Scholarlink

<p>High-Performance Computing (HPC) systems rely on efficient interconnection networks to provide fast and reliable communication between processing elements. The dragonfly network topology is a prominent solution for such environments due to its scalability and high performance. Path congestion in dragonfly networks degrades communication efficiency by intensifying buffer contention and routing conflicts among packets. In this paper, we address the critical issue of path congestion in dragonfly networks. We introduce and evaluate multiple algorithms designed to detect and mitigate path congestion, including Path Congestion Detection and Path Congestion Detection with Disjoint Paths. These algorithms are specifically designed for the extreme concurrency, high-radix routers, and low latencies of current and emerging supercomputers. We exploit precomputed disjoint paths and real-time congestion monitoring to reduce packet stalls under global synchronization and bursty traffic patterns in HPC workloads. Our simulations demonstrate that the proposed methods significantly outperform existing solutions in terms of latency, throughput, and packet delivery reliability.</p>

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

Path congestion detection and disjoint path for improving communication efficiency in dragonfly interconnection networks

  • 期刊论文

  • 发表日期：

    2025年12月3日

查看全文
  • Yaodong Wang,
  • Yamin Li

摘要

High-Performance Computing (HPC) systems rely on efficient interconnection networks to provide fast and reliable communication between processing elements. The dragonfly network topology is a prominent solution for such environments due to its scalability and high performance. Path congestion in dragonfly networks degrades communication efficiency by intensifying buffer contention and routing conflicts among packets. In this paper, we address the critical issue of path congestion in dragonfly networks. We introduce and evaluate multiple algorithms designed to detect and mitigate path congestion, including Path Congestion Detection and Path Congestion Detection with Disjoint Paths. These algorithms are specifically designed for the extreme concurrency, high-radix routers, and low latencies of current and emerging supercomputers. We exploit precomputed disjoint paths and real-time congestion monitoring to reduce packet stalls under global synchronization and bursty traffic patterns in HPC workloads. Our simulations demonstrate that the proposed methods significantly outperform existing solutions in terms of latency, throughput, and packet delivery reliability.