Jethro's Braindump


All-reduce is a architecture for the distributed machine learning training.

In a single iteration of all-reduce, gradients are computed by GPUs independently across different machines, and the gradients are aggregated by the all-reduce primitive.

The all-reduce architecture is bandwidth-optimal in the absence of CPUs, but with additional CPU and bandwidth, this optimality no longer holds.

How it Works

Here we describe the operation of Ring, the most popular all-reduce algorithm.

An all-reduce operation can be split into 2 phases. First, the all-scatter operation splits \(M\) bytes into \(n\) parts, and uses \(n\) rings with different starting and ending point to reduce the \(n\) parts$. Each node sends \(\frac{(n-1)M}{n}\) traffic. While each of the other \(n-1\) rings sends \(M/n\) bytes.

In the all-gather phase, each node broadcasts its reduced part to the other \(n-1\) nodes using a ring. Each node sends \(\frac{(n-1)M}{n}\) of traffic.