Non-invasive characterization of user behavior in high-performance computing systems: The case of the Santos Dumont supercomputer

User behavior characterization is a key step towards increasing system utilization and user satisfaction in HPC systems. Different variables can be used for characterizing user behavior, including job geometry, dependency between jobs, and think time. In this work, we propose a non-invasive characterization methodology centered around the distribution of different job types in distinct user groups. The methodology is defined as non-invasive because it relies solely on accounting information made available by resource managers, without resorting to the monitoring of the applications’ execution. We used the Santos Dumont supercomputer at the National Laboratory for Scientific Computing (LNCC) in Brazil as a case study for our methodology. We made a cohort of the accounting logs of Santos Dumont’s resource manager (Slurm) during the period from 01/01/2023 to 07/31/2023. We identified 6 different job types and a distinctive distribution of them for different projects—the hierarchical organization of the supercomputer users in groups—, emphasizing the use of different strategies to explore the available capability of the supercomputer, as well as different styles of under-utilization of its resources. We believe such a methodology can help in identifying opportunities for improving resource allocation and reducing queue waiting times in HPC systems like the Santos Dumont supercomputer.