In last year’s ICDE panel in New Orleans [l], we examined the question of whether database research is able to provide leadership to database industries. There was a consensus that with the maturing of the field, we should now focus on new areas where we can leverage off our rich experience in database research. The question of what problem to work on next has always been a difficult one to answer and we suspect that it will not get easier. Even then, it will be rewarding to examine the question of how the successful and not so successful threads of research came into being and what caught our fancy and why. Specifically, we will seek the perspective of the panelists on the following questions:

Received June 1993. revised September 1993.

Though in general, current database systems adequately support application development and operation for online transaction processing (OLTP), increasing complexity of applications and throughput requirements reveal a number of weaknesses with respect to the data model and implementation techniques used. By presenting the experiences gained from a case study of a large, high volume stock trading system, representative for a broad class of OLTP applications, it is shown, that this particularly holds for dealing with high frequency access to a small number of data elements (hot spots). As a result, we propose extended data types and several novel mechanisms, which are easy to use and highly increase the expressional power of transaction oriented programming, that effectively cope with hot spots. Moreover, their usefulness and their ability to increased parallelism is exemplified by the stock trading application.

In shared database management systems, there is a need for dynamically setting a large number of operational parameters, like transaction routing tables, assignment of central resources, multiprogramming level, etc. It is shown that static techniques as they are applied in current centralized DBMSs are inadequate for exploiting the performance potential of a shared DBMS. We therefore plead for using dynamic load control methods for adaptive modification of the DBMS's operational parameters. After a brief survey of the state of the art in database load control, we outline an approach for load balancing in a shared system, which consists of both analytic models and heuristics. The key concepts are illustrated by a simple example, and a number of areas for future research are identified.

Various logging and recovery techniques for centralized transaction-oriented database systems under performance aspects are described and discussed. The classification of functional principles that has been developed in a companion paper is used as a terminological basis. In the main sections, a set of analytic models is introduced and evaluated in order to compare the performance characteristics of nine different recovery techniques with respect to four key parameters and a set of other parameters with less influence. Finally, the results of model evaluation as well as the limitations of the models themselves are discussed.

This paper describes the concepts and implementation of a design aid for the internal schema of an existing CODASYL-like database system. It allows for tailoring the storage structure level to a given logical schema and a specified workload. According to the 1978 CODASYL report, our DBMS provides two levels of schema declaration, the DDL-level for logical schema description and a DSDL-like level for specifying the storage structures to implement the objects of the logical schema. The repertoire of storage structures supported by our system is a good internal schema are basically heuristic. This approaach is justified by weighing its advantages and shortcomings against those of analytic models and simulation. Finally, some preliminary user experiences with a pilot version are related.

In many large database applications there are certain elements mostly containing aggregate information, which are very frequently referred to (read and modified) by many transactions. If access to such fields has to obey to conventional two-phase lock protocols (1,2), transactions will be serialized in front of these "hot spots", i.e. the degree of parallelism is reduced. To avoid this kind of lock contention some improved lock protocols have been proposed, the most interesting of which is the one implemented in IMS Fast Path (3,4), where add and subtract may be performed concurrently on numerical fields, since backout is always possible with the unique inverse of each operand. A similar scheme is proposed in (10). We expand this idea to parallel readers and writers on numerical data types, proving that under certain conditions the result of such concurrent operations is consistent in the sense that it is equal to some serial schedule (2,5).

A special UNDO log algorithm combining the advantages of update in place strategies with the shadow page concept is introduced. It is to support applications demanding high rates of possibly long update transactions as well as fast UNDO recovery. By doubling the disk space for the database, UNDO loggng can be performed with a minimum of additional I/O operations. Thus, we gain a performance comparable to systems without any UNDO loggg. Since in this algorithm one logical slot consists of two physical blocks, the name TWIST (twin slot) algorithm is proposed.