Communication-aware Job Scheduling using SLURM

Job schedulers play an important role in selecting optimal resources for the submitted jobs. However, most of the current job schedulers do not consider job-specific characteristics such as communication patterns during resource allocation. This often leads to sub-optimal node allocations. We propose three node allocation algorithms that consider the job’s communication behavior to improve the performance of communication-intensive jobs. We develop our algorithms for tree-based network topologies. The proposed algorithms aim at minimizing network contention by allocating nodes on the least contended switches. We also show that allocating nodes in powers of two leads to a decrease in inter-switch communication for MPI communications, which further improves performance. We implement and evaluate our algorithms using SLURM, a widely-used and well-known job scheduler. We show that the proposed algorithms can reduce the execution times of communication-intensive jobs by 9% (326 hours) on average. The average wait time of jobs is reduced by 31% across three supercomputer job logs.