Architecture-Level Software Performance Models for Online Performance Prediction

Modern enterprise systems often have distributed application architectures composed of many independent services running in a heterogeneous environment (Papazoglou et al., 2007). In such systems, applications are customized and new services are composed and deployed on-the-fly subjecting the system resources to varying workloads. Moreover, existing services, given their loosely-coupled nature, can evolve independently of one another. Managing system resources in such environments to ensure acceptable end-to-end performance and availability, while at the same time optimizing resource utilization, is a challenge (Armbrust et al., 2009; Brooks, 2011). Service providers are often faced with questions such as: What would be the performance of a given service if the workload continues to evolve as currently observed? What performance would a new service deployed on the infrastructure exhibit and how much resources should be allocated to it? How should the workloads of the running services be partitioned among the available resources such that performance objectives are met and resources are utilized e ciently? Answering such questions requires the ability to predict at run-time, i.e., during system operation, how the performance of running services would be a↵ected if the workload or the system configuration changes. We refer to this as online performance prediction. Existing approaches to online performance prediction (e.g., Nou et al. (2009); Li et al. (2009); Jung et al. (2010)) are based on predictive stochastic performance models such as (layered) Queueing Networks or Queueing Petri Nets. Such models normally abstract the system at the resource level without explicitly taking into account the software architecture. Services are typically modeled as black boxes, i.e., the control flow and the dependencies between software components are neglected. Detailed models that explicitly capture the software architecture and configuration exist in the literature (e.g., Becker et al. (2009); Grassi et al. (2007); Bertolino and Mirandola (2004)). They are typically software architecture models annotated with descriptions of the system’s performance-relevant behavior. Such models, often referred to as architecture-level performance models, are intended to evaluate alternative system designs at design-time and/or predict the system performance for capacity planning purposes. However, there are fundamental di↵erences between o✏ine and online scenarios for performance prediction. This leads to di↵erent requirements on the underlying performance abstractions of the system architecture and the respective performance prediction techniques suitable for use at design-time versus run-time. To realize proactive performance and resource management, overload situations should be anticipated and suitable reconfiguration actions must be found and triggered before Service Level Agreements (SLAs) are violated. In this context, there are situations where performance prediction results need to be available very fast to adapt the system before performance issues arise, as well as situations where a fine-grained prediction is needed to find an e cient system configuration. In this thesis, we develop novel architecturelevel performance model abstractions for component-based software systems specifically designed for online use. We provide modeling and prediction facilities that enable online performance prediction during system operation. Performance questions can be answered