Eunomia: a Performance-Variation-Aware Fair Job Scheduler with Placement Constraints for Heterogeneous Datacenters

Due to hardware upgrades and server consolidation, it is not uncommon to witness a few generations of servers deployed in the same datacenters. As a result, variants of fair job schedulers are proposed to enforce fairness for constrained jobs that have hardware or software constraints on task placement. However, the other important characteristics resulted from server heterogeneity, performance variation, is unfortunately overlooked by state-of-art fair job schedulers with placement constraints. In this paper, we propose Eunomia, a performance-variation-aware fair job scheduler, to address the unfairness issue due to performance variation in heterogenous clusters. Eunomia introduces a key metric, called progress share, which is defined as the ratio between the accumulated task progress given the current allocation and the accumulated task progress if the user can monopolize the cluster. Eunomia aims to equalize progress share of jobs as much as possible, so as to achieve the same slowdown of jobs from different users due to resource sharing and placement constraints, regardless of performance variation. Evaluation results show that Eunomia is able to deliver better share fairness compared with state-of-art schedulers without performance loss.