Get Started
Home
Topics
Search
Library
Research questionHow can distributed-training communication collectives adapt to external congestion in shared cloud clusters without network control?A single congested flow can delay an entire distributed-training communication round. In shared clouds, a user's job may encounter traffic from other jobs or background workloads without controlling their schedules or the network.
AI
Machine Learning
Research Paper
Technology
Latest papersRecent research connected to this question, newest first.Tuning Collective Patterns to Alleviate Congestion in Shared AI ClustersApplies to distributed AI training across GPU nodes when congestion comes from traffic beyond the user's control. The evidence covers a prototype shim over NCCL, evaluation on a shared academic GPU cluster, and simulations across congestion scenarios; it does not establish broader deployment results.research paper · Sep 3, 2026
Related questions
How can shared GPUs schedule concurrent heterogeneous AI inference without combinatorial profiling as workloads change?How should partitioning, placement, and scheduling be coordinated to reduce bubbles in heterogeneous LLM training?How can multi-GPU serving meet chunk-latency targets for stateful, bursty video-generation sessions?How can enterprises consolidate heterogeneous internal workloads onto one self-hosted LLM under data-residency constraints?