Organizations are striving for competitive advantage. As a result, business optimization is being pushed to new heights in terms of volume and speed. Areas such as customer profitability, campaign profitability or customer insight for better service are driving new analytical challenges as well as new algorithms over large volumes of data.

Object-relational database systems allow users to define new types and functions. This typing extension allows structural semantics to be applied to new data types (non-SQL92 data types). However, defining new, large, complex objects within a tuple is not always practical. Practicality alone requires that large object data structures be used. This implies that predicate evaluation can be performed on tuple columns that are physically stored outside of a tuple's physical record. This out-of-line storage mechanism is commonly used today for column values that are physically large (implemented with a large object descriptor in the tuple pointing to a large object data structure). This type of predicate-based evaluation presents new optimizer and run-time challenges to the database system on clustered and massively parallel processor shared-nothing architectures.

This paper introduces a novel approach to optimizing and monitoring database queries which involve operations on multiple data types in a parallel multimedia engine. Our approach uses dataflow graphs to represent the multimedia operations in a query. We have extended the Actors parallel programming model by designing an agent model for query execution that incorporates extensions for efficient data streaming, agent migration, and agent cloning. We incorporate algorithms for dynamic run-time workload control of both the number of queries in the system as well as the number of instances of agents executing multimedia operators. We describe an approach for capturing cost statistics for user-defined functions that is used by the system to estimate and schedule the execution of nested user-defined functions.

The Prospector Multimedia Object Manager prototype is a general-purpose content analysis multimedia server designed for massively parallel processor environments. Prospector defines and manipulates user defined functions which are invoked in parallel to analyze/manipulate the contents of multimedia objects. Several computationally intensive applications of this technology based on large persistent datasets include: fingerprint matching, signature verification, face recognition, and speech recognition/translation [OIS96].

The Teradata Multimedia Object Manager is a general-purpose content analysis multimedia server designed for symmetric multiprocessing and massively parallel processing environments. The Multimedia Object Manager defines and manipulates user-defined functions (UDFs), which are invoked in parallel to analyze or manipulate the contents of multimedia objects. Several computationally intensive applications of this technology, which use large persistent datasets, include fingerprint matching, signature verification, face recognition, and speech recognition/translation.