EVENT DOSSIER
The researchers proposed a system called REACT, aimed at addressing network congestion issues in shared AI clusters. This system operates at the application layer, using existing traffic statistics to detect congestion in real time and dynamically modifying communication sets (such as aggregate nodes in AllReduce trees). It can be deployed unilaterally by users without relying on underlying network infrastructure. The prototype was implemented on NCCL and tested in shared academic GPU clusters. Results showed that algorithm bandwidth performance increased by 13% to 38% under network congestion; simulations across various congestion scenarios further indicated that its performance could reach up to 75%.
本文提出名为 REACT 的系统，可在共享 AI 集群中通过调整 GPU 节点间的数据交换模式来缓解网络拥塞。该系统运行于应用层，利用现有流量统计实时检测拥塞并动态修改通信集合（如 AllReduce 树中的聚合节点），无需底层网络基础设施支持即可由用户单方面部署。原型在 NCCL 上实现，并在共享学术 GPU 集群中测试，结果显示在网络拥塞下算法带宽性能提升 13%-38%；跨多种拥塞场景的模拟进一步表明性能最高可达 75%。