An Integration Framework for Sensor Networks and DSMSs

Daniel Abadi, Wolfgang Lindner, Samuel Madden & Michael Stonebraker

What

DSMSs [1,4,2] were developed to support an emerging class of applications -monitoring applications - that proved problematic for traditional DBMSs. Monitoring applications are applications that monitor continuous streams of data. Their fundamental data-active/human-passive, real-time, trigger oriented model is difficult to support in the traditional human-active/data-passive trigger-as-secondclass-citizen DBMS model. DSMSs are better suited to supporting these applications by organizing query operators in a work-flow diagram and allowing data to actively stream through these operators, which transform the input data according to a continuous query plan. The performance of a DSMS can be measured using a Quality of Service (QoS) metric in which an application can specify the utility of observed latency, throughput, or quality of result tuples that reach the application.

Sensor networks consist of multiple sensor nodes: small, battery-powered, wireless computers that can contain any number of sensors that can measure the surrounding environment (eg. temperature, light, acceleration, or position). Power is if utmost importance. If used naively, individual sensor nodes will deplete their energy supplies in only a few days. To save power, communication distances are usually reduced to below loo feet, resulting in most real deployments making use of multi-hop communication where intermediate nodes relay information for their peers. Sensor networks can serve as a data source to monitoring applications and have limited query processing power (simple query operators can be performed on tuples as they are relayed from node to node on their way to a basestation) [3].

Our work attempts to integrate these two systems so that application specified queries are transformed to unified, optimized query plans that can be executed across sensor/DSMS boundaries. This integration is non-trivial because of the different architectures and assumptions of these systems. The sensor network performs the same operation in parallel across many different nodes in a lossy and power constrained environment while the DSMS performs only one instance of an operation on a server node with fewer power, CPU, and storage constraints. Optimization of this query plan is particularly difficult due to the applications' insensitivity of the constraints of the sensor network.

Why

Integration of sensor networks and DSMSs is important because sensor networks serve as a natural data source to DSMSs. In return DSMSs are capable of executing much more complex operations on the data than nodes in the sensor network, allowing a wider variety of queries to be performed on sensor produced data.

How

We integrate sensor networks with DSMSs by inserting into the architecture a proxy and wrappers at the interface between the sensor network and the DSMS system, as in Figure 1.

Figure 1 : Integration Architecture

The proxy serves to relay the capabilities and constraints of what query processing can be done in the sensor network. This allows the DSMS to allocate operations to the sensor network in its query plans that the network is capable of performing. The proxy also converts Squal [1] operators into operations that the sensor network can perform (such as adding sample operators and converting filter boxes into selection operators).

The wrappers keep statistics of individual properties of each sensor network such as average loss rate, power utilization, and what attributes the sensors in the network are able to sample. In addition, it controls the mapping of operators to individual nodes and organizes the distribution of tuples from static tables to their correct node destinations (in a static join). It passes all the statistics that it maintains to the proxy which can then forward them to the DSMS for use in optimization.

The DSMS optimizer needs these statistics to consider the consequences of operator movements from the DSMS to the sensor network and sample rate adjustments in the sensor network. For example, if application QoS indicates that output quality is critical, moving an operation such as a join into the sensor network might result in an increased amount of dropped tuples, decreasing quality further. The DSMS can use the wireless loss statistics from the sensor network to which it is considering moving the join to estimate the effects on QoS of such a move. If QoS indicates that network lifetime needs to be optimized, the DSMS can attempt to move high cardinality operators out of the network or low cardinality operators into the network to reduce power utilization (by reducing network transmissions). Alternatively, it can request that the sensors sample at a lower rate.Part of this integration effort includes extending sensor network query processing capabilities, such as adding a join operator to [3].

Progress:

We built a first prototype of the integration architecture which was demonstrated at the VLDB 2004 conference [5].

Research Support:

We gratefully acknowledge funding from NSF under grant number lTR 115-0325703.

References:

[1] Daniel J. Abadi, Don Carney, Ugur Çetintemel, Mitch Cherniack, Christian Convey, Sangdon Lee, Michael Stonebraker, Nesime Tatbul, and Stan Zdonik. Aurora: A new model and architecture for data stream management. VLDB Journal, September 2003.

[2] 5. Chandrasekaran, A. Deshpande, M. Franklin, J. Hellerstein, W. Hong, S. Krishramurthy, S. Madden, V. Raman, F. Reiss, and M. Shah. TelegraphCQ: Continuous Dataflow Processing for an Uncertain World. In CIDRConference, January 2003.

[3] Samuel R. Madden, Michael J. Franklin, Joseph M. Hellerstein, and Wei Hong. The Design of an Acquisitional Query Processor for Sensor Networks. In ACM SIGMOD Conference, June 2003.

[4] R. Motwani, J. Widom, A. Arasu, B. Babcock, S. Babu, M. Datar, G. Manku, C. Olston, J. Rosenstein, and R. Varma. Query Processing, Approximation, and Resource Management in a Data Stream Management System. In CIDR Conference, January 2003.

[5] Daniel J. Abadi, Wolfgang Lindner, Samuel Madden, and Joerg Schuler. An Integration Framework for Sensor Networks and Data Stream Management Systems. In VLDB Conference, August 2004.