[IEEE 22nd International Conference on Data Engineering (ICDE'06) - Atlanta, GA, USA (2006.04.3-2006.04.7)] 22nd International Conference on Data Engineering (ICDE'06) - Energy-Efficient Continuous Isoline Queries in Sensor Networks

Download [IEEE 22nd International Conference on Data Engineering (ICDE'06) - Atlanta, GA, USA (2006.04.3-2006.04.7)] 22nd International Conference on Data Engineering (ICDE'06) - Energy-Efficient Continuous Isoline Queries in Sensor Networks

Post on 14-Apr-2017

215 views

Category:

Documents

0 download

Embed Size (px)

TRANSCRIPT

<ul><li><p>Energy-Efficient Continuous Isoline Queries in Sensor Networks</p><p>Adam Silberstein Rebecca Braynard Jun YangDepartment of Computer Science, Duke University, Durham, NC 27708, USA</p><p>{adam,rebecca,junyang}@cs.duke.edu</p><p>Abstract</p><p>Environmental monitoring is a promising application forsensor networks. Many scenarios produce geographicallycorrelated readings, making them visually interesting andgood targets for the isoline query. This query depicts bound-aries showing how values change in the network. Temporaland spatial suppression provide opportunities for reducingthe cost of maintaining the query result. We combine bothtechniques for maximal benefit by monitoring node and edgeconstraints. A monitored node triggers a report if its valuechanges. A monitored edge triggers a report if the differencebetween its nodes values changes. The root collects reportsand derives all node values, from which the query result isgenerated. We fully exploit this strategy in our algorithm,CONCH, which maintains the set of node and edge constraintsthat minimizes message traffic.</p><p>1 IntroductionConserving battery power is paramount to prolonging the lifeof a wireless sensor network. Network nodes communicatebetween each other and to the base station, or root, via ra-dio. Radio dominates energy consumption. Query algorithmsmust be sensitive to this cost, and attempt to minimize com-munication during execution. A primary strategy for min-imizing communication is suppression, which, in general,keeps nodes from transmitting data unless monitored condi-tions change. This strategy is effective in many scenarios,such as environmental, where conditions change slowly, orindustrial, where change reflects abnormal behavior.</p><p>We now examine the isoline query, which maintains a mapof the sensor network with lines drawn bounding areas ofnodes with identical or similar measured values, such as tem-perature. If two nodes are close enough in value, we saythey are in the same tier. We increase query resolution bydecreasing tier width. The isoline query can be used to vi-sualize the moving trends affecting the monitored area. Itis particularly useful when data is geographically correlated,and nearby nodes have a high likelihood of being in the sametier. The resulting map is then likely to be visually interesting</p><p>The authors are supported by NSF CAREER award IIS-0238386 andgrant CNS-0540347.</p><p>with longer isolines, rather than disjoint points or short linesegments. Fundamentally, the query maintains an estimate ofeach node value.</p><p>Temporal Suppression The most basic version of suppres-sion is temporal, where each node only transmits its moni-tored value to the root if its tier has changed since its lasttransmission. The root, in turn, fills in all unreported valueswith their previous ones. If most nodes are unchanging, thisscheme is effective. When all nodes change, though, all trans-mit, incurring a high cost.</p><p>Spatial Suppression This scheme lets nodes suppress if theirvalues are the same as their neighbors. The scheme exploitslocal data correlation, since nearby nodes are most likely tohave similar values. Spatial suppression can be done by hav-ing all nodes attempt to report their values at different slotsduring a timestep. Nodes overhear reports sent by neighbors.When a nodes slot comes up, it first computes the average ofvalues overheard so far. If its value is close to this average,it suppresses. The root then fills in missing values with theaverage of neighbors. This approach has a significant flaw. Toaccurately derive a nodes value, the root must know which ofits neighbors were averaged to trigger suppression. The aver-age of all neighbors may differ from the average of that subset.Rectifying the problem requires a global reporting order of allnodes, known by the root so it can correctly derive a nodesvalue by averaging only the values that precede it in the order.Due to clock skew among nodes, holding to a global order islikely expensive in practice.</p><p>Spatio-Temporal Suppression This policy attempts to com-bine the advantages of both schemes. The obvious policy is tosuppress nodes if they qualify for either type. This approachis problematic, however, because this oring creates ambi-guity; the root has no way of knowing if a missing value istemporally or spatially suppressed. If these are in conflict,the correct value cannot be derived. One solution is to andthe schemes and suppress only when a node qualifies for both.The resulting approach would be less effective than using onlyone suppression type, since it can only restrict suppressionchances. Therefore, there is no benefit to this combination.</p><p>Nevertheless, there is potential in combined suppression.If a nodes value does not change, it should not report. Addi-tionally, if a nodes value does change, but its relationship to</p><p>Proceedings of the 22nd International Conference on Data Engineering (ICDE06) 8-7695-2570-9/06 $20.00 2006 IEEE </p></li><li><p>its neighbors does not, it should not report. Consider the net-work in Figure 1. In one timestep, a band of nodes all rise invalue (shift from white to black), causing the isoline bound-ing the high value area to move from left to right. Such a shiftis a typical example of nearby nodes acting in a correlatedfashion. The nodes in the interior of the band change value,but maintain the same relative relationships with their neigh-bors. Ideally, only one node, on the boundary of the band,should need to report to detect the bands movement. Thisscheme would achieve a dramatic improvement over requir-ing all nodes to report, as with temporal, or one node fromevery local cluster of nodes within the band, as with spatial.Our Contribution As we have shown, neither temporal norspatial are alone sufficient to exploit all suppression opportu-nities. It is not obvious, however, how to implement effectivespatio-temporal suppression. To meet this challenge, we havedeveloped the monitoring algorithm CONCH, short for con-straint chaining. It employs spatio-temporal suppression bymonitoring changes to both node values and value differencesalong edges. We view each monitored node or edge as a con-straint. By maintaining a chain of constraints, we can globallyderive all values in order to support the isoline query.</p><p>2 Related WorkThe isoline query is targeted in a number of papers. One ap-proach similar to ours is event contour [7]. This method com-bines temporal and spatial suppression, but in the problematicway where the root cannot be sure which suppression type isused. In addition, their overhearing-driven spatial suppressionmeans that even when all nodes have the same value, somemust always report. Further, if two nodes have correlated butvery different values, both will be reported, even though oneis sufficient to derive the other. Other isoline-related problemsare discussed in [4] and [1].</p><p>Temporal suppression based on value changes is used forcontinuous aggregation queries in [8]. A more sophisticatedform of temporal suppression, proposed in [6], uses KalmanFilters to prevent a node from reporting only if a change in itsvalue diverges from what a model would predict. The generalidea of model-based suppression is orthogonal to ours; insteadof using value-based constraints, we can use constraints on theparameter values of models that predict how values change.</p><p>3 PreliminariesOur network consists of n fixed-location nodes u1, . . . , un,all measuring some feature such as temperature. An edge, eij ,exists for each pair of nodes, ui and uj , within radio commu-nication range of each other. The network is rooted at a basestation node, and messages can be sent using some existingrouting protocol [5, 10].</p><p>The primary source of energy usage is radio communica-tion. We optimize for energy efficiency by minimizing thenumber and size of messages sent through the network, andevaluate algorithms accordingly. We use the energy specifi-cations of MICA2 motes [2], and account for both per-bytesending/receiving costs and per-message startup costs.</p><p>The isoline query is essentially a continuous query request-ing the tier values of all nodes. We simplify the notion of acontinuous query with rounds. Each round is sufficiently longfor all nodes to communicate among themselves and to theroot. We assume the root initially disseminates the query planinto the network, but nodes act autonomously in each round.The root may infrequently update the plan in the network.</p><p>3.1 A First Cut: Neighborhood ApproachBefore presenting CONCH, we introduce a neighborhood-based algorithm that supports an improved spatio-temporalsuppression scheme. This algorithm returns the correct result,but has performance shortcomings we will further address inCONCH. In the neighborhood approach, each node maintainsvalue information for each node within its communication dis-tance. This set of nodes is the nodes neighborhood.</p><p>We apply temporal suppression by giving all nodes the op-portunity to regularly broadcast their values, but only havingthem do so if their tiers have changed since last broadcast. Ifui broadcasts a change in tier, all nodes that hear the messageupdate their values for ui in their list of neighbors. We applyspatial suppression as follows. ui tracks the difference be-tween its own tier, vi and the tiers of each of its neighborhoodnodes. Periodically, ui has the opportunity to send an updatereport to the root. ui remembers the last reported tier differ-ence between itself and uj , (voldi voldj ) and calculates thecurrent difference, (vi vj). If these calculations differ, uiincludes the new difference in its report. This process repeatsfor each neighbor.</p><p>This suppression policy captures the features we sought inSection 1 to improve spatial suppression. Now, whether twoneighbors are in the same or different tiers, if they move to-gether, the policy suppresses. We expect such behavior dueto local correlation. The scheme is spatio-temporal: spatialbecause it suppresses neighbors when their relationship staysthe same, and temporal because it suppresses the need to con-stantly report neighbor relationships as time progresses.Shortcomings The neighborhood approach has some keyshortcomings. Say we have a collection of 8 nodes all in thesame tier, vl, but at the next timestep, 4 nodes jump to a highertier, vh. This example is depicted in Figure 2(a), with theblack nodes having higher values. A horizontal axis is drawnto show the split between low and high values. The dottedlines indicate neighbor relationships between pairs of nodes.When the black nodes rise in value, many tier differences aredetected between black and white nodes. These differencescorrespond to the dotted edges crossing the axis. Accordingto the neighborhood approach, each node incident to such adotted edge sends a message to the base station. The size ofthe message is proportional to the number of incident edges.</p><p>This illustration reveals the redundancy inherent in this ap-proach. Each node contributes to several difference calcula-tions, and is sent several times over. Although there are a lotof changes, the overall effect can be described succinctly: allnodes above the axis move to vh. Our intuition is we shouldonly need a single message to capture these changes.</p><p>Proceedings of the 22nd International Conference on Data Engineering (ICDE06) 8-7695-2570-9/06 $20.00 2006 IEEE </p></li><li><p>Figure 1. Moving isoline.(a) (b)</p><p>Figure 2. Edge monitoring.</p><p>a</p><p>b</p><p>f</p><p>d</p><p>e hc</p><p>gi k</p><p>jl</p><p>m</p><p>Figure 3. CONCH forest.</p><p>The problem can be explained by considering edges. Theneighborhood approach monitors all edges between nodeswithin communication distance. This scheme causes the largenumber of messages triggered in our example, instead of ourgoal of only one to describe the change.</p><p>4 CONCHWe now present CONCH, the main contribution of this paper.The crux of CONCH is that we can support the spatio-temporalsuppression strategy by monitoring far fewer edges than theneighborhood scheme, with a minimal spanning forest whoseedges correspond to monitored constraints. We refer to Figure2(b), where a spanning tree now connects the nodes. We no-tice that far fewer edges are monitored compared to the neigh-borhood approach. When the nodes above the axis move tovh, only one edge observes a change in tier difference betweenits nodes, so only one message is sent. The black nodes newvalues can be derived from this one message. Because noneof the edges between the black nodes were reported, we knowthat each connected pair must continue to be in the same tier.If the left-most black node is now in vh, then all the blacknodes must be in vh. We achieve energy savings by reduc-ing the number of monitored edges while, through constraintchaining, being able to derive all node values.</p><p>4.1 CONCH PlanFormally, a CONCH plan is a spanning forestSF over the sen-sor network. All nodes are covered by some tree in SF . Somesubset of nodes (always including the base station) are desig-nated as tree roots, and are the set of monitored nodes in thenetwork, roots(SF ). The edges in the forest determine theset of monitored edges in the network, edges(SF ). For eache edges(SF ), one incident node is designated the updaterand the other the reporter. By convention, for eij , ui is theupdater. Nothing prevents a node from serving either role onbehalf of any of its incident edges, but each node serves onlyone role per edge.</p><p>The plan maintains constraints on roots(SF ) andedges(SF ). For each ui roots(SF ), if vi changes, viis sent to the base station. For each eij edges(SF ),if vi changes, ui sends an update to uj . If the difference(vi vj) changes, uj sends the new difference to the basestation. Hence, CONCH combines two suppression strategies.Tree roots are maintained temporally, while edges are main-tained spatio-temporally.</p><p>It is important to note that the spanning forest used inCONCH is usually different from the spanning tree used forrouting. A CONCH forest is constructed in a way (discussed indetail in [9]) to minimize the expected cost of monitoring theconstraints implied by the forest. On the other hand, a routingtree is constructed in a way to allow messages to reach thebase station in as few hops as possible. For example, consideran example CONCH forest in Figure 3. The part of the forestrooted at node a contains a very long path to f , and wouldserve poorly as a routing tree. This separation of the CONCHforest and routing tree does not impact communication effi-ciency, however, because we still route all messages using therouting tree.</p><p>4.2 CONCH AlgorithmCONCH is divided into a start-up phase and continual phase.Start-up A spanning forest SF covering the network is con-structed from the set of all possible nodes and edges. It is builtand maintained either by the base station, or in a distributedmanner using techniques such as those in [3]. In either case,the base station is always informed of updates to SF . Eache SF is assigned a reporter and updater. Figure 3 shows anexample CONCH forest. Gray nodes are tree roots. The edgeroles...</p></li></ul>

Recommended

View more >