面向BCube拓扑的All-Reduce在网计算加速算法OA

As the communication performance requirements of large-scale parallel computing and deep learning tasks continue to grow,traditional host-based collective communication algorithms have gradually become a bottleneck in system performance.In-network computing,which offloads All-Reduce operations to network devices for execution,has emerged as a crucial approach for overcoming communication bottlenecks.However,most existing mainstream in-network computing approaches rely heavily on switch-centric topologies(e.g.,FatTree)and are difficult to adapt to server-centric topolo-gies(e.g.,BCube).This paper proposes BCube-INC,an in-network computing acceleration algorithm for All-Reduce targeting the server-centric BCube topology.Building on the hierarchical structure of BCube,the algorithm constructs multiple static reduction trees and applies a process-level path schedu-ling strategy to isolate their physical paths,enabling the reduction and broad cast processes to run con-currently.Performance evaluations and comparisons on the BCube(3,2)topology show that the laten-cies of Ring All-Reduce,BML,and the proposed algorithm are 104.85 μs,18.12 μs,and 16.5 μs,respec-tively,with the proposed algorithm achieving a maximum speedup of 1.93 over BML.Compared with the BML algorithm,the proposed algorithm achieves nearly 2 times bandwidth improvement.