A highly configurable and efficient simulator for job schedulers on supercomputers

摘要

The supercomputers maintained by the Jülich Supercomputing Centre (JSC) are run in batch mode and are shared among all active users. Jobs are submitted to the job scheduler, which determines for each job the time of execution. It manages running and waiting jobs, sorts them by a given priority and executes the jobs by fitting them into the current system state. For users it is often difficult to comprehend the job scheduler's algorithm. Thus, a prediction of the future system allocation of currently running and queued jobs is valuable. It helps users to plan job submissions and supports administrators by optimising the system load. This talk summarises the design and implementation of a configurable simulation for various job schedulers. The developed simulation program called JuFo focuses on the job schedulers Loadleveler and Moab as these are the most important batch systems for the supercomputers provided by the JSC. A major design goal is to keep the simulation independent from special job schedulers, which is achieved by the generic configuration of JuFo. Finally, a test framework is developed, which allows for evaluating the accuracy of the generated schedules.