计算机科学
网络拥塞
计算机网络
控制(管理)
计算机安全
人工智能
网络数据包
标识
DOI:10.1109/infcomw.2019.8845230
摘要
Data center network speeds are advancing on a yearly basis. As of now, servers have network interfaces of more than 25 Gbps. However, high-speed interfaces cannot be sufficiently utilized because the current Transmission Control Protocol (TCP) has a lot of software overhead. Therefore, Remote Direct Memory Access (RDMA) technologies, which decrease Central Processing Unit (CPU) overhead and eliminate data copies, are being used in Ethernet. In Ethernet, RDMA is enabled by RDMA over Converged Ethernet v2 (RoCEv2), which uses Data Center Quantized Congestion Notification (DCQCN) as a congestion control method. On the other hand, RoCE enables Non-volatile Memory Express over Fabrics (NVMe-oF), which can access remote Non-volatile Memory (NVM) at high speeds comparable with those of local NVM. However, to achieve suitable remote access performance, network congestion needs to be alleviated. In this paper, we show an example of how remote access performance is decreased by network congestion. We also find that the existing microburst-aware congestion control method for TCP does not work for RDMA. We propose Microburst Aware Congestion Control for Storage Traffic (MARCS) as a congestion control method for RDMA. In MARCS, a flow uses a dedicated transmission queue when the flow is at the beginning stage. Furthermore, by controlling the bandwidth of the queue, the performance of other continuous flows is kept high. MARCS also utilizes a mechanism to prevent packet reordering that heavily affects the performance of RDMA. Using simulation and real devices, we show that MARCS dramatically improves the throughput when congestion occurs.
科研通智能强力驱动
Strongly Powered by AbleSci AI