Design of efficient congestion control method based on RDMA

Journal Information

J. Hefei Univ. Tech. (Nat. Sci.), Oct.2025, Vol.48 No.10: 1344-1351

DOI: 10.3969/j.issn.1003-5060.2025.10.007

Authors

WANG Fanghui $ ^{1} $, HUANG Zhengfeng $ ^{1} $, QIU Linya $ ^{1} $, GUO Erhui $ ^{1,2} $

(1. School of Microelectronics, Hefei University of Technology, Hefei 230601, China; 2. Wuxi Stars Micro System Technologies Co., Ltd., Wuxi 214000, China)

Abstract and Keywords

Abstract: This paper studies the network congestion control problem of remote direct memory access (RDMA) technology in data centers, and proposes a parameterized data center quantized congestion notification(DCQCN-p) algorithm to solve the problems of slow convergence speed and lack of hardware implementation schemes of the mainstream congestion control algorithm DCQCN. The DCQCN-p algorithm optimizes the velocity factors a and g of the congested flow to adjust the speed ratio $ R_{c} $, and reduces the frequency of speed reduction through circuit design. By establishing the algorithm model and building the NS-3 simulation platform, the performance of the DCQCN-p algorithm in terms of the speed adjustment of a single scheduled flow and the delay and throughput of multiple scheduled flows in the face of congestion is compared. The simulation results show that the data transmission rate of the DCQCN-p algorithm is increased by 50% compared to the DCQCN algorithm when a single flow is congested. In addition, the DCQCN-p algorithm achieves a minimum link rate of 13.28 Gbit/s, representing a 24% increase over DCQCN, 48% over TIMELY, and 23% over data. center transmission control protocol(DCTCP). The fairness of bandwidth allocation of DCQCN-p algorithm(65% variance) is significantly improved compared to TIMELY(216%) and DCTCP(191%).

Keywords: remote direct memory access(RDMA); parameterized data center quantized congestion notification(DCQCN-p) algorithm; circuit design; multi-stream efficient; network emulation

作者信息

王芳慧 $ ^{1} $，黄正峰 $ ^{1} $，邱麟雅 $ ^{1} $，郭二辉 $ ^{1,2} $

(1. 合肥工业大学微电子学院，安徽合肥 230601；2. 无锡众星微系统技术有限公司，江苏无锡 214000)

摘要和关键词

摘要: 文章研究并解决数据中心的远程内存直接读取(remote direct memory access, RDMA)技术的网络拥塞控制问题。针对主流拥塞控制算法数据中心量化拥塞通知(data center quantized congestion notification, DCQCN)的收敛速度慢和缺乏硬件实现方案的不足, 提出可参数硬件化的数据中心量化拥塞通知(parameterized DCQCN, DCQCN-p)算法, 该算法通过优化拥塞流的速度因子a、g调整速度比例 $ R_{c} $, 并通过电路设计减少降速的频次; 通过建立算法模型和搭建网络仿真NS-3平台, 对比DCQCN-p算法在面临拥塞时单个调度流速度调整的性能以及多个调度流并发情况下的时延和吞吐量。仿真结果表明: 在单个流面临拥塞时, DCQCN-p算法的数据传输速率比DCQCN算法的提高了50%; DCQCN-p算法在链路上最小速率为13.28 Gbit/s, 相较于DCQCN、TIMELY、数据中心传输控制协议(data center transmission control protocol, DCTCP)算法, 分别增长了24%、48%、23%; DCQCN-p算法(方差65%)的带宽分配公平性相较于TIMELY算法(方差216%)和DCTCP算法(方差191%)表现出显著的性能提升。

关键词: 远程内存直接读取(RDMA);可参数硬件化的数据中心量化拥塞通知(DCQCN-p)算法;电路设计;多流高效;网络仿真

Funding

国家自然科学基金资助项目（62274052；62374049）；安徽省重点研究与开发计划资助项目（202304a05020003）和安徽高校协同创新资助项目（GXXT-2023-011）

合肥工业大学学报(自然科学版)

Navigation