Scholarlink

In this chapter, we present our distributed GNN training solution. Specifically, to mitigate the communication overhead induced by expensive data movement across workers, we propose Sancus, a staleness-aware, communication-avoiding, decentralized GNN system. By introducing a set of novel bounded embedding staleness metrics and adaptively skipping broadcasts, Sancus abstracts decentralized GNN computation as sequential matrix multiplication and leverages cached historical embeddings. Theoretically, we establish bounded approximation errors for both embeddings and gradients, together with convergence guarantees. Empirically, we evaluate Sancus with common GNN models under diverse system configurations on large-scale benchmark datasets. Compared to SOTA works, Sancus avoids up to 74% communication and achieves at least \(1.86\times \) higher throughput on average, without accuracy loss.

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

Distributed Graph Neural Network Training

  • 章节

  • 发表日期：

    2026年

查看全文
  • Jingzhi Fang,
  • Jingshu Peng

摘要

In this chapter, we present our distributed GNN training solution. Specifically, to mitigate the communication overhead induced by expensive data movement across workers, we propose Sancus, a staleness-aware, communication-avoiding, decentralized GNN system. By introducing a set of novel bounded embedding staleness metrics and adaptively skipping broadcasts, Sancus abstracts decentralized GNN computation as sequential matrix multiplication and leverages cached historical embeddings. Theoretically, we establish bounded approximation errors for both embeddings and gradients, together with convergence guarantees. Empirically, we evaluate Sancus with common GNN models under diverse system configurations on large-scale benchmark datasets. Compared to SOTA works, Sancus avoids up to 74% communication and achieves at least \(1.86\times \) higher throughput on average, without accuracy loss.