Game theoretic analysis of the slurm scheduler model

In the context of High Performance Computing, scheduling is a necessary tool to ensure that there exists acceptable quality of service for the many users of the processing power available. The scheduling process can vary from a simple First Comes First Served model to a wide variety of more complex implementations that tend to satisfy specific requirements from each group of users. Slurm is an open source, faulttolerant, and highly scalable cluster management system for large and small Linux clusters [1]. MareNostrum 4, a High Performance Computer, implements it to manage the execution of jobs send to it by a variety of users [2]. Previous work has been done from an algorithmic approach that attempts at directly reduce queuing times among other costs [3][4]. We consider that there is utility at looking at the problem also from a Game Theoretic perspective to define clearly the mechanics involved in the system, and also those that define the influx of tasks that the scheduler manages. We model the Slurm scheduling mechanism using Game Theoretic concepts, tools, and reasonable simplifications in an attempt to formally characterize and study it. We identify variables that play a significant role in the scheduling process and also experiment with changes in the model that could make users behave in a way that would improve overall quality of service. We recognize that the complexity of the models might derive in difficulty to theoretically analyze them, so we make use of usage data derived from real usage from BSC-CNS users to measure performance. The real usage data is extracted from Autosubmit [5], a workflow manager developed at the Earth Science Department at BSC-CNS. This is a convenient choice, given that we also attempt to measure the influence of an external agent (e.g. a workflow manager) could have in the overall quality of service if it imposes restrictions, and the nature of these restrictions