【作者】 魏宁；

【导师】 周晓波；

【作者基本信息】 天津大学 ， 计算机科学与技术， 2019， 硕士

【摘要】 越来越多的应用程序(例如Web搜索,数据挖掘和推荐系统)部署到数据中心,同时它们越来越依赖于高性能数据中心网络来满足用户不断增长的服务体验质量要求。现有的流量调度方法都是建立在假设数据流的信息已知(如大小和截止时间)或拓扑为对称拓扑(如Fat-Tree等)的情况下。但在实际生产过程中,数据流的信息是很难在开始发送时获取到,即使是准确的预测也是十分困难的。在这种情况下,最小化数据中心网络中的平均流完成时间是一个巨大的挑战。现有的解决方案可以潜在地提供理想的性能,但是它们需要特殊的硬件支持或者对服务器和主机端的TCP/IP协议栈进行修改,而这在实践中是很难实现。本文为具有非对称拓扑挑战的数据中心网络提出了一种基于Flowlet的信息不可知情况下的流量调度机制。该方法的实现是轻量级但有效的,不需要对服务器和主机端的TCP/IP协议栈进行修改。主要思想是利用交换机中的多个优先级队列,在Flowlet级别进行调度,动态降低数据流的优先级。更具体地说,数据流最开始被赋予最高的优先级,然后根据其已发送的Flowlet的数量逐渐降低其优先级,以模拟最短的作业优先原则。接着本文将最小化平均流完成时间问题建模为一个非线性比率和问题,并设计了两种启发式方法来得出次优的降级阈值。实验结果表明,与最新的流量调度方法相比,在实际工作负载下,提出的方法可以降低平均流完成时间多达15.35%。接着本文考虑到静态阈值不匹配问题可能带来的性能损失,以及静态阈值的更新周期较长,在前文工作上基础上提出了基于深度强化学习的信息不可知流量调度方法,将最小化平均流完成时间问题建模为一个深度强化学习问题,改进了深度确定性策略梯度方法来训练该模型。实验结果表明,基于深度强化学习的方法在实际工作负载下可以将平均流完成时间降低多达10.4%。更多还原

【Abstract】 As more and more applications,like web search,data mining and recommendation systems,are deployed in data centers,they rely on high performance data center networks(DCNs)to meet users’ increasing quality of the experience(Qo E)requirements.Hence,minimizing the average flow completion time(FCT)has been one of the most important goals for DCNs.However,existing traffic scheduling methods assume either prior knowledge of flows(i.e.,sizes and deadlines)or symmetric topologies(i.e.,FatTree or Bcube).In practice,it is difficult to obtain the information of flows.In this case,it is a great challenge to minimize the average FCT in DCNs.Existing solution can potentially provide ideal performance,but they require non-trivial hardware modifications or modifications to TCP/IP stack of end hosts which are hard to implement in practice.In this paper,we propose a flowlet based information-agnostic traffic scheduling mechanism for DCNs with asymmetric topologies.The mechanism is light-weight yet effective,and doesn’t require modifications to TCP/IP stack of end hosts.The key idea of our method is leveraging multiple priority queues in switches to demote the priority of flows dynamically on the flowlet level.More specifically,the priority of a flow will be demoted according to the number of flowlets it has sent,which follows the shortest job first discipline.We formulate the average FCT minimization problem as a nonlinear Sum-of-Ratios problem and design two heuristic methods to derive the sub-optimal demotion thresholds.Experiment results show that our method can reduce the average FCT by up to 15.35% with a realistic workload,as compared to the state-of-the-art traffic scheduling methods.Then,considering the possible mismatch problem caused by the static threshold used in the above work.We propose a deep reinforcement learning based method on the basis of the previous article.We formulate the average FCT minimization problem as a deep reinforcement learning(DRL)problem and upgrade the DRL method named Deep Deterministic Policy Gradient(DDPG)to train this model.Experiment results show that DRL method can reduce the average FCT by up to 10.4% with a realistic workload.更多还原