NCCL

NCCL is NVIDIA’s GPU collective-communication library and a recurring execution substrate for data-parallel, tensor-parallel, and expert-parallel systems in this corpus.

是什么

NCCL provides collectives such as AllReduce, AllGather, and ReduceScatter over GPU interconnects. In paper pages it is usually an implementation substrate rather than the contribution itself: a system either schedules around its collectives, diagnoses their failures, changes buffer/layout behavior, or compares a new transport path against the assumptions of the standard stack.

NCCL throughput and failure behavior depend on topology, GPU generation, driver/runtime version, message size, concurrent traffic, and collective algorithm selection. A result measured with NCCL therefore should not be generalized to an abstract network claim without its platform boundary.

关键观察 / 隐含假设

  • 观察:collective behavior is coupled to workload and topology. fabric-lib-MLSys26 and veScale-FSDP-MLSys26 treat buffer layout and communication scheduling as first-class performance constraints.
  • 观察:communication failures and stragglers are operational concerns, not only performance noise. Guard-MLSys26 and Greyhound-ATC25 study related reliability/diagnostic boundaries.
  • 假设:the standard collective API is sufficient for an optimization. Obscura-ATC25 and Mercury-SOSP25 show that deployment constraints can require additional transport, scheduling, or observability mechanisms.

演进时间线

  • 2025 ATC:Greyhound-ATC25 — investigates distributed execution behavior involving collective communication.
  • 2025 SOSP:Mercury-SOSP25 — places communication/runtime behavior in a system-management context.
  • 2026 MLSys:fabric-lib-MLSys26 — treats collective implementation and scheduling as part of performance engineering.

相关概念

相关论文