Scholarlink

The contemporary data center networks predominantly adopt INT-based congestion control algorithms owing to their commendable performance. These algorithms capitalize on switch-provided insights to swiftly adjust rates to target values within a single Round-Trip Time (RTT), reacting fast to network congestion while maintaining low latency. Nevertheless, in the prevalent use of Remote Direct Memory Access (RDMA) networks within data centers, the dynamic traffic patterns and the propensity for launching new flows at line rate often lead to observed disparities in bandwidth allocation among flows. This inequity persists even after reaching a steady state for most of INT-based algorithms. This phenomenon arises from the algorithms’ reluctance to make significant rate adjustments when link bandwidth is fully utilized and queues remain nearly empty. To address this issue, this study introduces a fairness identification module embedded within switches. This module employs statistical methods to categorize flows’ sending rates with bounded state. And then upon detecting unfairness, signals target rate to senders. This mechanism enables senders to compete equitably with other flows at the designated rate. Through our evaluation, the proposed solution showcases a substantial over 80% reduction in the time to convergence fairly compared to the state-of-the-art solutions with virtually no loss in overall link utilization.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing Equity: A Switch-Assisted Strategy for Improving Fairness in RDMA Networks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Quanwei Sun,
  • Xingbo Gao,
  • Zerui Tian,
  • Sen Liu,
  • Yang Xu,
  • H. Jonathan Chao

摘要

The contemporary data center networks predominantly adopt INT-based congestion control algorithms owing to their commendable performance. These algorithms capitalize on switch-provided insights to swiftly adjust rates to target values within a single Round-Trip Time (RTT), reacting fast to network congestion while maintaining low latency. Nevertheless, in the prevalent use of Remote Direct Memory Access (RDMA) networks within data centers, the dynamic traffic patterns and the propensity for launching new flows at line rate often lead to observed disparities in bandwidth allocation among flows. This inequity persists even after reaching a steady state for most of INT-based algorithms. This phenomenon arises from the algorithms’ reluctance to make significant rate adjustments when link bandwidth is fully utilized and queues remain nearly empty. To address this issue, this study introduces a fairness identification module embedded within switches. This module employs statistical methods to categorize flows’ sending rates with bounded state. And then upon detecting unfairness, signals target rate to senders. This mechanism enables senders to compete equitably with other flows at the designated rate. Through our evaluation, the proposed solution showcases a substantial over 80% reduction in the time to convergence fairly compared to the state-of-the-art solutions with virtually no loss in overall link utilization.