Skip to content

并行机器学习中(ring—structure),efficient all—reduce,可否在 O( number_of(node) / Network_bandwidth ) 复杂度下实现? #6

Description

@David-Dingle

您的视频中,四个节点,六轮(每轮通讯四条链路同时利用)通讯,共24次通讯,复杂度是(number_of_parameter / wide——bandwidth)

有A B C D四台机器,分别在本地维护维护一个 w 矩阵,和一个同纬度的sum_local 矩阵,一个计数器 i 初始化为1。

sum_local = w // 根据本地 w 的拷贝初始化 sum_local
i = 1 // sum_local 已收集了一个矩阵
while( i != number_of_Node_in_ring )
{
throw w to next node
get new w from previous node // 每次接到上个节点传来的 w,下次循环时只将接收到的w再传递出去
++i
sum_local += w
}

这个方法可行么,共计四轮(每轮,四条链路同时利用)通讯,共16 次通讯(复杂度是 number_of_node / bandwidth)

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions