The COMFORT prototype

Abstract

1 Project Goals COMFORT stands for "Gs?mfortable Perffnmmce liming". The long-term goal that we pursue in the COMFORT project is to automate, to the largest possible extent, the performance tuning of database systems. Tuning of database systems depends critically on the expertise and experience of system administrators and other human tuning experts which are responsible for the setting of system parameters. The purpxe of such system parameters, or "tuning knobs", is to adapt the system to the speeific characteristics of a given workload. With a wider use of OLTP and other multi-user database applications, on the one hand, and a lack of qualified tuning experts, on the other hand, there is a strong need for simplifying the tricky job of human administrators and ideally automating at least some critical tuning decisions. Our approach is to derive appropriate tuning heuristics, or "rules of thumb", from quantitative performance models for individual tuning problems, and to incorporate such heuristics in an adaptive or "self-tuning" system architecture. Thus, the goal of automatic performance tuning entails two lines of research. On the one hand, we are investigating specific tuning problems, with emphasis on the challenging problems that are posed by multi-user parallel database systems. On the other hand, we are aiming at architectural principles of a database system that can automatically adapt itself to the workload. Such a system should ideally be q self-tuning in that it does not require human intervention, q responsive in that it can quickly react to signit3cant changes of the workload, q robust in that it guarantees acceptable performance even under "stress" conditions, and q Iow-cost in that it can make decisions on-line at acceptable costs of bookkeeping, gathering of statistics, evaluation of analytic models, etc. Within this framework of a self-tuning database system architecture, we have been addressing the tuning issues of data placement in multklisk architectures, load control, parallelization and optimization of complex queries, and processor allocation for multi-user pamllel database systems [1].