The MicroPulse Framework for Adaptive Waking Windows in Sensor Networks

Transcription

1 The MicroPulse Framework for Aaptive Waking Winows in Sensor Networks Demetrios Zeinalipour-Yazti, Panayiotis Anreou, Panos K. Chrysanthis, George Samaras, Anreas Pitsillies Department of Computer Science, University of Cyprus, P.O. Box 20537, 1678 Nicosia, Cyprus Department of Computer Science, University of Pittsburgh, Pittsburgh, PA 15260, USA Abstract In this paper we present MicroPulse, a novel framework for aapting the waking winow of a sensing evice S base on the ata workloa incurre by a query Q. Assuming a typical tree-base aggregation scenario, the waking winow is efine as the time interval τ uring which S enables its transceiver in orer to collect the results from its chilren. Minimizing the length of τ enables S to conserve energy that can be use to prolong the longevity of the network an hence the quality of results. Our metho is establishe on profiling recent ata acquisition activity an on ientifying the bottlenecks using an in-network execution of the Critical Path Metho. We show through traceriven experimentation with a real ataset that MicroPulse can reuce the energy cost of the waking winow by three orers of magnitue. Inex Terms Critical Path Metho, Waking Winow, Scheuling, Sensor Networks. I. INTRODUCTION Recent avances in embee computing have mae it feasible to prouce small scale sensors, actuators an processors that can be use for a-hoc eployments of environmental monitoring infrastructures [13], [5], [9]. The longevity of a Wireless Sensor Network (WSN) heavily relies on the existence of power-efficient algorithms for the acquisition, aggregation an storage of the sensor reaings. Communicating over the raio in a WSN is the most energy emaning factor among all other functions, such as storage [17] an processing [9]. The energy consumption for transmitting 1 bit of ata using the MICA mote is approximately equivalent to processing 1000 CPU instructions [9]. One way to cope with the energy challenge is to power own the raio transceiver uring perios of inactivity. In particular, it has been shown that sensors operating at a 2% uty cycle can achieve lifetimes of 6-months using two AA batteries [8]. The continuous interval uring which a sensor noe S enables its transceiver, collects an aggregates the results from its chilren, an then forwars them all together to its own parent is efine as the waking winow τ. Note that τ is continuous because it woul be very energy-emaning to suspen the transceiver more than once uring the interval of an epoch, which specifies the amount of time that sensors have to wait before re-computing a continuous query. It is important to mention that the exact value of τ is queryspecific an can not be etermine accurately using current techniques. For instance a sensor oes not know in avance how many tuples it will receive from its chilren. Choosing the correct value for τ is a challenging task as any wrong estimate might isrupt the synchrony of the query routing tree. The objective of this work is to automatically tune τ, locally at each sensor without any a priori knowlege or user intervention. Note that in efining τ we are challenge with the following trae-off: Early-Off Transceiver: Shall S power-off the transceiver too early reuces energy consumption but also increases the number of tuples that are not elivere to the, the root of the routing tree. As a result the will generate an erroneous answer to the query Q; an Late-Off Transceiver: Shall S keep the transceiver active for too long ecreases the number of tuples that are lost ue to powering own the transceiver too early but also increases energy consumption. Thus, the network will consume more energy than necessary which is not esirable given the scarce energy buget of each sensor. In this paper we present MicroPulse, a novel framework for aapting the waking winow of a sensing evice S base on the ata workloa incurre by a query Q. Our ieas are establishe on profiling recent ata acquisition activity an on ientifying the bottlenecks using an in-network execution of the Critical Path Metho. The Critical Path Metho (CPM) [10] is a graph-theoretic algorithm for scheuling project activities. It is wiely use in project planning (construction, prouct evelopment, plant maintenance, software evelopment an research projects). The core iea of CPM is to associate each project milestone with a vertex v an then efine the epenencies between the given vertices using activities. For instance, the activity v i v j enotes that the completion of v i epens on the completion of v j. Each activity is associate with a weight (enote as weight ) which quantifies the amount of time that is require to complete v i assuming that v j is complete. The critical path allows us to efine the minimum time or otherwise the maximum path that is require to complete a project (i.e.,

2 SINK (s0) 40 s1 s2 s s5 s Critical Path s0<=s1<=s3<=s8 U= s4 4 s7 s8 s9 Fig. 1. The figure illustrates nine sensing evices (shown as vertices) an the respective time cost (shown as eges) to transfer the results from an arbitrary an continuous query Q to the (s 0 ). MicroPulse utilizes this information in orer to locally aapt the waking winow of each sensor using the Critical Path Metho. milestone v 0 ). Any elay in the activities of the critical path will cause a elay for the whole project. In orer to aapt the iscussion to a sensor network context assume that each sensor s i is represente by a CPM vertex. More formally, we map each s i to the elements of the vertex set V = {v 1, v 2,..., v n } using a 1:1 mapping function f : s i v i, i n. Also, let the escenent-ancestor relations of the sensor network be enote as eges in this graph. An example with 9 sensing evices {s 1,...,s 9 } is illustrate in Figure 1. The weights on the eges efine the require time to propagate the query results between the respective pairs. It is easy to see that the total time to answer the query at the in the given network is at least ψ = 99, since the critical path is s 0 s 1 s 3 s 8. Having this information at han enables the scheuling of transmission between sensors. In particular, sensor s i can be scheule to wake-up an transmit at the following ealines (w i ): w 1 = ψ 40 = 59, w 2 = w 1 13 = 46, w 3 = w 1 30 = 29, w 4 = w 1 22 = 37 while s 0 an s 1 will be listening for these transmissions uring the intervals τ 0 =[59..99) an τ 1 =[29..59) respectively. The same intuition also applies for the leaf noes, e.g., s 5 starts transmission at w 5 = w 2 11 = 35 an s 2 listens for this transmission in the range τ 2 =[35..46). Aitionally, the critical path enables a sensor x to estimate the interval uring which its parent y will have its transceiver enable. This is very useful because in the subsequent epochs an uner a ifferent workloa, x can fin out if it can eliver the new workloa without first asking y to ajust its waking winow. It shoul be note that the eges in Figure 1 have ifferent weights. This is very typical for a sensor network as the link quality can vary across the network [13]. Another reason is that some sensors might return more tuples than other sensors. Note that our scheuling scheme is istribute which makes it funamentally ifferent from centralize scheuling approaches like DTA [16] an TD-DES [1] that generate collision-free query plans at a centralize noe. Aitionally, our approach is also ifferent from techniques such as [12] which segment the sensor network into sectors in orer to minimize collisions uring ata acquisition. Our Contributions In this paper we make the following contributions: We formulate the problem of aapting the waking winow τ of a sensing evice in orer to conserve energy that can be use to prolong the longevity of the network an hence the quality of results. We solve the waking winow problem by combining a custom profiling structure an the critical path metho. We experimentally valiate the efficiency of our solution using real sensor ataset from Intel Research [6]. The remainer of the paper is organize as follows: Section II stuies the waking winow mechanism of popular ata acquisition systems an iscusses the avantages an shortages for each of these systems. Section III presents the unerlying algorithms of the MicroPulse Framework. Section IV presents the experimental stuy using a trace-riven simulator an Section V conclues the paper. II. BACKGROUND AND RELATED WORK In this section we stuy the waking winow mechanism of the two most popular eclarative acquisition frameworks: TAG [9], [8] an Cougar [15]. We start out our escription by assuming that the query Q has been isseminate to n sensors. Tiny Aggregation (TAG): In this approach, the epoch e is ivie into a number of fixe-length time intervals {e 1, e 2,..., e }, where is the epth of the routing tree, roote at the, that conceptually interconnects the n sensors. The core iea of this framework is summarize as follows: when noes at level i+1 transmit then noes at level i listen. More formally, a sensor s i enables its transceiver at chronon w i = e/ ( epth(s i )) an keeps the transceiver active for τ i = e/ chronons. Note that 0 i= (e i) provies a lower-boun on e, thus the answer will always arrive at the before the en of the epoch. Setting e as a prime number ensures the following inequality 0 i= (e i) < e, which is esirable given that the answer has to reach the at chronon e. For instance, if the epoch is 31 secons an we have a three-tiere network (i.e., =3) like Figure 2 (top, left), then the epoch is slice into three segments {10,10,10}. During interval [0..10), noes at level 3 will transmit while noes at level 2 will listen; uring interval [10..20) level 2 noes transmit while level 1 noes listen; an finally uring [20..30), level 1 noes transmit an the (level 0) listens. Thus, the answer will be reay prior the completion of chronon 31 which is the en of the epoch. The parent wake-up winow τ is an over-estimation (in the above example 10 secons!) of the actual time that is require to transmit between the chilren an a parent. The rationale behin this over-estimation is to offset the limitations in the quality of the clock synchronization algorithms [9] but in reality it is too coarse. In the experimental Section IV, we foun that this over-estimation is three orers of magnitue larger than necessary.

3 Fig. 2. Level 0 Level 1 Level 2 Level 3 Level 0 Level 1 Level 2 Level 3 TAG Micropulse Cougar 0 e 0 e 0 e Listening Processing Transmitting The Waking (Listening) Winow (τ) in TAG, Cougar an MicroPulse. Aitionally, it is not clear how τ is set uner a variable workloa which occurs uner the following circumstances: i) from a non-balance topology, where some noes have many chilren an thus require more time to collect the results from their epenents; an ii) from multi-tuple answers, which are generate because some noes return more tuples than other noes (e.g. because of the query preicate). The MicroPulse framework presente in this paper gracefully hanles both cases of variable workloa by utilizing the Critical Path Metho. Our framework, like TAG, utilizes the TinyOS MAC layer [14] to hanle the collisions that will occur if noes in the same vicinity transmit uring the same interval. Cougar: In this approach, each sensor maintains a waiting list that specifies the chilren for each noe. Such a list can be constructe by having each chil explicitly acknowleging its parent uring the query issemination phase. Having the list of chilren enables a sensor to shut own its transceiver as soon as all chilren have answere. This yiels a set of non-uniform waking winows {τ 1, τ 2,...} as oppose to TAG where we have a single τ which is uniform for all sensors (i.e., e/ ). The main rawback of Cougar is that a parent noe has to keep its transceiver active from the beginning of the epoch until all chilren have answere. In particular, it hols that τ i > τ j if epth(v i ) < epth(v j ). In orer to cope with chilren sensor that may not respon, Cougar eploys a timeout h. To unerstan the rawback of Cougar consier Figure 2 (top, right), where level 2 an level 1 noes have activate their transceivers at chronon zero an wait for the leaf noes to respon. If we have a failure at any given noe x, then each noe on the path x... s 0 (), has to keep its raio active for h aitional secons. III. THE MICROPULSE FRAMEWORK In this section we escribe the unerlying algorithms of the MicroPulse Framework. We ivie our escription in the following three conceptual phases: 1) Construction Phase, execute once prior the execution of Q, uring which the constructs the routing tree an becomes aware of the critical path cost ψ. 2) Pulse Phase, execute once prior the execution of Q, uring which each sensor s i tunes its wake-up chronon w i an waking winow τ i accoring to the value ψ. 3) Aaptation Phase, execute when a topology or workloa change occurs which results in a new critical path. A. The Critical Path Construction Phase ) an one counter per chil connection (enote as s in i,j ), where j enotes the ientifier of the chil. These counters measure the time that is require to transmit the tuples between the respective sensors an will be utilize as inicators of the link workloa in the subsequent epochs. Note that these counters account for more time than what is require ha we assume a collision-free MAC channel. Aitionally, it is important to mention that we coul have eploye a more This phase starts out by having each noe s j select one noe s i as its parent. This results in a waiting list similarly to Cougar [15]. To accomplish this task, the parent s i is notifie through an explicit acknowlegement or becomes aware of s j s ecision by snooping the raio. Note that in both TAG [9] an Cougar [15] noes select as their parent whichever sensor forware the query first. Alternatively, noes coul have chosen as their parent the neighbor with the smallest hop count from the or the one with the highest signal strength. In more recent frameworks, like GANC [11] an Multi-Criteria Routing [7], sensors select their parents base on query semantics, power consumption, remaining energy an others. In more unstable topologies a noe can maintain several parents [4] in orer to achieve fault tolerance but this might impose some limitations on the type of supporte queries. Nevertheless, all these alternatives are supplementary to this step. In the next step, we profile the activity of the incoming an outgoing links, an then propagate this information towars the. In particular, we execute one roun of ata acquisition where each sensor s i maintains one counter for its parent connection (enote as s out i complex structure rather than the counters s out i an s in i,j. That woul allow a sensor to obtain a better statistical inicator of the link activity. By projecting the time costs obtaine for each ege to a virtual spanning tree creates a istribute structure, similar to the one epicte in Figure 1. The final step is to percolate these local ege costs to the by recursively executing the following in-network function at each sensor s i : ψ i = { 0 if s i is a leaf noe, max j chilren(si)(ψ j + s in i,j ) otherwise. The critical path cost is then ψ 0 (enote for brevity as ψ). Using our working example of Figure 1, we will en up with the following values : ψ 5 i 9 = 0, ψ 4 = 4, ψ 3 = 29, ψ 2 = 11, ψ 1 = 59 an ψ 0 = 99. B. The Pulse Phase In this phase we shape the waking winow τ i an the wakeup chronon w i of each sensor by isseminating ψ, constructe uring phase 1, in the network. Algorithm 1 presents the main steps of this proceure which has a message complexity of

4 Algorithm 1 : MicroPulse: The Pulse Phase Input: n sensing evices {s 1, s 2,..., s n } an the s 0, the Critical Path cost ψ, the epoch e. Output: A set of n waking winows {τ 1, τ 2,..., τ n } an n wake-up times {w 1, w 2,..., w n }. Execute these steps beginning from s 0 (top-own): 1) If ψ > e then abort The Critical Path is larger than the Epoch. 2) Fin the maximum s in i,j in s i s chilren an enote the ientifier of this sensor as maxchil. Now calculate the waking time w i as follows: w i = ψ s in i,maxchil a b c, (1) where a, b an c are three variables which offset the costs of processing, the inaccurate clock an collisions at the MAC layer. The waking winow is the interval: τ i = [w i..(w i + s in i,maxchil )) (2) 3) Now isseminate ψ to s i s chilren. Upon receiving ψ, each chil noe s j ecreases ψ locally, as follows: ψ = ψ s out j (3) 4) At the same time with step 3, isseminate s in i,maxchil to s i s chilren. This information, will be useful to efine the latency tolerance (λ i ) of s i in the next aaptation phase. 5) Repeat steps 2-4, recursively until all sensors in the network have set w i an τ i respectively (i n). O(n). At the en of the algorithm execution each sensor knows exactly when it shoul wake up (i.e., w i ) an for how long (i.e., τ i ). To facilitate our presentation assume that the processing, the inaccurate clock an collisions at the MAC layer costs, enote as a, b an c respectively, are all equal to zero. Executing Algorithm 1 on the example of Figure 1 which has a ψ equal to 99 along with an epoch e of 100 yiels the following triples (s i, w i, τ i ): { (s 0, 59, [59..99)), (s 1, 29, [29..59)), (s 2, 46, [46..59)), (s 3, 29, [29..59)), (s 4, 37, [37..59)), (s 5, 35, [35..46)), (s 6, 39, [39..46)), (s 7, 27, [27..29)), (s 8, 0, [0..29)), (s 9, 33, [33..37)) } C. Aaptation Phase In this section we summarize the main ieas behin the aaptation phase which ajusts the critical path in cases of workloa or topology changes. To motivate our iscussion, consier the scenario where a chil in roun (r+1) requires more time to eliver the results to its parent than in roun r (i.e., an increase workloa). In the worst case this might require a complete reconstruction of the critical path cost. Fortunately, the structures eploye by MicroPulse enable a chil to know the interval uring which its parent will have an active transceiver. Therefore, the given chil may be able to start elivering the workloa earlier, if so, succeeing in completing the transmission on-time. In particular, each sensor s i knows the maximum incoming ege of its parent X from step 4 of Algorithm 1, enote as s in X,maxchil. The chil s i then calculates the latency tolerance 1 λ i of its parent X as follows: λ i = s in X,maxchil sout i (4) Note that λ i provies a sensor s i with the maximum leeway from the workloa inicator s out i. For instance in Figure 1, s 7 calculates λ 7 = 29 2 = 27. Thus, s 7 s workloa can increase by 27 chronons without affecting the synchrony of the query routing tree. The same also applies to the opposite case, where we have a ecrease in the workloa s out i. If s i is not on the critical path then there is absolutely no consequence on efficiency. If on the other han, s i is on the critical path then a ecrease in s out i might result in some waste of energy as s i s parent will have a larger τ than necessary. In orer to cope with such cases an the fact that s in X,maxchil might change or λ i might become negative, we reconstruct an re-pulse the critical path in the network perioically after a certain number of changes, failures or collisions. However, these upates represent the minority of the cases as our framework is tailore for stationary wireless sensor networks where the critical path will not change frequently. In the future we plan to evise more elaborate algorithms to cope with the aaptation uner more ynamic environments. IV. EXPERIMENTAL EVALUATION In this section we escribe our experimental methoology an the results of our evaluation. A. Experimental Methoology We aopt a trace-riven experimental methoology in which a real ataset from n sensors is fe into our custom-built simulator. Our methoology is as following: Algorithms: We have implemente the TAG, Cougar an MicroPulse waking winow algorithms. We utilize a failure rate of 20% where a sensor has a probability of 0.2 to not participate in a given epoch. We set the chil waiting timer h to 200ms. Sensing Device: We use the energy moel of Crossbow s new generation TelosB [2] sensor evice to valiate our ieas. TelosB is a ultra-low power wireless sensor equippe with a 8 MHz MSP430 core, 1MB of external flash storage, an a 250Kbps RF Transceiver that consumes 23mA when the raio is on, 1.8mA in active moe with the raio off an 5.1µA in sleep moe. Our performance measure is Energy, in Joules, that is require at each iscrete chronon to resolve the query. The energy formula is as following: Energy(Joules) = V olts Amperes Secons. For instance the energy to transmit 30 bytes at 1.8V is : 1.8V A 30 8bits/250kbps = 39µJ. 1 The latency tolerance is often also referre to as slack.

5 Aggregate Energy Consumption (mj) (processing + listening) Intel Berkeley Dataset - Energy Consumption (for all n sensors) (Graph:real, n=52, =14, e=31, link=250kbps) TAG - 20% failures Cougar - 20% failures Cougar - no failures MicroPulse - 20% failures Number of Epoch Fig. 3. Energy Consumption: Comparing the aggregate energy cost of the three algorithms using the TelosB energy moel. Dataset: We utilize a real ataset from Intel Berkeley Research [6]. This ataset contains ata that is collecte from 58 sensors eploye at the premises of the Intel Research in Berkeley between February 28th an April 5th, The motes utilize in the eployment were equippe with weather boars an collecte time-stampe topology information along with humiity, temperature, light an voltage values once every 31 secons (i.e., the epoch). The ataset inclues 2.3 million reaings collecte from these sensors. We use reaings from the 52 sensors that ha the largest amount of local reaings since some of them ha many missing values. Our query is SELECT motei, temp FROM sensors. The epth of the query spanning tree is 14. B. Energy Consumption In orer to assess the efficiency of the MicroPulse algorithm we measure the energy that is spent on the waking winow for the n sensors in isolation from the rest components (flash, weather boar, etc.). In particular, we measure the time that is spent in each MicroPulse phase an then multiply this by the respective current an voltage. Figure 3 shows that the aggregate energy consumption for the 52 sensors uner TAG, requires 7,984mJ. This is three orers of magnitue more than what the MicroPulse algorithm requires (13.75±0.58mJ). The big ifference in performance is clearly attribute to the uniform size of τ in TAG which is 2.21 secons (i.e., 31 (secons) / 14 (epth)), while in MicroPulse τ is only 146ms on average. The same figure also shows that the Cougar algorithm requires on average ±24.42mJ which is one orer of magnitue more than the energy require by MicroPulse. The isavantage of the Cougar algorithm originates from the fact that the parents keep their transceivers enable until all the chilren have answere or until the local timer h has expire (in cases of failures). Thus, any failure is automatically translate into a chain of elaye waking winows all of which consume more energy than necessary. This is shown by the thir line in which Cougar uner no failures requires only 42mJ which is only the 14% of what Cougar requires uner failures. It is interesting to highlight that MicroPulse maintains a competitive avantage even over Cougar (no failures). V. CONCLUSIONS AND FUTURE WORK This paper stuies the problem of optimizing the length of the waking winow in sensor networks in orer to conserve energy. Our ieas are establishe on profiling recent ata acquisition activity an on ientifying the bottlenecks using the Critical Path Metho. We have escribe the core ieas of our framework an some preliminary experimental results using our trace-riven simulator with real atasets from Intel Research. We foun that MicroPulse offers tremenous energy reuctions uner realistic conitions. In the future we plan to evise efficient aaptation algorithms which will be triggere in cases of variable workloa between consecutive chronons. Acknowlegements: This work was supporte in part by the US National Science Founation uner projects S-CITI (#ANI ) an AQSIOS (#IIS ), the European Union uner the project mpower (#034707) an the Cyprus Research Founation uner the project GEITONIA (#PLYP- 0506). REFERENCES [1] Cetintemel U., Fliners A., Sun Y., Power-efficient ata issemination in wireless sensor networks, In ACM MOBIDE, [2] Crossbow Technology, Inc. [3] Hill J., Szewczyk R., Woo A., Hollar S., Culler D., Pister K., System Architecture Directions for Networke Sensors, In ASPLOS [4] Consiine J., Li F., Kollios G., an Byers J. Approximate Aggregation Techniques for Sensor Databases, In IEEE ICDE [5] Intanagonwiwat C., Govinan R. Estrin D. Directe iffusion: A scalable an robust communication paraigm for sensor networks, In ACM MOBICOM [6] Intel Lab Data [7] Li Q., Beaver J., Amer A., Chrysanthis P., Labriniis A. Multi- Criteria Routing in Wireless Sensor-Base Pervasive Environments, In Pervasive Computing [8] Maen S.R., Franklin M.J., Hellerstein J.M., Hong W., The Design of an Acquisitional Query Processor for Sensor Networks, In ACM SIGMOD [9] Maen S.R., Franklin M.J., Hellerstein J.M., Hong W., TAG: a Tiny AGgregation Service for A-Hoc Sensor Networks, In OSDI [10] Gross J.L, an Yellen J., Graph Theory & Its Application, by Chapman & Hall/CRC Press, ISBN: X, [11] Sharaf A.M., Beaver J., Labriniis A., Chrysanthis P.K., Balancing Energy Efficiency an Quality of Aggregate Data in Sensor Networks, In VLDB [12] Sharma D. an Zaorozhny V.I. an Chrysanthis P.K, Timely ata elivery in sensor networks using whirlpool, In DMSN [13] Szewczyk R., Mainwaring A., Polastre J., Anerson J., Culler D., An Analysis of a Large Scale Habitat Monitoring Application, In ACM SenSys [14] Woo A., Culler D.E. A transmission control scheme for meia access in sensor networks, In MOBICOM [15] Yao Y., Gehrke J.E., Query Processing in Sensor Networks, In CIDR [16] Zaorozhny V. an Chrysanthis P.K an Labriniis A., Algebraic Optimization of Data Delivery Patterns in Mobile Sensor Networks, In DEXA [17] Zeinalipour-Yazti D., Lin S., Kalogeraki V., Gunopulos D., Najjar W., MicroHash: An Efficient Inex Structure for Flash-Base Sensor Devices, In USENIX FAST 2005.