Title:Agon: A Scalable Competitive Scheduler for Large Heterogeneous Systems

Abstract:This work proposes a competitive scheduling approach, designed to scale to large heterogeneous multicore systems. This scheduler overcomes the challenges of (1) the high computation overhead of near-optimal schedulers, and (2) the error introduced by inaccurate performance predictions. This paper presents Agon, a neural network-based classifier that selects from a range of schedulers, from simple to very accurate, and learns which scheduler provides the right balance of accuracy and overhead for each scheduling interval. Agon also employs a de-noising frontend allowing the individual schedulers to be tolerant towards noise in performance predictions, producing better overall schedules. By avoiding expensive scheduling overheads, Agon improves average system performance by 6\% on average, approaching the performance of an oracular scheduler (99.1% of oracle performance).