Predicting software performance in symmetric multi-core and multiprocessor environments

With today’s rise of multi-core processors in the desktop market and in common server systems, concurrency becomes a ubiquitous challenge in software development. Concurrency allows the improvement of software performance by exploiting available processor cores. However, possible performance gained by concurrency can be limited by software bottlenecks or inherently sequential parts of the application. Performance prediction methods help software architects to find software bottlenecks in multiprocessing environments during the design phase. Unfortunately, in highly concurrent systems, current prediction methods result in errors up to several orders of magnitude. They neglect important influences of the operating system schedulers and middleware leading to erroneous predictions. Therefore, this thesis addresses the challenge of performance prediction in symmetric multiprocessing environments. It proposes a performance modelling framework for operating system schedulers such as implemented in Windows or Linux. The modelling approach reflects the performance-relevant features of operating system schedulers and can be customised to represent the system under study. It can also be combined with other prediction methods to increase their prediction accuracy. The influence of the middleware on software performance is addressed by a performance modelling approach to message-oriented middleware. The approach combines abstract performance models with measurements and, thus, can omit implementation details of vendors. A series of case studies conducted in the scope of this thesis demonstrates that both techniques reduce the prediction error to less than 5% to 10% in most cases.