Swarm: Playground for Large-scale Decentralized Learning Simulations

Abstract

Decentralized learning is an emerging field of research that opens doors for many novel pervasive computing applications. In decentralized learning, model training is offloaded to devices in the edge, and in some approaches, functions entirely without a central controller. Swarm is a tool for fast and large-scale simulations to test the performance of the practical implementations of decentralized learning algorithms that underlie many pervasive computing applications. In Swarm, developers can easily launch simulations for their algorithms by simply writing code that defines the behavior of a device when collaborating with others. The developer delegates to Swarm the emulation of the devices’ encounters, given a pervasive computing scenario. By decoupling the encounter emulation and the learning algorithm execution, Swarm makes the configuration of diverse application scenarios easy and their simulations repeatable. Moreover, developers can evaluate the scalability of an algorithm in diverse and large-scale application contexts as Swarm can automatically deploy and manage multiple worker nodes. Finally, the Swarm Dashboard provides a visualization of the simulation progress and the algorithm performance.