English(EN) Tuning Collective Patterns to Alleviate Congestion in Shared AI Clusters

新的REACT系统在应用层缓解AI集群拥塞

作者 PulseAugur 编辑部 · [1 个来源] · 2026-09-07 04:00

研究人员开发了一个名为REACT的系统，该系统解决了分布式训练期间共享AI集群中的拥塞问题。REACT在应用层运行，使用流统计信息实时检测网络拥塞，并动态调整通信集体模式。这种方法不需要特殊的网络基础设施支持，并且可以由单个用户部署。REACT被原型化为NCCL之上的一个Shim层，在共享学术GPU集群的拥塞情况下，通信性能提高了13%-38%，模拟显示在各种场景下可提高高达75%。 AI

影响提高了共享AI集群中的通信性能，可能加速分布式训练。

排序理由关于AI基础设施新系统的学术论文。[lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv cs.LG 阅读 →

NCCL
REACT

AI 生成摘要 · Google Gemini · 来自 1 个来源。我们如何撰写摘要 →

本文如何被排名

Signal score: 32 / 100; Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket: Tool; 关于AI基础设施新系统的学术论文。[lever_c_demoted from research: ic=1 ai=1.0]
Source corroboration: Single-source cluster; Only one publisher covered this so far. Single-source stories can still rank when the publisher is high-authority, but they lack cross-source corroboration.
Topics: infra, paper; Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance: High; Clearly on-topic for AI-industry coverage.
Story freshness: Breaking (< 6h); Fresh story with cross-source coverage still developing. Ranking may shift as more sources report.

完整方法见我们的编辑标准。

报道来源 [1]

arXiv cs.LG TIER_1 English(EN) · Eashan Gupta, Yongzhou Chen, Apoorve Mohan, Pavlos Maniotis, Abdullah Kayi, Radhika Mittal · 2026-09-07 04:00

调整集体模式以缓解共享AI集群中的拥塞

arXiv:2609.04417v1 Announce Type: cross Abstract: Distributed AI training involves recurring rounds of data exchange between multiple pairs of GPU nodes. Slowdown in even one flow due to congestion can cause the entire communication round to slowdown. Current approaches for evadi…

本文如何被排名

报道来源 [1]

调整集体模式以缓解共享AI集群中的拥塞

相关实体

相关话题