Scholarlink

Graph neural networks (GNNs) demonstrate potential in processing non-Euclidean data, but face communication bottlenecks when training on edge devices with graph sampling. We propose EDGaE, a distributed GNN training system addressing this challenge through two key designs: (1) Memory-aware graph partitioning that divides large graphs into edge-device-processable partitions; (2) Frequency-based feature caching that prioritizes high-access vertices. Evaluations show EDGaE achieves 2.54–25.57× speedup over Deep Graph Library (DGL) while maintaining scalability with device count and graph size.

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

EDGaE: Efficient Distributed Graph Neural Network Training System at the Edge

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Song Jin,
  • Xiantao Cai,
  • Jiawei Jiang

摘要

Graph neural networks (GNNs) demonstrate potential in processing non-Euclidean data, but face communication bottlenecks when training on edge devices with graph sampling. We propose EDGaE, a distributed GNN training system addressing this challenge through two key designs: (1) Memory-aware graph partitioning that divides large graphs into edge-device-processable partitions; (2) Frequency-based feature caching that prioritizes high-access vertices. Evaluations show EDGaE achieves 2.54–25.57× speedup over Deep Graph Library (DGL) while maintaining scalability with device count and graph size.