Synthetic workload generation for stress testing session-based systems

A large portion of the world's computing infrastructure is dedicated to the support of enterprise applications such as e-commerce and business process management. Such applications are often business critical and require effective stress testing techniques to understand how they will perform under various workload conditions. The workloads of these systems are characterized by sessions of interdependent requests that often cause and exploit dynamically generated responses. Inter-request dependencies must be reflected in synthetic workloads for these systems to stress application functions in a representative manner. In addition, to support tasks such as capacity planning and service level management, it should be possible to have flexible control over many different workload characteristics that are likely to affect system performance. Simultaneously satisfying these requirements can make manual construction of synthetic workloads difficult and time consuming.;This thesis presents a synthetic workload generation approach that allows session-based systems to be stress tested in a representative, flexible, and automated manner. The proposed approach uses a workload model which includes attributes that can influence the performance of session-based systems. Statistical characterizations can be specified for each of these attributes. The approach employs a trace generation algorithm that automates the tasks of ensuring correct inter-request dependencies and simultaneously incorporating the desired characteristics for many different attributes into a synthetic workload. The chief advantage of the approach is the fine control it offers over workload characteristics. For example, it permits the characterizations of one or more attributes to be changed at a time while keeping those of others unchanged so that a system's sensitivity to those characteristics alone can be established.;The approach is validated in a number of ways through a stress testing case study of an e-commerce system. The study shows that the approach can produce synthetic workloads with characteristics that match those observed in real systems. The study also demonstrates that a synthetic workload created with the approach causes a nearly identical performance-impact on the system as the original workload it modeled. Finally, by using the approach to generate various workloads that differ from one another in carefully controlled ways, the study reveals that the system's performance is sensitive to the characteristics of attributes included in the workload model.