面向新应用的数据中心网络DCN的自适应多粒度传输研究

Due to some new applications such as low latency and cluster computing for big data, DCN has shown some new features such as mixed DC, mixed elephant flow/mice flow/Coflow and coexisted link/cache/compute resource in the network structure, traffic characteristics, and network resources, respectively. The performance of traditional transmission methods in DCN is poor, due to lacking of deep coupling between them and these new features. In order to address some currently challenges, such as the contradiction between differentiated traffic characteristic and no-adjustable control scale, the contradiction between diversified requirements for performance and difficult allocating resource, the contradiction between diversified resource metric and difficult reusing of multiple resource, this project will research the following: 1. Building a hierarchical distributed intelligent SDN control platform based on NFV to flexibly supply controller resources. 2. Through the mapping between resource and delay, a method for reuse/distribution of multiple resource is proposed to achieve the conversion from requiring performance to requiring resource. 3. Based on deep learning, a flow type prediction mechanism with self-learning, which is pre-classification at edge + fine classification at center is proposed, to achieve fine-grained flow classification as soon as possible. 4. SDN+P4+INT can achieve the programmable and fine-grained control for resources and messages, and the deep reinforcement learning can taught itself the network, thus a dedicated route set that can be adaptive to diversified requirement for performance is established. 5. A hybrid scheme with distributed and centralized + packet and flow level is proposed to achieve multi-scale transmission with different policy, which can meet the diversified performance requirement from various types of flow at the same time. This project can greatly improve data transmission efficiency for DCN.

为应对低时延、大数据集群计算等新应用，DCN在网络结构、流量特征、网络资源分别出现混合DC、大象流/老鼠流/Coflow分化、链路/缓存/计算多资源并存等新特征，传统传输方法缺乏与新特征的深度耦合在DCN中性能低。为解决现有的流量特征分化与控制粒度调整难、差异化的性能要求与资源分配难、差异化的资源度量与多资源复用难等矛盾，本项目研究以下内容：1基于NFV构建层次化分布式智能SDN控制面，实现控制器资源的弹性供给。2通过资源-时延的映射提出多资源复用与配置方法，实现性能要求到资源需求的转换。3基于深度学习提出边缘预分类+中心精分类的自学习流类型预测机制，实现尽早地细粒度流分类。4基于SDN+P4+INT对资源和报文细粒度的可编程控制，通过深度强化学习对网络的自学习，建立可自适应于各类流性能要求的专有路由集。5提出分布式与集中式+报文与流的多粒度分策略传输机制，同时满足各类流差异化的性能要求。

为应对低时延、大数据集群计算等新应用，DCN在网络结构、流量特征、网络资源分别出现混合DC、大象流/老鼠流分化、链路/缓存/计算多资源并存等新特征，传统传输方法缺乏与新特征的深度耦合在DCN中性能低。为解决现有的流量特征分化与控制粒度调整难、差异化的性能要求与资源分配难、差异化的资源度量与多资源复用难等矛盾，本项目以准确地、及早地预测流类型为基础，研究实现自适应的多粒度传输机制——老鼠流、大象流等在各自专有链路集内，以报文/流、单个流/一组流等多个粒度实现分策略地传输。并且，利用深度强化学习对网络的自学习，根据各类流的性能要求为所属专有链路集自适应地分配网络资源，以实现同时满足各类流差异化的性能要求。.本项目经过四年潜心深入的研究，研究工作进展顺利，完成了预期的研究目标，取得了7项研究成果：基于深度强化学习的自适应多粒度传输方法、多资源复用/配置的方法、基于深度学习的多级流类型预测方法等。共发表学术论文17篇，其中，SCI期刊论文10篇(中科院1区论文4篇)，包括IEEE Network、IEEE Transactions on Network and Service Management、Journal of Network and Computer Applications、International Journal of Intelligent Systems等国际著名期刊；CCF 推荐学术会议论文4篇。IEEE Network是全球通信网络领域具有很高声誉和影响力的杂志之一，中科院1区，2021年影响因子达10.294。已获授权国家发明专利9件，其中有2件发明专利完成成果转化。项目组成员赵赛、欧阳海滨、杨钊等三人晋升副教授，欧阳海滨、赵赛分别于2019年、2020年获批国家自然科学基金青年项目。