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).
翻译:这项工作提出了一种竞争性的排期办法,旨在推广到大型的多种多核心系统。这个排期器克服了以下挑战:(1) 近最佳排程员的高计算间接费用,(2) 性能预测不准确所带来的错误。本文介绍Agon,一个基于神经网络的分类器,从一系列的排期器中从简单到非常精确地进行选择,并学习了哪个排期器为每个排期间隔提供准确性和管理费用的适当平衡。Agon还使用一个去注的前端,使单排程员能够容忍性能预测中的噪音,产生更好的总体排期。Agon通过避免昂贵的排期间接费用,平均将平均系统性能提高6 ⁇,接近一个排程器的性能(占压轴性能的99.1%)。