面向网络功能虚拟化的资源调度与效能优化研究

By softwarizing traditional dedicated hardware based functions to virtualized network function (VNF) instance that can run on standard commodity servers, network function virtualization (NFV) technology promises increased efficiency, flexibility and scalability. To network service providers, two primary concerns are on the service provision cost and quality-of-service (QoS). Both are highly influenced by the VNF instance placement and resource management. With NFV, multiple instances for the same function may coexist in the substrate network. This raises an optimization dimension as the number of instances that shall be placed shall be also considered. In this project, we first study a cost efficient VNF instance placement, resource allocation and network flow balancing strategy for guaranteed QoS. After placing the instances, we shall further consider how to dynamically schedule the packet processing instances for each packet. With the consideration of coexistence of multiple instances for one network function, after processing a packet on one instance, there are multiple next-hop processing instance candidates. In addition, a network flow shall go through a chain of instances according to the network service chain definition. The scheduling on one instance not only directly influences the processing latency on the next-hop instance but also all other instances thereafter. The process can be described using a multiple-hop queuing network model. We will investigate how to systematically and dynamically schedule the packets and allocate resources accordingly so as to minimize the overall service latency. During network operation, it is inevitable that many dynamic events (e.g., new QoS requirement, node or link failure, traffic load increasing or decreasing, etc.) may happen. In order to guarantee the predefined QoS, we shall reschedule the instance placement and resource allocation accordingly in a cost-efficient manner. To this end, we will study and propose online rescheduling algorithms to deal with these events, with a special emphasis on how to re-explore existing placement and allocation. After adjusting the instance placement, some flows shall also be migrated accordingly from their original instance to the new one. If these flows are with states, we shall further consider the state migration and synchronization. A new instance can process a stateful flow only after it obtains the corresponding flow states from the original instance. During this process, these flows shall be temporarily cached. It is significant to study how to quickly migrate flow by considering destination instance selection and cache node decision. Besides, flows sharing a group state must synchronize the state among their instances and this incurs synchronization cost. As a result, the flow migration decision may also influence the synchronization cost. We will propose a multi-objective optimization model to address the tradeoff between the migration latency and the synchronization cost. By this project, we are expected to promote the development of NFV and build a solid foundation and technical support for practical NFV management.

针对传统网络中资源部署低效、管理复杂的问题，网络功能虚拟化技术应运而生。通过将网络功能和硬件平台解耦，网络功能虚拟化技术提升了网络弹性。如何权衡资源分配、任务调度、服务质量和服务成本之间的关系，实现高效能的网络管理，是网络功能虚拟化应用中一个亟待解决的重要问题。本项目拟从理论、模型以及算法层面对该问题进行研究。首先研究如何通过虚拟网络功能实例部署、资源分配以及流负载均衡，在保证服务质量的同时优化服务成本；在此基础上，进一步提出以最小化网络服务时延为目标的数据包实时调度策略；接着设计在线实例部署与资源分配算法，以应对网络中的各类动态事件；最后分析状态复制与同步问题，以最小化流迁移时延与同步成本为目标，研究带状态的网络流迁移策略。通过本项目的研究，可望在理论上揭示资源调度与服务质量的关联性，探索高效能网络功能虚拟化的资源分配与任务调度的优化模型，在工程上为网络功能虚拟化的应用提供技术支撑。

申请人的主要的研究方向是网络功能虚拟化技术，该技术通过解耦网络服务和物理硬件，使得硬件功能以虚拟机的形式按需部署在通用服务器上。当网络功能脱离了硬件束缚，那么如何管理网络功能就成为了优化服务效能的关键问题和核心挑战。申请人针对现在网络功能虚拟化研究工作中存在的非联合优化、粗粒度调度、定规则控制的现状造成的低效用、多拥塞、弱智能问题展开了讨论，从分析框架、调度策略和控制算法层面展开研究，主要包括以下三方面，（1）多语义跨状态的一体化分析框架：我们首先分析了不同语义和状态需求的处理过程，提出一个服务逻辑的扩展结构，把不同的处理语义和状态同化成一体化的分析框架。基于此框架上，分析不同的服务质量需求，定制网络功能的数量和部署和网络流的分配方案，降低了27%的网络服务开销。（2）网络功能和网络流制定动态调度策略：申请人首先以服务链上的各个网络功能为元单位，建立了新的多跳动态队列的网络模型，分析各个网络功能队列对服务质量的影响，提出基于CPU+GPU的细粒度网络包分发策略。其次，申请人提出了网络流量突发的节点和时机的判定机制，并实现了“单服务多模式”的实际系统，使得“批处理”和“流处理”模式能够共存并按需切换，针对流量突发进行优化，提升5倍的处理速度。（3）智能化的网络功能编排与流量控制算法：长期服务优化是一个马尔科夫决策过程，申请人借鉴了强化学习的理念，设计了最大化服务长期效用的控制决策。在此基础上，根据问题特性和决策关联，设计了新的动作探索方法和激活函数，用数学模型来辅助动作探索和经验回放过程，实现了训练速度的23倍提升。本项目研究的成果在IEEE JSAC、IWQoS、ICDCS、SECON等国际知名学术期刊和会议上发表和录用了论文13篇，其中CCF-A类论文2篇，CCF-B类论文3篇，SCI论文8篇。在项目支撑下申请了专利1项，参加国际会议和学术交流6次，和国内外同行专家进行交流合作。同时申请人基于本项目培养了6名研究生，其中4名已毕业。