【Resource Management】论文整理_simple packing strategies

【Resource Management】

Resource Management with Deep Reinforcement Learning

提出DeepRM，Policy Gradient Reinforcement Learning，理想化资源分配，N种资源，任务需求给定。RL来决定分配方法。Baseline，对比启发式Shortest Job First，packing heuristic in [17]，Tetris⋆

[SIGCOMM '19] Learning scheduling algorithms for data processing clusters

研究的问题是，DAG任务依赖调度问题。决策，并行度和work运行哪个任务。

fair scheduling [8, 31],shortest-job-first,simple packing strategies [34],Many systems encode job stages and their depen- dencies as directed acyclic graphs (DAGs)[10, 19, 42, 80],Efficiently scheduling DAGs leads to hard algorithmic problems whose optimal solutions are intractable [36].coarse-grained fair sharing [8, 16, 31, 32], rigid priority levels [77], and manual specification of each job’s parallelism [68, §5].For example, in a recent paper, Agrawal et al. [5] showed that two simple DAG scheduling policies (shortest-job-first and latest- arrival-processor-sharing) have constant competitive ratio in a basic model with one task per job stage.Existing sched- ulers ignore this challenge: they enqueue tasks from a stage as soon as it becomes available, or run stages in an arbitrary order.Our method is based on graph convolutional neural networks [12, 23, 46] but customized for scheduling. Table 1 defines ou