Analyzing users in parallel computing: A user-oriented study
The performance of parallel schedulers is a crucial factor in the efficiency of high performance computing environments. Scheduler designs for practical application focusing on improving certain metrics can only be achieved, if they are evaluated in realistic testing environments. Since real users submit jobs to their respective system, special attention needs to be spent on their job submission behavior and the causes of that behavior. In this work, we investigate the impact of dynamic user behavior on parallel computing performances and analyze the significance of feedback between system performance and future user behavior. Therefore, we present a user-based dynamic workload model for generative simulations, which we use to analyze the impact of dynamically changing think times on simulations. We run several such simulations with widely known scheduling techniques FCFS and EASY, providing first insights on the influence of our approach on scheduling performances. Additionally, we analyze the performances by means of different metrics allowing a discussion on user satisfying performance measures.