Goal-oriented Performance Control for Transaction Processing

The performance of current transaction processing systems largely depends on human experts for administration and tuning. These experts have to specify a multitude of internal control parameters in different subsystems for which finding appropriate settings is very difficult. Another shortcoming of manual system administration lies in its inability to quickly react to changing system and workload conditions. In order to overcome these limitations we advocate for an automatic and adaptive performance control. To simplify the administration we pursue a "goaloriented" approach that aims at automatically enforcing external performance objectives, in particular response time goals. The effective implementation of such a scheme poses a multitude of largely unsolved challenges. Our approach is based on a feedback loop for automatic detection and correction of performance problems and requires comparatively few extensions over existing TP systems. We have implemented various control strategies for bottleneck determination and resolution within a detailed simulation system. Results of some initial simulation experiments are analyzed.