RDMA容器云架构下面向应用感知的网络资源管理与优化机制研究

With the emergence and development of modern networking applications including artificial intelligence, big data processing and microservices, it has become an inevitable trend to facilitate applications’ demands of high elasticity and network performance based on RDMA-enabled container cloud. However, the characteristic of thin middleware in the architecture of RDMA-enabled container cloud has brought significant challenges to network resource management. To tackle this challenge, this project, addressing three characteristics of modern networking applications, namely, topology heterogeneity, dynamic nature and performance sensitivity, conducts the research on the application aware network resource management and optimization for RDMA-enabled container cloud from the aspects of container orchestration optimization, network control optimization and data transferring optimization, respectively. First, a smart container orchestration mechanism based on the end-to-end model is investigated. Second, a flexible dynamic resource driven bandwidth adjustment mechanism is proposed. Third, an adaptive parameter tuning mechanism is further researched to satisfy the diverse network performance preference of applications. Based on the above research, a network resource management prototype system will be implemented to verify the theoretical results and provide a general, fast and efficient platform for modern networking applications.

随着人工智能、大数据处理、微服务应用等一批新型网络应用不断演进，基于RDMA容器云平台支撑上层应用弹性扩展和高效互联的需求是必然趋势。然而，RDMA容器云架构中间层扁平化的特点给复杂应用场景下充分利用网络资源的性能带来挑战。本文从应用感知的网络资源管理角度出发，针对大规模分布式应用存在拓扑异构性、流量动态性、性能敏感性等特点，分别从容器编排优化、网络控制优化、报文传输优化三个方面入手，研究RDMA容器云架构下面向应用感知的网络资源管理和优化机制，实现细粒度、高效能的网络资源管理。具体包括：（1）研究基于端到端学习模型的智能容器编排机制；（2）研究应用动态资源需求驱动的容器带宽柔性调整机制；（3）研究面向应用异构性能偏好的RDMA自适应参数选择机制。在此基础上设计并开发面向应用感知的RDMA容器云网络资源管理系统，对相关成果进行验证并为新型网络应用提供通用、快速、高效的支撑平台。

本项目以支撑大规模分布式应用在RDMA容器云中的高效执行为目标，从应用感知的网络资源管理角度出发，针对上层应用三个粒度的特征带来的挑战，结合RDMA容器云架构三个平面的特点和优势，分别从容器编排优化、网络控制优化、报文传输优化三个方面入手，研究RDMA容器云架构下应用感知的网络资源管理和优化机制，实现细粒度、高效能的资源管理。通过项目实施，项目负责人发表论文8篇，包括CCF A/B类高水平论文6篇，培养硕士研究生3名，博士生1名，依托本项目支持申请获批国家自然基金面上项目1项，同时申请国家发明专利2项，获得国家发明专利授权2项，完成了预期目标。