Optimal Number of Edge Devices in Distributed Learning Over Wireless Channels

We consider a distributed learning system, where a parameter server (PS) assigns data and computational tasks to edge devices to build a global model. Distributing data to multiple workers involves communication between PS and edge devices and entails a fundamental tradeoff between computation and communication. In this paper, we aim at characterizing the optimal number of edge devices required for guaranteeing convergence and for achieving a certain accuracy within a finite time horizon.