Abstract: Abstract: In recent years, RDMA has provided high-performance data transmission for data centers. However, the current software and hardware design and communication protocols of RDMA cannot be applied to cloud-edge communications in edge computing. The main problems include: the RDMA network card cannot cope with high latency, interference between the data center and cloud side flows, etc. This paper proposes a RDMA based edge computing cloud-edge data communication technology. By redesigning the cloud-edge communication of edge computing, the transmission is divided into three transmission parts within the data center, at both ends of the cloud-edge dedicated line, and in the cloud-edge dedicated line. According to the transmission characteristics of different parts, a credit based flow control at the access end of the dedicated line network and a high delay communication cache are constructed, and a stepped congestion processing is designed. In addition, a cloud-edge communication RDMA network accelerator card has been developed, and a collaborative accelerator for both ends of the cloud-edge network has been designed to ensure high real-time performance through hardware acceleration cards. The lab measurement and simulation experiments verify that the RDMA based edge computing cloud-edge data communication technology significantly improves the cloud-edge communication throughput by more than 80%, the greater the load, the more obvious the increase, and ensures the fairness between each flow.

Key words: RDMA, edge computing, flow control, congestion algorithm

摘要： 近年来，RDMA为数据中心提供了高性能数据传输, 但当前RDMA的软硬件设计、通信协议都无法适用于边缘计算的云边通信，主要问题包括：RDMA网卡无法应对高时延、数据中心内和云边流之间相互干扰等。本文提出了一种基于RDMA的边缘计算云边数据通信技术。通过对边缘计算的云边通信进行重新设计，将传输分为数据中心内、云边专线两端、云边专线中三个传输部分，根据不同部分的传输特点，构建基于信用的专线网络接入端流控，以及高时延通信缓存，并设计了阶梯式拥塞处理。此外，还研发了云边通信RDMA网络加速卡，并设计了云边两端网络协同加速器，通过硬件加速卡保障高实时性。经实验室实测与仿真实验验证，基于RDMA的边缘计算云边数据通信技术提高云边通信吞吐量80%以上，负载越大提升越明显，且保障了各个流之间的公平性。

关键词: 关键词: RDMA, 边缘计算, 流量控制, 拥塞算法