Find in Library

Abstract Serverless computing is stepping forward to provide a cloud environment that mainly focuses on managing infrastructure, resources and configurations on the behalf of a user. Research in this field can’t rely on commercial providers such as AWS and Azure, as their inflexibility and cost often limits the required levels of reproducibility and scalability. Therefore, simulators have been opted as an alternative solution by the research community. They offer a reduced-cost and easy-setup environment. To get respectable precision, simulators use real traces collected and offered by commercial providers. These traces represent comprehensive information of executed tasks that reflect user behaviour. Due to serverless computing’s recency, typical workload traces employed by IaaS simulators are not well adoptable to the new computing model. In this paper, we propose an approach for generating realistic serverless traces. We enhance our previous generator approach that was based on the Azure Functions dataset. Our new, genetic algorithm based approach improves the statistical properties of the generated traces. We also enabled arbitrary scaling of the workload, while maintaining real users’ behaviour. These advances further support reproducibility in the serverless research community. We validated the results of our generator approach using the coefficient of determination ( $$R^2$$ R 2 ), which shows that our generated workload closely matches the original dataset’s characteristics in terms of execution time, memory utilisation as well as user participation percentage. To demonstrate the benefits of the reusability of the generated traces, we applied them with a diverse set of simulators and shown that they offer reproducible results independently of the simulator used.