Introducing New Backfill-based Scheduler for SLURM Resource Manager

The work proposes a design for a new external scheduler for SLURM (Simple Linux Utility for Resource Management). Schedulers, included in SLURM by default are good enough for many sites, but big supercomputers serving many users meet limitations of standard SLURM schedulers. In this work we discover methods to break these limitations via implementing new portable SLURM scheduler. We address the problem of maximizing the number of users, whose requests are processed in each given moment of time, and decrease start time of user's first task. Our approach is based on standard backfill algorithm and includes additional features, such as simplification of SLURM priority system, replacing slow SQL-based accounting checks by faster ACL checks and upgrading cluster administration convenience.