Rule-based systems have been used extensively by the AI community in implementing knowledge-based expert systems. A current trend in the database community is to use rules for the purpose of providing inferential capabilities for large database applications. However, in the database context, performance of rule-program processing has proved to be a major stumbling block particularly in data-intensive and real-time applications. Similar work in the AI community has demonstrated the same problems mainly due to the predominantly sequential semantics of the underlying rule languages. In a previous paper, we presented an incremental evaluation algorithm forDatalog
⌝* programs. We show the general applicability of the incremental evaluation algorithm to a broad class ofDatalog-style rule languages with varying operational semantics, denoted byDatalog** to emphasize this generality. We present the proof of correctness of the algorithm within the Datalog** frame work, and its extension for the PARULEL rule language, a principle component of the PARADISER (PARAllel and DIStributed Environment for Rules) rule processing environment for databases. We discuss the meta-rule formalism of PARULEL and argue that it provides a means for programmable operational semantics which separates control from the base logic of a rule program. This allows the realization of a wide range of operational semantics including those of Datalog and OPS5.

Incremental database rule processing in PARADISER

An overview is given on the development history of the key technologies of data memory and storage. For each of them the essential characteristics for their application in computer systems and also their relation to competing preceeding or replacing technologies is described. Comparisons of characteristic values of speed and packaging densities along with a list of historical milestones are added. As a conclusion this overview extending over many decades shows that the progress of this field, which is of dominating importance for system design and application, has not yet reached saturation of progress now and for many years to come. 1. The Role of Data Memory and Storage From the very beginning of data processing, data memory and storage has always played a dominant role in the architecture, the design, the implementation and the operation of data processing systems, be it that we regard a system of the historical past, of the early times of electronic computers, or of our times. The economic importance of data memory and storage is substantial: Almost one third of the world-wide yearly production value of the total systems hardware is devoted to this sector. The demand of computer users for more data memory and storage at improved cost/performance can still not be satisfied, even with the enormous investments in research, development and industry in the past decades

Computer Systems and Software Engineering

Query optimization for sequential execution of non-recursive queries has reached a high level of sophistication in commercial DBMS. The successful application of parallel processing for the evaluation of recursive queries will require a query optimizer of comparable sophistication. The groundwork for creating this new breed of query optimizer will consist of a combination of theoretical insight and empirical investigation. Restricting our attention to linear recursive queries, we illustrate this process by developing a family of query processing strategies and, through experiments on a parallel computer, obtaining the basic information needed for an optimizer's heuristics.

Strategies for Parallel Linear Recursive Query Processing

In the last decade and half, scientific and commercial databases have grown at a fast pace as increasingly many enterprises have begun offering and using database-centered services. Concomitantly, new application requirements have placed unusual processing demands on existing databases. Many researchers believe that next generation database systems will be required to provide efficient and scalable facilities for high level inference and complex query processing.
In this thesis, we address the efficiency and scalability problem in the context of Database Rule Processing Systems (DRPS). We propose a scalable strategy for the efficient parallel processing of rule programs applied to massive databases. We introduce predictive load balancing techniques in a synchronous model of rule execution, where the variance in runtime of the distributed sites is minimized per cycle of rule processing, thus increasing utilization and speedup. We demonstrate that static load balancing techniques are insufficient, and thus low overhead dynamic load balancing is the key to successful scaling. We present a form of dynamic load balancing that is based upon predicting future system loads, rather than conventional demand-driven or adaptive approaches that monitor current system state. We utilize a combination of meta-data on the database and observed system performance to predictively formulate load balancing decisions over a parallel ensemble. We analyze a number of possible predictive dynamic load balancing protocols by isoefficiency analysis to guide the design of a distributed rule processing system. Our results are outlined as follows. We develop a set-oriented rule language with parallel execution semantics to overcome the constraints imposed by tuple-oriented rule languages. We initially assume a replicated database configuration for simplicity, and we explore automatic work partitioning techniques, including heuristic determination of partitioning attributes, predictive dynamic load balancing protocols to apply them, a "self-monitoring" principle for system modelling that predicts system performance in the near term based upon recent performance history, and extensive use of statistics on the underlying database (meta data) to achieve load balancing and hence scaling. We then extend our basic results and develop comprehensive algorithms for predictively balanced inference support and complex query processing over distributed database configurations. Our approach enables large-scale expert databases. It is expected that the resulting quantitative improvements will lead to qualitative improvements in the applications built within this framework.

Runtime reorganization of parallel and distributed expert database systems

Logic languages such as Datalog have been proposed as a method for specifying flexible and customisable static analysers. Using Datalog, various classes of static analyses can be expressed precisely and succinctly, requiring fewer lines of code than hand-crafted analysers. In this paradigm, a static analysis specification is encoded by a set of declarative logic rules and an o -the-shelf solver is used to compute the result of the static analysis. Unfortunately, when large-scale analyses are employed, Datalog-based tools currently fail to scale in comparison to hand-crafted static analysers. As a result, Datalog-based analysers have largely remained an academic curiosity, rather than industrially respectful tools. This thesis outlines our e orts in understanding the sources of performance limitations in Datalog-based tools. We propose a novel evaluation technique that is predicated on the fact that in the case of static analysis, the logical specification is a design time artefact and hence does not change during evaluation. Thus, instead of directly evaluating Datalog rules, our approach leverages partial evaluation to synthesise a specialised static analyser from these rules. This approach enables a novel indexing optimisations that automatically selects an optimal set of indexes to speedup and minimise memory usage in the Datalog computation. Lastly, we explore the case of more expressive logics, namely, constrained Horn clause and their use in proving the correctness of programs. We identify a bottleneck in various symbolic evaluation algorithms that centre around Craig interpolation. We propose a method of improving these evaluation algorithms by a proposing a method of guiding theorem provers to discover relevant interpolants with respect to the input logic specification. The culmination of our work is implemented in a general-purpose and highperformance tool called Souffl´e. We describe Souffl´e and evaluate its performance experimentally, showing significant improvement over alternative techniques and its scalability in real-world industrial use cases.

Simona R. Cohen

Papers

Why a single parallelization trategy is not enough in knowledge bases