Smart Sensors and Sensor Networks

Smart Sensors and Sensor Networks Lecture8 Data aggregation

Smart Sensors and Sensor Networks Data centricity • Data-centric networking (DCN) vs. identity-centric networking (ICN) • In DCN, the interest of the application is not in the identity of a particular sensor node, it is much rather in the actual information reported about the physical environment; • This is especially the case when a WSN is redundantly deployed such that any event could be reported by multiple nodes; • The application is not interested precisely which of these nodes is providing data; data and not nodes is addressed in requests; • The set of nodes that is involved in such a data-centric address is implicitly defined by the property that a node can contribute data to such an address; • In ICN, the requesting node has to find out all nodes that provide the requested capability and address them explicitly; • Implementation of data-centric networking by publish/ subscribe:

Smart Sensors and Sensor Networks • All nodes are connected to a software bus; • On this bus, data is made publicly available via a publish action; • Those nodes that have previously announced their interest in that particular kind of data by an appropriate subscribe action are then notified their interest about the availability of this data; • If a subscriber is no longer interested, it can simply unsubscribe from any kind of event and will no longer be notified of such events; • The publish/ subscribe interaction pattern provides three essential properties concerning their relationship between providers and subscribers of information: • Decoupling in space: publishers and subscribers need not be aware of each other, they can be oblivious of their mutual identities and numbers; • Decoupling in time: publishing and notification of data can happen at different times; the software bus has to provide intermediate storage; • Decoupling in flows: interactions with the software bus can happen asynchronously without blocking; • Implementation of the concepts of publishers and subscribers can be done centrally or distributed;

Smart Sensors and Sensor Networks • Addressing data • When subscribing to data or when publishing it, the problem is how to refer to this data; • The topic-based solution: • A set of keywords exist into which the set of all data is grouped and publications and subscriptions happen using a chosen keyword; • A typical example for topics are names of stocks traded at a stock exchange; when the price of a given stock changes, a notification for the corresponding topic is generated; • The topic abstraction is simple but it lacks flexibility; • The content-based solution: • Arbitrary predicates on the content of the entire data are evaluated; • An example: “is temperature reported in the events larger than 250C?”; • Primitive predicates can be combined into more complex ones with standard logical operators (and, or, not); they can also be scoped in time (“only provide notifications within a certain time”) or in space (“only provide notifications originating from a certain room”); • These predicates can be seen as filters describing which events/ notifications shall be delivered to each individual subscriber;

Smart Sensors and Sensor Networks • Implementation options and problems: • The centralized solution: • All subscribers and publications are sent to a central node that evaluates the data-centric addresses to decide to which node which publication to be sent; inadequate for WSNs; • Group communication; • The idea is to form one topic per group and to use multicasting to distribute publications to their subscribers in the respective group; • Problems: • The number of multicast groups would be highly dynamic and even difficult to determine as the content-determining predicates change; • Using a smaller number of summarizing groups to distribute notifications would result in needless network traffic; • Using a large number of predetermined groups, reflecting the finest granularity of data that the predicates could resolve, would necessitate publishing to many groups and managing such a large number of groups would result in large administrative overhead; • Content-based networking cannot easily be mapped to the usual identity-based multicast groups; • Instead, content-based forwarding and routing are required; • Each node has to store, for each of its neighbors, a predicate describing the set of predicates that nodes in that subnetwork have subscribed to;

Smart Sensors and Sensor Networks • Convergecast: • Gathering of data from many or all nodes to a single or few sinks; • A convergecast tree is closely related to a multicast tree; • Major problem: the implosion: there is a lot of traffic converging at the root node of the tree; this effect is accentuated when redundant data is reported to the sink via multiple roots; • Optimization: the aggregate of the data to reduce the amount of data that has to be transported; • Data-centric storage (DCS) • Which entity needs to know each data? • DCS means to let the data itself describe where it is stored; more exactly, the name of the data is used to represent a key under which the data can be looked up; • Under this key, that is under the node that is identified by the key, the actual data is stored; • A query can then be routed directly to the node corresponding to the name of the data and directly retrieved; • This concept saves cost to distribute either data or queries in the entire network, replacing a flood by a unicast communication;

Smart Sensors and Sensor Networks • The precise role of a key can be defined in several ways; • A solution is to use a geographic interpretation of the key: the name of the data corresponds to a particular location and the mapping rule is known to all nodes; • Both producer and querier of data can easily compute the storage location of a given data item and route their packets toward it, using some geographic routing protocol; • Problems: • Nodes not available at the hashed location: when no node is available at precisely the location to which the key matches, it will not be possible to deliver the packet to such a node; instead, the closest node to the computed position should represent a good proxy; how this proxy can be determined depends on the geographic routing protocol; • Handling limited storage per node: it can happen that too much data ends up at a single node, either from a single key or from different keys; a solution is the use of mirror nodes in the vicinity of the home node to store such data; storage space is obtained at the cost of some additional local communication; • Handling failing and new nodes: when a node fails, all the key-value pairs stored in it would be lost; to avoid this situation, the data should be replicated, ideally to neighbor nodes; a simple solution is the perimeter mechanism; to detect a failed node and to keep the replicas consistent, all nodes occasionally send refresh packets around the perimeter; the same mechanism can be used to integrate new nodes;

Smart Sensors and Sensor Networks Data aggregation • Data aggregation is a form of in-network processing; • Data is computed (aggregated) at an intermediate node between sources and sink and the aggregated data is sent to the sink; there is no loss of information; • Examples: average, minimum or maximum values from a set of sensed values; more complex examples: approximating contours of regions or equivalent values, approximating lines of polygons that separate different regions etc.

Smart Sensors and Sensor Networks • Data aggregation as a tree formulation: • Data flows from sources to a sink along a tree; intermediate nodes in the tree apply some form of aggregation function to data they have collected from some or all of their children; • This aggregated value, possibly along with additional administrative values (for example, the number of nodes that have contributed to a mean value) is forwarded; • The efficiency of data aggregation can be judged using metrics: • Accuracy: the difference between the resulting value at the sink and the true value, since not all data is delivered to the sink any longer; can be expressed as differences, ratios, statistics etc. • Latency: aggregation can increase the latency of reporting as intermediate nodes might have to wait for data; • Message overhead: is reduced resulting in an improved energy efficiency and network lifetime; aggregation protocols can often deliberately trade off between accuracy, message overhead and latency and only provide estimates of the actual aggregated value;

Smart Sensors and Sensor Networks • Main open questions: • Which aggregation functions can be used, what categories exist? • How can the tree be formed, where and how should aggregation points be placed? • How long should a node wait for data from its children/ neighbors? • How should an interface look like that allows to easily express aggregation actions? • Categories of aggregation operations: • Representation of intermediate results: • When computing aggregates in an intermediate node, it is, in general, insufficient to only communicate the result of the actual aggregation function; • An example: to compute an average, a tuple <average, count> should be exchanged between nodes; • In some cases, it is simpler: if only sum and count are exchanged or when minimum or maximum functions are necessary; • The nature of the aggregation function determines the precise form of the intermediate state record;

Smart Sensors and Sensor Networks • Aggregation functions: • Given two partial state records <x> and <y>, either received or locally measured, an aggregation function f computes a new state record <z> = f(<x>, <y>); • Properties of f to be an aggregation function: • Duplicate sensitive: is the aggregation result changed if the measured value of a particular device, or some intermediate aggregate, is used in the composition more than once? • The answer should influence the aggregate structure; • Example of duplicate sensitive aggregate: the average value; • Example of not sensitive duplicate: the minimum or maximum values; • Summary or exemplary: • An exemplary aggregate is a single, representative, value out of a set of values; examples: maximum and average values; • A summary aggregate is a function of the entire set and, typically, does not strongly depend on individual values; example: sum value; • Timing aspects: • The aggregation should be done early or late? • Aggregation should be done in one shot or with certain periodicity? • Aggregation in time or in space?

Smart Sensors and Sensor Networks • Composable: • An aggregation function is said to be composable if the result of f applied to a set W of measurements can be computed by applying f to some partition of W = W1 U W2, usually with W1∩ W2 = 0, using a known helper function; formally: f(W) = g(f(W1), f(W2)) • Behavior of partial state record: the partial state necessary to compute the aggregated value varies for different aggregation functions; cases: • Distributive: the partial state is the aggregated value of a set of partial measurements; no additional function has to be applied at the sink; the partial state is of constant size; MIN is a typical distributive aggregate; • Algebraic: the partial state is of constant size, often with the actual aggregate a part of it or a simple function; AVG is a typical example; • Content-sensitive: the size and structure of the partial state depends on the values that have been actually measured; example: a histogram of the measured values; • Holistic: the partial state needs to reflect all measured values; • Unique: these functions are similar to holistic ones except that here the partial state size is only proportional to the number of distinct values that have been observed (rather than all of observed values); • In a practical sense, aggregation functions with distributive and algebraic partial state are well amenable to aggregation; content-sensitive functions may or may not be; holistic functions cannot practically be aggregated and for unique functions, aggregation might be applicable depending on the behavior of the values to be observed;

Smart Sensors and Sensor Networks • Placement of aggregation points • The aggregation points have to be well placed for maximum benefit; • Aggregation should happen close to the sources and many sinks should be aggregated as early as possible; the tree should have long trunks and bushy leaves; • When to stop waiting for more data? • When aggregating data, an intermediate node, as well as the sink, has to decide how long to wait for data from each of its children in a convergecast tree; • In the simplest case, a node knows which of its neighbors are its children, by means of an acknowledgement of the invitation messages during tree formation, and waits for answers from all of them; • This can take a long time because of fluctuations in the radio channel with ensuing high error rates, temporary node failures or because of a very imbalanced tree; • Waiting a long time will result in more data entering, thus higher accuracy but it will also increase delay and, potentially, energy consumption because of the required idling of the radio receiver; • A compromise has to be found;

Smart Sensors and Sensor Networks • Broadcasting an aggregated value • Aggregation can also be applied to broadcasting an aggregated value of all sensor readings within the entire WSN; • Example: to inform all nodes of the currently highest measured temp.; • A gossiping-based solution • The case: to provide an estimate of a composable, algebraic aggregation function to all nodes in the network; the task is complicated by potential failures of both links and nodes; • A simple fully distributed solution: every node sends its measurements to every other node in the network; is robust but does not scale; • The alternative is the hierarchy; it introduces scalability with acceptable message overhead; • The nodes are partitioned into groups with K members on average, K being a constant; K of these groups are again collected into a group of the next hierarchy level and so on until a single group remains; • In each group at each hierarchy level, a leader is elected; • The aggregate is computed bottom-up by group members sending their measured values/ their computed estimates to their respective group leader;

Smart Sensors and Sensor Networks • At the end, the leader of the top-most group knows the aggregated value, because the aggregation function is supposed to be composable and can redistribute it downward; • Building such a hierarchy in a sensor network should exploit geographic or, even better, radio proximity; • The hierarchy can also be used for a gossip-based approach, circumventing the inherent reliability problems of leader-election-based solution; • Continuous, exemplary aggregation with adaptive accuracy • The goal is to distribute aggregated value to all nodes, continuously and based on a special form of gossiping; • It is suitable only for exemplary aggregation functions (minimum and maximum) and it does not need the construction of hierarchical grouping; • Each node not only maintains an estimate of the current aggregated value, it also stores an estimate of the confidence a node puts into this value; this confidence for an aggregated scalar value A could be the variance of A; • The intelligence of this scheme lies in aggregating the local estimate with new information either from local measurement or received from a neighboring node; in addition, rules are described to regulate when such a new local estimate is to be sent to neighbors; 3 different rules are required:

Smart Sensors and Sensor Networks • Aggregating local estimate with local measurement: • Suppose a sensor node has some local estimate of the global aggregation function and it performs some local measurement, contributing to this aggregated value; a heuristic is needed to upload the local estimate, taking into account the concrete aggregation function to be computed; • Aggregating local estimate with remote information; • Deciding whether to transmit local estimate after update: • Once a node has updated its local estimate, the question is whether to distribute this new estimate to its neighbors; • The intuition is to only distribute estimates if they would bring about significant changes in the neighbors’ estimates in turn; • To decide this, every node X keeps track of the last estimates it has heard from each of its neighbors; node X combines its own, new estimate with each of these remembered neighbor estimates, calculating the new value that the neighboring node would obtain if it were informed about the update at node X; node X can determine the results that the other nodes would compute since all nodes use the same fusing algorithm; • Once the new result at a given neighbor Y is determined, node X computes the difference between node Y’s current and its prospective, new estimate; only if this difference, either in value or confidence, is larger than a predefined threshold will node X transmit its new estimate to node Y; • The justification is that these speculative calculations consume much less energy than simply transmitting the packet;

Smart Sensors and Sensor Networks • Data funneling • The scenario is that of a data sink requesting all sensors in a given geographical region to periodically send their measured values; the data sink is interested in obtaining all values of each, uniquely identified node; • The main idea is data funneling; • After the request, or interest, has been flooded, using any geocasting method, the nodes in the region of interest can send their data to a node on the border of that very region; this node will act as an aggregation point and forward the readings to the data sink; • To protect this border node from exhausting its energy resources, its role is rotated occasionally among all border nodes that face the data sink; • To achieve this load balancing, all interior nodes must know when to send their data to which border node; this is possible by having all the border nodes flood the interest packet, along with their own identifiers, into the region; • Then, all interior nodes know about all eligible border nodes and can apply a selection function to this record of information; • Since all interior nodes apply the same function to the same data, assuming no border node announcements are lost in the region, at a given time all data will be sent to the same border node;

Smart Sensors and Sensor Networks • Architectures for sensor integration • The real implementation of the sensor procesing system falls within one of the Flynn classes for computer architecture:

Smart Sensors and Sensor Networks • Traditionally, in cases in which the volume of information from the sensors is high, hierarchical architectures have been used for processing; • These architectures are implemented in three or more levels, each with an MIMD architecture, so the global system can be viewed as an interconnected cluster of MIMD systems; • However, the option of a hierarchical architecture may not always be the best; the best choice depends on factors such as the problem itself, available technology, reliability, complexity, cost and performance of the solution; • One needs a parameter independent of the particular implementation to indicate which of the solutions (e.g. parallel or hierarchical) is more complex, difficult and expensive to implement; • Merit factor of a system: allows comparison of the systems: • It is defined as the product of the bandwidth, BW, times the processing power, PC, needed to be able to solve a certain problem: MF (Mbytes/s x MIPS) = BWr x PCr • To be able to obtain the necessary expressions for parallel and hierarchical systems it is necessary to know the MF value of the association of a certain number of processors each with its specific MF in serial or in parallel;

Smart Sensors and Sensor Networks • Merit factor in a parallel system • It is assumed a number of N parallel connected processors, each with a certain MFi value; • The bandwidth of the system is the sum of the individual bandwidths, BWi, of each one of the processors: • The computing power is the sum of the capacities, PCi, of each one of the processors: • Therefore the merit factor of a parallel system in function of the MF of each processor is:

Smart Sensors and Sensor Networks • Merit factor in a series system • The processors are connected sequentially; • For N processors, each one with its bandwidth BWi, and computing power PCi, the equivalent system with just a single processor will have the bandwidth of the first one of the series: • The equivalent computing power will be determined by the total processing time, tN, and the number of operations to carry out in this time: D is the number of data and opi is the number of operations carried out by processor i; • The required quality of the equivalent system is:

Smart Sensors and Sensor Networks • Parameterization of parallel and hierarchical architectures • A certain processing problem can be parameterized, indicating the total bandwidth, BW, and the processing power necessary, PC, for their computation; • Solving the problem by means of a parallel architecture yields the following expressions: • However, it is common to have processing units with some processing power, PCi = PCpu and with bandwidth ABi = ABpu; therefore, the parameter to determine will be the number of units, N, with those characteristics necessary to implement the system:

Smart Sensors and Sensor Networks • The aggregating element (aggregator) picks up the decisions of the N processing elements and elaborates a final one; • The ways of getting this decision are varied; a solution can be a yes/no binary decision elaborated performing the logical AND of the decisions; • The required bandwidth in the aggregator depends on the size of the partial decisions of the N processing elements and the time, T, to make the final decision; • The processing power depends on the number of and operations necessary to carry out the decision and of the time T necessary to carry them out; using two input and operations, the number of the operations to obtain the result with N inputs is N – 1; • Final and partial decisions are overlapped in time if the system works in a pipeline way: first the system gets the N partial decisions in T seconds and after, the aggregator final decision is obtained in the same T time, when the next partial ones are received; • The bandwidth and processing power of the fuser can be expressed as: Sdec = the size in bytes of the partial decision; Sdat = the number of bytes for each datum coming from the sensors; ξ= the relationship between the size of the partial decision and the size of the data from each sensor; D = the number of data; op = the number of operations per datum;

Smart Sensors and Sensor Networks • If the resolution of the problem is outlined by means of the employment of a hierarchical structure, it will be necessary to introduce a new parameter to be able to obtain the expressions for BWi and PCi that will be a function of the level in the architecture;

Smart Sensors and Sensor Networks • This parameter is: • The reduction factor in the data volume due to the extraction of the information from the received raw data from sensors, in the case of a measurement system or • The proportion of data discarded by not completing certain requirements if it is a detection system; • At each level, the bandwidth, BWi, is reduced in a factor similar to the reduction due to the extraction of information or to the elimination of data not interesting (with a factor p); • In this case, suppose that, from a level to the next level, all data (“yes” decision) or no data (“no” decision) pass; • This decision has a probability p so that the time for the information arrival between levels is increased by the inverse of that probability; • Because all the data are sent, the effect is more time to send the same quantity of data, which implies a reduction of the bandwidth; • Assume that the processing power at each level i can be expressed as a proportion, αi, of the total processing capacity, PC; • Each one of the levels in the hierarchical system is thought of as a parallel system with a fuser; • The result of the decision makes the entirely of the data pass or not toward the next level;

Smart Sensors and Sensor Networks • Evaluation of Merit factor for: • Parallel systems, • Hierarchical systems; • Evaluation of Merit factor for parallel systems • The MF for the parallel system has two terms: • MFproc : depends on the used processors and • MF fus : depends on the fuser; • The equivalent of the parallel system is an association of N parallel processors in series with a fuser; • Suppose that all the processors have the same characteristics: N = the number of units, MFpu = the MF of each processing unit;

Smart Sensors and Sensor Networks • The MF of the parallel system will be the series of both calculated, that is: a and b are the relationships between the processing capacity, PCpu, and the bandwidth, BWpu, respectively, of each processor and the total of the problem; • Evaluation of Merit factor for hierarchical systems • The hierarchical system is a series of E sequential connected parallel systems; in each level, the system is implemented with N processors and an aggregator; therefore:

Smart Sensors and Sensor Networks • MFParali is the MF of each level, expressed as (it was assumed that the aggregator needs the same time to get a decision as the processors on the same level): • The MF for the processors on the level is the parallel association of N of them, that is: ai and bi are defined the same way as in the parallel system, and N is defined as:

Smart Sensors and Sensor Networks • The figure with the hierarchical system shows how to calculate the MF of the aggregator; • The bandwidth of the aggregator for level i is: • The processing power necessary in the aggregator of level i is: • Therefore, the MF of the aggregator of level i is:

Smart Sensors and Sensor Networks • Substituting in the top eq. from slide 28 yields: • Therefore, the MF of the hierarchical system, MFHier, is

Smart Sensors and Sensor Networks • The last equation becomes identical with the equation corresponding to he parallel system, when the number of levels is one; • The following step is to carry out the comparison between the expressions for the parallel and the hierarchical systems and to obtain an analytic expression that follows the decision of the better solution for a given problem; • If the MF of the parallel case is taken as a reference, it results an equation and 4E variables, which would determine infinite solutions; • A different approach to compare both systems is the following: • The parallel solution is determined and its MF is calculated; • A hierarchical system is designed and its MF is calculated; • Both results are compared; if one is interested in a hierarchical solution and its MF is bigger than the one for the parallel solution, the hierarchical parameters (a, b and P) can be adjusted and the process repeated until the MF is smaller; • However, it is possible to obtain an analytic expression if what is known is the value of the MF of the parallel system and the values of the parameters of E – 1 levels of the hierarchical system of E levels; imposing that the MF in the hierarchical system is smaller than the one in the parallel system, the following expression is obtained to calculate the values of the parameters of the last level:

Smart Sensors and Sensor Networks • Preprocessing elements: • The introduction of preprocessing elements improve the system performance due to the reduction of their processing loads; • When the number of sensor channels to process is very high, the required processing power and bandwidth in the levels of the system can be too high; • In image processing, hierarchical processing systems are used in which successive levels carry out the processing with higher resolution; • Particularly in analysis of video images, because of the great correlation that exists between frames, a technique is used based on the location of areas (region of interest, RoI) of the image that have changed from one frame to the next; • Processing takes place only on those regions with the consequent reduction in the processing time; • The idea is to apply the concept of RoI to a hierarchical processing sensor system with a great volume of information to release the computational load in the levels where this is possible; • The RoI can be defined as a group of sensor channels of the system; it can represent a one-dimensional, two-dimensional or thee-dimensional space of measure of the physical environment;

Smart Sensors and Sensor Networks • Data clustering • Another of the improvements that can be introduced in the system is to try to avoid the dispersion of the data to process in each processing element at each level in the hierarchy; • The use of a RoI suggests that all its channels should be processed in a combined way because the elaboration of the decision will be made on the basis of existing relationships among the values of the channels; • If the levels of the hierarchy are implemented like parallel systems, it can happen that the channels of the RoI may be distributed among several processors, making intercommunication necessary, and thus increasing the time necessary for processing and reducing the performance of the system; • A solution is to gather the data so that one maximizes the probability that all the channels of an RoI are sent to only one processing unit of the parallel system inside the level; • For that, one must discover channels that will be part of an RoI with bigger probability; this implies that the physical process has a certain bias and is present more probably in certain subspaces of the measure space; • If this is not the case, then two options exist: • To use the information of the RoI to carry out a dynamic routing of the data; the solution introduces some delay; • To gather the channels in a static form, following some relationship with the physical phenomenon that is being observed;

Smart Sensors and Sensor Networks • Example: • CERN, the European Laboratory for Particle Physics, is the most important laboratory in the world for the study of particle physics; it holds the bigger accelerator, the LHC large hadron collider, in which two beams of protons will collide with an energy near 14 TeV (tera electron volts) to study the origin of mass by searching a new particle; • For analysis of the collision results, two big accelerators, ATLAS and CMS, have been constructed; • ATLAS is a huge toroid, 22 m long and 32 m high, with more than 170 million electronic channels to read, coming from sensors inside the detector; • All these channels sum a total of 1.3 Mbytes to be read every 25 ns, giving a rate of 50 Tbytes/s; the total processing capacity needed to perform all he operations is estimated at 5.1010 MIPS; • A parallel solution would require, assuming 100 MIPS and 200 MB/s processors, 500 million processing units (eq. 24.9); MF = 2.5 x 1021 MIPS x MB/s; • A three-level hierarchical solution can be implemented using 40 MIPS and 200 MB/s processors for the first level and 100 MIPS and 200 MB/s processors for the next two levels; the MF total is 1.45 x 1014;

Smart Sensors and Sensor Networks • By introducing RoIs, the hierarchical system was improved; in this care, only level 2 will include RoI; simulations showed that the average number of RoIs per acquisition in level 2 would be 5, each one with 135 000 channels; the new number of processing units at level 2 is 1440 and a new value for RoI was obtained:

Smart Sensors and Sensor Networks