ucla oct 2002

Quality Aware Sensor Database (QUASAR) Project** : Quality Aware Sensor Database (QUASAR) Project** Sharad Mehrotra Department of Information and Computer Science University of California, Irvine **Supported in part by a collaborative NSF ITR grant entitled “real-time data capture, analysis, and querying of dynamic spatio-temporal events” in collaboration with UCLA, U. Maryland, U. Chicago


Talk Outline : Talk Outline Quasar Project motivation and background data collection and archival components query processing tracking application using QUASAR framework challenges and ongoing work Brief overview of other research projects MARS Project - incorporating similarity retrieval and refinement over structured and semi-structured data to aid interactive data analysis/mining Database as a Service (DAS) Project - supporting the application service provider model for data management


Emerging Computing Infrastructure… : Emerging Computing Infrastructure… Instrumented wide-area spaces In-body, in-cell, in-vitro spaces Generational advances to computing infrastructure sensors will be everywhere Emerging applications with limitless possibilities real-time monitoring and control, analysis New challenges limited bandwidth & energy highly dynamic systems System architectures are due for an overhaul at all levels of the system OS, middleware, databases, applications


Impact to Data Management … : Impact to Data Management … Traditional data management client-server architecture efficient approaches to data storage & querying query shipping versus data shipping data changes with explicit update Emerging Challenge data producers must be considered as “first class” entities sensors generate continuously changing highly dynamic data sensors may store, process, and communicate data


Data Management Architecture Issues : Data Management Architecture Issues Where to store data? Do not store -- stream model not suitable if we wish to archive data for future analysis or if data is too important to lose at the producers limited storage, network, compute resources at the servers server may not be able to cope with high data production rates. May lead to data staleness and/or wasted resources Where to compute? At the client, server, data producers Data/query request Data/query result server producer cache


Quasar Architecture : Quasar Architecture Hierarchical architecture data flows from producers to server to clients periodically queries flow the other way: If client cache does not suffices, then query routed to appropriate server If server cache does not suffice, then access current data at producer This is a logical architecture-- producers could also be clients. server Client cache Server cache & archive producer cache data flow Query flow


Quasar: Observations & Approach : Quasar: Observations & Approach Applications can tolerate errors in sensor data applications may not require exact answers: small errors in location during tracking or error in answer to query result may be OK data cannot be precise due to measurement errors, transmission delays, etc. Communication is the dominant cost limited wireless bandwidth, source of major energy drain Quasar Approach exploit application error tolerance to reduce communication between producer and server Two approaches Minimize resource usage given quality constraints Maximize quality given resource constraints


Quality-based Data Collection Problem : Quality-based Data Collection Problem Let P = < p[1], p[2], …, p[n] > be a sequence of environmental measurements (time series) generated by the producer, where n = now Let S = be the server side representation of the sequence A within- quality data collection protocol guarantees that for all i error(p[i], s[i]) <   is derived from application quality tolerance Sensor time series …p[n], p[n-1], …, p[1]


Simple Data Collection Protocol : Simple Data Collection Protocol sensor Logic (at time step n) Let p’ = last value sent to server if error(p[n], p’) >  send p[n] to server server logic (at time step n) If new update p[n] received at step n s[n] = p[n] Else s[n] = last update sent by sensor guarantees maximum error at server less than equal to  Sensor time series …p[n], p[n-1], …, p[1]


Exploiting Prediction Models : Exploiting Prediction Models Producer and server agree upon a prediction model (M, ) Let spred[i] be the predicted value at time i based on (M, ) sensor Logic (at time step n) if error(p[n], spred[n] ) >  send p[n] to server server logic (at time step n) If new update p[n] received at step n s[n] = p[n] Else s[n] = spred[n] based on model (M, )


Challenges in Prediction : Challenges in Prediction Simple versus complex models? Complex and more accurate models require more parameters (that will need to be transmitted). Goal is to minimize communication not necessarily best prediction How is a model M generated? static -- one out of a fixed set of models dynamic -- dynamically learn a model from data When should a model M or parameters  be changed? immediately on model violation: too aggressive -- violation may be a temporary phenomena never changed: too conservative -- data rarely follows a single model


Challenges in Prediction (cont.) : Challenges in Prediction (cont.) who does the model update? Server Long-haul prediction models possible, since server maintains history might not predict recent behavior well since server does not know exact S sequence; server has only samples extra communication to inform the producer Producer better knowledge of recent history long haul models not feasible since producer does not have history producers share computation load Both server looks for new models, sensor performs parameter fitting given existing models.


Archiving Sensor Data : Archiving Sensor Data Often sensor-based applications are built with only the real-time utility of time series data. Values at time instants <

Problem Formulation : Problem Formulation Let P = < p[1], p[2], …, p[n] > be the sensor time series Let S = < s[1], s[2], …, s[n] > be the server side representation A within archive quality data archival protocol guarantees that error(p[i], s[i]) < archive Trivial Solution: modify collection protocol to collect data at quality guarantee of min(archive , collect) then prediction model by itself will provide a archive quality data stream that can be archived. Better solutions possible since archived data not needed for immediate access by real-time or forecasting applications (such as monitoring, tracking) compression can be used to reduce data transfer


Data Archival Protocol : Data Archival Protocol Sensors compresses observed time series p[1:n] and sends a lossy compression to the server At time n : p[1:n-nlag] is at the server in compressed form s’ [1:n-nlag] within-archive s[n-nlag+1:n] is estimated via a predictive model (M, ) collection protocol guarantees that this remains within- collect s[n+1:] can be predicted but its quality is not guaranteed (because it is in the future and thus the sensor has not observed these values) …p[n], p[n-1], .. compress Sensor memory buffer Sensor updates for data collection Compressed representation for archiving processing at sensor exploited to reduce communication cost and hence battery drain


Slide16 : Piecewise Constant Approximation (PCA) Given a time series Sn = s[1:n] a piecewise constant approximation of it is a sequence PCA(Sn) = < (ci, ei) > that allows us to estimate s[j] as: scapt [j] = ci if j in [ei-1+1, ei] = c1 if j

Slide17 : Online Compression using PCA Goal: Given stream of sensor values, generate a within-archive PCA representation of a time series Approach (PMC-midrange) Maintain m, M as the minimum/maximum values of observed samples since last segment On processing p[n], update m and M if needed if M - m > 2archive , output a segment ((m+M )/2, n) Time Value Example: archive = 1.5 1 2 3 4 5


Slide18 : Online Compression using PCA PMC-MR … guarantees that each segment compresses the corresponding time series segment to within-archive requires O(1) storage is instance optimal no other PCA representation with fewer segments can meet the within-archive constraint Variant of PMC-MR PMC-MEAN, which takes the mean of the samples seen thus far instead of mid range.


Slide19 : Improving PMC using Prediction Observation: Prediction models guarantee a within- collect version of the time series at server even before the compressed time series arrives from the producer. Can the prediction model be exploited to reduce the overhead of compression. If archive> collect no additional effort is required for archival --> simply archive the predicted model. Approach: Define an error time series E[i] = p[i]-spred[i] Compress E[1:n] to within-archive instead of compressing p[1:n] The archive contains the prediction parameters and the compressed error time series Within-archive of E[I] + (M, ) can be used to reconstruct a within- archive version of p


Slide20 : Combing Compression and Prediction (Example)


Slide21 : Estimating Time Series Values Historical samples (before n-nlag) is maintained at the server within-archive Recent samples (between n-nlag+1 and n) is maintained by the sensor and predicted at the server. If an application requires q precision, then: if q  collect then it must wait for  time in case a parameter refresh is en route if q  archive but q < collect then it may probe the sensor or wait for a compressed segment Otherwise only probing meets precision For future samples (after n) immediate probing not available as an option


Slide22 : Experiments Data sets: Synthetic Random-Walk x[1] = 0 and x[i]=x[i-1]+sn where sn drawn uniformly from [-1,1] Oceanographic Buoy Data Environmental attributes (temperature, salinity, wind-speed, etc.) sampled at 10min intervals from a buoy in the Pacific Ocean (Tropical Atmosphere Ocean Project, Pacific Marine Environment Laboratory) GPS data collected using IPAQs Experiments to test: Compression Performance of PMC Benefits of Model Selection Query Accuracy over Compressed Data Benefits of Prediction/Compression Combination


Slide23 : Compression Performance K/n ratio: number of segments/number of samples


Slide24 : Query Performance Over Compressed Data “How many sensors have values >v?” (Mean selectivity = 50)


Slide25 : Impact of Model Selection K/n ratio: number of segments/number of samples. pred is the localization tolerance in meters Objects moved at approximately constant speed (+ measurement noise) Three models used: loc[n] = c loc[n] = c+vt loc[n] = c+vt+0.5at2 Parameters v, a were estimated at sensor over moving-window of 5 samples


Slide26 : Combining Prediction with Compression K/n ratio: number of segments/number of samples


Slide27 : QUASAR Client Time Series GPS Mobility Data from Mobile Clients (iPAQs) Latitude Time Series: 1800 samples Compressed Time Series (PMC-MR, ICDE 2003) Accuracy of ~100 m 130 segments


Query Processing in Quasar : Query Processing in Quasar Problem Definition Given sensor time series with quality-guarantees captured at the server A query with a specified quality-tolerance Return query results incurring least cost Techniques depend upon nature of queries Cost measures resource consumption -- energy, communication, I/O query response time


Aggregate Queries : Aggregate Queries S minQ = 2 maxQ = 7 countQ = 3 sumQ = 2+7+6 = 15 avgQ = 15/3 = 5


Processing Aggregate Queries (minimize producer probe) : Processing Aggregate Queries (minimize producer probe) MIN Query c = minj(si.high) b = c - query Probe all sensors where sj.low < b only s1 and s3 will be probed Sum Query select a minimal subset S’  S such that si in S’ (jpred) >= si in S(jpred)- query If query = 15, only s1 will be probed Let S = be set of sensors that meet the query criteria si.high = sipred[t] + jpred sj.low = sipred[t] - jpred a b c s1 s2 sn s3 s1 s2 s5 s3 s4 10 5 2 5 3


Minimizing Cost at Server : Minimizing Cost at Server Error tolerance of queries can be exploited to reduce processing at server. Key Idea Use a multi-resolution index structure (MRA-tree) for processing aggregate queries at server. An MRA-Tree is a modified multi-dimensional index trees (R-Tree, quadtree, Hybrid tree, etc.) A non-leaf node contains (for each of its subtrees) four aggregates {MIN,MAX,COUNT,SUM} A leaf node contains the actual data points (sensor models)


MRA Tree Data Structure : MRA Tree Data Structure Spatial View A B C D E F G A B C D E F G Tree Structure View


Slide33 : Non-Leaf Node Disk Page Pointers (each costs 1 I/O) Leaf Node Probe “Pointers” (each costs 2 messages) MRA-Tree Node Structure


Slide34 : Two sets of nodes: NP (partial contribution to the query) NC (complete contribution) Node Classification


Aggregate Queries using MRA Tree : Aggregate Queries using MRA Tree Initialize NP with the root At each iteration: Remove one node N from NP and for each Nchild of its children discard, if Nchild disjoint with Q insert into NP if Q is contained or partially overlaps with Nchild “insert” into NC if Q contains Nchild (we only need to maintain aggNC) compute the best estimate based on contents of NP and NC Q N


MIN (and MAX) : MIN (and MAX) 3 9 4 5 Interval minNC = min { 4, 5 } = 4 minNP = min { 3, 9 } = 3 L = min {minNC, minNP} = 3 H = minNC = 4 hence, I = [3, 4] Estimate Lower bound: E(minQ) = L = 3


Slide37 : MRA Tree Traversal A B C D E F G Progressive answer refinement until NP is exhausted Greedy priority-based local decision for next node to be explored based on: Cost (1 I/O or 2 messages) Benefit (Expected Reduction in answer uncertainty)


Adaptive Tracking of mobile objects in sensor networks : Adaptive Tracking of mobile objects in sensor networks Tracking Architecture A network of wireless acoustic sensors arranged as a grid transmitting via a base station to server A track of the mobile object generated at the base station or server Objective Track a mobile object at the server such that the track deviates from the real trajectory within a user defined error threshold track with minimum communication overhead.


Sensor Model : Sensor Model Wireless sensors : battery operated, energy constrained Operate on the received acoustic waveforms Signal attenuation of target object given by : Is(t) = P /4 r2 P : source object power r= distance of object from sensor Is(t) = intensity reading at time t at ith sensor Ith : Intensity threshold at ith sensor


Sensor States : Sensor States S0 : Monitor ( processor on, sensor on, radio off ) shift to S1 if intensity above threshold S1 : Active state ( processor on, sensor on, radio on) send intensity readings to base station. On receiving message from BS containing error tolerance shift to S2 S2 : Quasi-active (processor on, sensor on, radio intermittent) send intensity reading to BS if error from previous reading exceeds error threshold Quasar Collection approach used in Quasi-active state


Server side protocol : Server side protocol Server maintains: list of sensors in the active/ quasi-active state history of their intensity readings over a period of time Server Side Protocol convert track quality to a relative intensity error at sensors Send relative intensity error to sensor when sensor state = S1( quasi- active state) Triangulate using n sensor readings at discrete time intervals.


Basic Triangulation Algorithm (using 3 sensor readings) : Basic Triangulation Algorithm (using 3 sensor readings) P: source object power, Ii = intensity reading at ith sensor (x-x1)2 + (y- y1)2 = P/4 I1 (x-x2)2 + (y- y2)2 = P/4 I2 (x-x3)2 + (y- y3)2 = P/4 I3 Solving we get (x, y)=f(x1,x2,x3,y1,y2,y3, P,I1, I2 , I3, ) (x, y) More complex approaches to amalgamate more than three sensor readings possible Based on numerical methods -- do not provide a closed form equation between sensor reading and tracking location ! Server can use simple triangulation to convert track quality to sensor intensity quality tolerances and a more complex approach to track.


Adaptive Tracking : Mapping track quality to sensor reading : Adaptive Tracking : Mapping track quality to sensor reading Intensity ( I2 ) time  I2 Intensity ( I3 ) time  I3 t i t( i+1 ) t i t( i+1 ) t i t( i+1 ) X (m) Y (m) Claim 1 (power constant) Let Ii be the intensity value of sensor If then, track quality is guaranteed to be within track where and C is a constant derived from the known locations of the sensors and the power of the object. Claim 2 (power varies between [Pmin , Pmax ]) If then track quality is guaranteed to be within track where C’ = C/ P2 and is a constant . The above constraint is a conservative estimate. Better bounds possible  track


Adaptive Tracking: prediction to improve performance : Communication overhead further reduced by exploiting the predictability of the object being tracked Static Prediction : sensor & server agree on a set of prediction models only 2 models used: stationary & constant velocity Who Predicts: sensor based mobility prediction protocol Every sensor by default follows a stationary model Based on its history readings may change to constant velocity model (number of readings limited by sensor memory size) informs server of model switch Adaptive Tracking: prediction to improve performance


Actual Track versus track on Adaptive Tracking (error tolerance 20m) : Actual Track versus track on Adaptive Tracking (error tolerance 20m) A restricted random motion : the object starts at (0,d) and moves from one node to another randomly chosen node until it walks out of the grid.


Energy Savings due to Adaptive Tracking : Energy Savings due to Adaptive Tracking total energy consumption over all sensor nodes for random mobility model with varying track or track error. significant energy savings using adaptive precision protocol over non adaptive tracking ( constant line in graph) for a random model, prediction does not work well !


Energy consumption with Distance from BS : Energy consumption with Distance from BS total energy consumption over all sensor nodes for random mobility model with varying base station distance from sensor grid. As base station moves away, one can expect energy consumption to increase since transmission cost varies as d n ( n =2 ) adaptive precision algorithm gives us better results with increasing base station distance


Challenges & Ongoing Work : Challenges & Ongoing Work Ongoing Work: Supporting a larger class of SQL queries Supporting continuous monitoring queries Larger class of sensors (e.g., video sensors) Better approaches to model fitting/switching in prediction In the future: distributed Quasar architecture optimizing quality given resource constraints supporting applications with real-time constraints dealing with failures


The DAS Project** : The DAS Project** Goals: Support Database as a Service on the Internet Collaboration: IBM (Dr. Bala Iyer) UCI (Gene Tsudik) ** Supported in part by NSF ITR grant entitled “Privacy in Database as a Service” and by the IBM Corporation


Software as a Service : Software as a Service Get … what you need when you need Pay … what you use Don’t worry … how to deploy, implement, maintain, upgrade


Software As a Service: Why? : Software As a Service: Why? Advantages reduced cost to client pay for what you use and not for hardware, software infrastructure or personnel to deploy, maintain, upgrade… reduced overall cost cost amortization across users Better service leveraging experts across organizations Driving Forces Faster, cheaper, more accessible networks Virtualization in server and storage technologies Established e-business infrastructures Already in Market ERP and CRM (many examples) More horizontal storage services, disaster recovery services, e-mail services, rent-a-spreadsheet services etc. Sun ONE, Oracle Online Services, Microsoft .NET My Services etc Better Service for Cheaper


Database As a Service : Database As a Service Why? Most organizations need DBMSs DBMSs extremely complex to deploy, setup, maintain require skilled DBAs with high cost


What do we want to do? : What do we want to do? Database as a Service (DAS) Model DB management transferred to service provider for backup, administration, restoration, space management, upgrades etc. use the database “as a service” provided by an ASP use SW, HW, human resources of ASP, instead of your own Application Service Provider (ASP) Server BUT….


Challenges : Challenges Economic/business model? How to charge for service, what kind of service guarantees can be offered, costing of guarantees, liability of service provider. Powerful interfaces to support complete application development environment User Interface for SQL, support for embedded SQL programming, support for user defined interfaces, etc. Scalability in the web environment overheads due to network latency (data proxies?) Privacy and Security Protecting data at service providers from intruders and attacks. Protecting clients from misuse of data by service providers Ensuring result integrity


Data privacy from service provider : Data privacy from service provider Encrypted User Database User Data The problem is we do not trust “the service provider” for sensitive information! Fact 1: Theft of intellectual property due to database vulnerabilities costs American businesses $103 billion annually Fact 2: 45% of those attacks are conducted by insiders! (CSI/FBI Computer Crime and Security Survey, 2001) encrypt the data and store it but still be able to run queries over the encrypted data do most of the work at the server Server Application Service Provider Untrusted Server Site


System Architecture : System Architecture Server Site Original Query Server Side Query Encrypted Results Actual Results Service Provider Client Site Client Side Query ? ? ?


NetDB2 Service : NetDB2 Service Developed in collaboration with IBM Deployed on the Internet about 2 years ago Been used by 15 universities and more than 2500 students to help teaching database classes Currently offered through IBM Scholars Program


MARS Project** : MARS Project** Goals: integration of similarity retrieval and query refinement over structured and semi-structured databases to help interactive data analysis/mining **Supported in part by NSF CAREER award, NSF grant entitled “learning digital behavior” and a KDD grant entitled “Mining events and entities over large spatio-temporal data sets”


Similarity Search in Databases (SR) : Similarity Search in Databases (SR) Alice Honda sedan, inexpensive, after 1994, around LA


Query Refinement (QR) : Refined Results Query Refinement (QR)


Why are SR and QR important? : Why are SR and QR important? Most queries are similarity searches Specially in exploratory data analysis tasks (e.g., catalog search) Users have only a partial idea of their information need Existing Search technologies (text retrieval, SQL) do not provide appropriate support for SR and (almost) no support for QR. Users must artificially convert similarity queries to keyword-searches or exact-match queries Good mappings difficult or not feasible Lack of good knowledge of the underlying data or its structure Exact-match may be meaningless for certain data types (e.g., images, text, multimedia)


Slide62 : Similarity Access and Interactive Mining Architecture Search Client Query Session Manager Query Log Manager/Miner Similarity Query Processor Feedback-based Refinement Method ORDBMS Similarity Operators Types Feedback Table Database Query Log Answer Table Scores Table Refinement Manager History-based Refinement Method Query/Feedback Ranked Results Initial Query Feedback Ranked Results Query Results Schemes Ranking Rules Legend: --- logging __ Process


MARS Challenges... : MARS Challenges... Learning queries from user interactions user profiles past history of other users Efficient implementation of similarity queries refined queries Role of similarity queries in OLAP interactive data mining


Slide64 : Query-Session Manager -parse the query - check query validity -generate schema for support tables - maintain sessions registry Similarity Query Processor -executes query on ORDBMS - ranks results (e.g. can exclude already see tuples, etc) - logs query(query or Top-k) Refinement Manager - maintains a registry of query refinement policies (content/collaborative) - generates the scores table - identifies and invokes intra-predicate refiners. Query Log Manager/Miner - maintains query log . Initial-Final pair . Top-K results . Complete trajectory - Query-query similarity (can have multiple policies) - Query clustering


Text Search Technologies (Altavista, Verity, Vality, Infoseek) : Text Search Technologies (Altavista, Verity, Vality, Infoseek) Strengths support ranked retrieval can handle missing data, synonyms, data entry errors Approach convert enterprise structured data into a searchable text index. Limitations cannot capture semantics of relationships in data cannot capture semantics of non-text fields (e.g., multimedia) limited support for refinement or preferences in current systems cannot express similarity queries over structured or semi-structured data (e.g., price, location) Movies Actors Directors Al Pacino acted in a movie directed by Francis Ford Coppola


SQL-based Search TechnologiesOracle, Informix, DB2, Mercado : SQL-based Search Technologies Oracle, Informix, DB2, Mercado Approach translate similarity query into exact SQL query. Strengths support structured as well as semi-structured data support for arbitrary data types Scalable attribute-based lookup Limitations translation is difficult or not possible difficult to guess right ranges causes near misses not feasible for non-numeric fields cannot rank answers based on relevance does not account for user preference or query refinement select * from user_car_catalog where model = Honda Accord and 1993 £ year £1995 and dist(90210) £ 50 and price < 5000