KEWS: A Method Evaluation of Workload Simulation based on KPIs

For end-to-end performance testing, workload simulation is an important method to enhance the real workload while protecting user privacy. To ensure the effectiveness of the workload simulation, it is necessary to dynamically evaluate the similarity of system inner status using key performance indicators(KPIs), which provide a comprehensive record of the system status, between the simulated workload and real workload by injecting workload into the system. However, due to the characteristics of KPIs, including large data size, amplitude differences, phase shifts, non-smoothness, high dimension, and Large numerical span, it is unpractical to evaluation on the full volume of KPIs and is challenging to measure the similarity between KPIs. In this paper, we propose a similarity metric algorithm for KPIs, ESBD, which describes both shape and intensity similarity. Around ESBD, a KPIs-based quality evaluation of workload simulation(KEWS) was proposed, which consists of four steps: KPIs preprocessing, KPIs screening, KPIs clustering, and KPIs evaluation. These techniques help mitigate the negative impact of the KPIs characteristics and give a comprehensive evaluation result. The experiments conducted on Hipstershop, an open-source microservices application, show the effectiveness of the ESBD and KEWS.