基于DCQCN的拥塞控制算法研究

Remote Direct Memory Access (RDMA) is a technology used to solve server data processing latency in network transmission. It can effectively reduce latency without requiring the intervention of both user operating systems. Since DCQCN was proposed, it has gradually become a widely adopted congestion control solution for RDMA network. However, in data center networks that meet ultralow latency and high bandwidth requirements, DCQCN has insufficient cast processing capability and complex parameter adjustment issues when large-scale incast communication happens. DCQCN uses fixed period and steps for rate increase when probing for available bandwidth, which is prone to queuing, causing even packet loss in large-scale communications. Thus, this article proposes an improved algorithm DCQCN-Q based on the congestion queue at the switch. DCQCN-Q leverages in-network telemetry (INT) to obtain accurate switch queue information and feedbacks the congestion status of the switch to the sender. The sender dynamically adjusts the rate according to this feedback signal. The experimental results on the NS-3 simulation platform show that DCQCN-Q has strong ability of handling incast loads. When DCQCN-Q is implemented in the simulation under the workload ofWebSearch, the data latency is more than 13% lower than other mainstream algorithms.