Transaction-Aware Software Performance Prediction

Model-based Quality of Service (QoS) prediction approaches enable an early identification of bottlenecks in the design of complex software systems. In addition, they allow software architects to evaluate design decisions against each other. In data-intensive business applications, database accesses have a significant impact on a system’s QoS. Therefore, these accesses need to be considered in order to derive meaningful QoS predictions. Database management systems use concurrency control protocols to guarantee that concurrently executed transactions are isolated against each other. Their employed isolation mechanisms, e.g. locks, may affect the QoS of the whole software system. Palladio is a framework for QoS predictions on an architectural level. It comprises an architectural model, the Palladio Component Model (PCM), out of which QoS predictions can be derived via analytical and simulative techniques. Thus far, the PCM doesn’t include an explicit modeling of databases and database accesses. In order to derive precise QoS predictions for dataintensive systems, software architects and component developers would have to manually specify the system’s isolation mechanism in each PCM instance. Due to the complexity of isolation mechanisms, their QoS implications are hence commonly disregarded by PCM instances that describe data-intensive application systems. This master thesis extends Palladio to model database accesses and simulate concurrency control. A model based on the two-phase locking (2PL) protocol is developed. The model predicts conflicts between transactions on an abstraction level suited for architectural performance predictions. In order to enable the prediction of delays caused by concurrency control, the conflict model is coupled with the existing PCM simulator EventSim. An evaluation compares predictions of the enhanced simulator with measurements from a MySQL database. Initial differences between measured and predicted results could be traced back to properties specific to the 2PL implementation of MySQL. These differences could be considerably reduced through manual adjustments to the evaluated system’s modeling. The evaluation results indicate that our extension validly represents the performance impact of 2PL.