

























This paper describes a new algorithm called PAT, for Parallel Aggregated Trees, and which can be used to implement all-gather and reduce-scatter operations. This algorithm works on any number of ranks, has a logarithmic number of network transfers for small size operations, minimizes long-distance communication, and requires a logarithmic amount of internal buffers, independently from the total operation size. It is aimed at improving the performance of the NCCL library in cases where the ring algorithm would be inefficient, as its linear latency would show poor performance for small sizes and/or at scale.
此内容由惯性聚合(RSS阅读器)自动聚合整理,仅供阅读参考。 原文来自 — 版权归原作者所有。