6(3):239249, June 2001. Javier Gomez, Andrew T. Campbell, Mahmoud Naghshineh, and Chatschik Bisdikian. Conserving Transmission Power in Wireless Ad Hoc Networks. In Proceedings of International Conference on Network Protocols, pages 2434, Riverside, California, November 2001. IEEE Computer Society LAN MAN Standards Committee. Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specications, IEEE Std 802.11-1999. The Institute of Electrical and Electronics Engineers, New York, New York, 1999. Robin Kravets and P. Krishnan. Application-Driven Power Management for Mobile Communication. Wireless Networks, 6(4):263277, July 2000. Qun Li, Javed Aslam, and Daniela Rus. Online Power-aware Routing in Wireless Ad-hoc Networks. In Proceedings of the Seventh Annual International Conference on Mobile Computing and Networking (MobiCom 2001), pages 97107, Rome, Italy, July 2001. Michael J. McGlynn and Steven A. Borbash. Birthday Protocols for Low Energy deployment and Flexible Neighbor Discovery in Ad Hoc Wireless Networks. In Proceedings of the Second ACM Symposium on Mobile Ad Hoc Networking and Computing (MobiHoc 2001), pages 137145, Long Beach, California, October 2001. David Mills. Internet Time Synchronization: the Network Time Protocol. IEEE Transactions on Communications, 39(10):1482 1493, October 1991. Jeffrey Monks, Vaduvur Bharghavan, and Wen-Mei Hwu. A Power Controlled Multiple Access Protocol for Wireless Packet Networks. In Proceedings of INFOCOM 2001, volume 1, pages 219228, Anchorage, Alaska, April 2001. Santashil PalChaudhuri and David B. Johnson. Power Mode Scheduling for Ad Hoc Networks. In Proceedings of the 10th IEEE International Conference on Network Protocols (ICNP02), Paris, France, November 2002.

10

55 "4pkt_energy" 50

0.026 "4pkt_lat"

0.024 45 40
% Energy saving

0.022

Latency(sec)

35 30 25 20 15

0.02

0.018

0.016

0.014 10 5 10 20 30 40 50 Duty Cycle 60 70 80 90 0.012 10 20 30 40 50 Duty Cycle 60 70 80 90

(a) Percent Energy Saving Fig. 9. Adaptive Sleep Algorithm

(b) Delay Latency

[15] Curt Schurgers, Vlasios Tsiatsis, Saurabh Ganeriwal, and Mani Srivastava. Optimizing Sensor Networks in the Energy-LatencyDensity Design Space. IEEE Transactions on Mobile Computing, 1(1):7080, Jan-March 2002. [16] Eugene Shih, Seong-Hwan Cho, Nathan Ickes, Rex Min, Amit Sinha, Alice Wang, and Anantha Chandrakasan. Physical Layer Driven Protocol and Algorithm Design for Energy-Efcient Wireless Sensor Networks. In Proceedings of the Seventh Annual International Conference on Mobile Computing and Networking (MobiCom 2001), pages 272287, Rome, Italy, July 2001. [17] Suresh Singh, Mike Woo, and C. S. Raghavendra. PowerAware Routing in Mobile Ad Hoc Networks. In Proceedings of the Fourth Annual ACM/IEEE International Conference on Mobile Computing and Networking (MobiCom98), pages 181 190, Dallas, Texas, October 1998. [18] Agere Systems. ORiNOCO World PC Card Data Sheet, February 2002. http://www.orinocowireless.com. [19] Yu-Chee Tseng, Chih-Shun Hsu, and Ten-Yueng Hsieh. PowerSaving Protocols for IEEE 802.11-based Multi-Hop Ad Hoc Networks. In Proceedings of INFOCOM 2002, New York, New York, June 2002. [20] Roger Wattenhofer, Li Li, Paramvir Bahl, and Yi-Min Wang. Distributed Topology Control for Power Efcient Operation in Multihop Wireless Ad Hoc Networks. In Proceedings of INFOCOM 2001, volume 3, pages 13881397, Anchorage, Alaska, April 2001. [21] Hagen Woesner, Jean-Pierre Ebert, Morten Schlger, and Adam Wolisz. Power-Saving Mechanisms in Emerging Standards for Wireless LANs: The MAC Level Perspective. IEEE Personal Communications Magazine, 5(3):4048, June 1998. [22] Ya Xu, John Heidemann, and Deborah Estrin. Adaptive EnergyConserving Routing for Multihop Ad Hoc Networks. Technical Report TR 2000-527, October 2000. [23] Ya Xu, John Heidemann, and Deborah Estrin. Geographyinformed Energy Conservation for Ad Hoc Routing. In Proceedings of the Seventh Annual International Conference on Mobile Computing and Networking (MobiCom 2001), pages 70 84, Rome, Italy, July 2001. [24] Wei Ye, John Heidemann, and Deborah Estrin. An EnergyEfcient MAC protocol for Wireless Sensor Networks. In Proceedings of INFOCOM 2002, New York, New York, June 2002. [25] Rong Zheng, Jennifer Hou, and Lui Sha. Asynchronous Wakeup for Power Management in Ad Hoc Networks. In Proceedings of MobiHoc 2003, Annapolis, Maryland, June 2003.