Edgent: A Multi-Agent Framework for Simulation and Optimization in Edge Computing

Distributing requests in Edge Computing environments is a complex task, requiring the identification of the optimal Edge Server for each request. This thesis addresses this problem by maximizing the number of completed requests and minimizing resource waste, while maintaining strict latency constraints. To address this challenge, we propose Edgent, a decentralized solution for request distribution that utilizes a Multi-Agent System approach. In Edgent, autonomous, context aware sub-agents run on each server, collaborating to optimize a common objective function. This enables intelligent migration of requests, shutdowns of instances to prevent resource waste, and the powering off of unnecessary servers. Edgent’s validation is conducted using a realistic Multi-Agent Simulation, developed with the Mesa Framework in Python. The results demonstrate that the Edgent approach exhibits high adaptability and resilience, even when external factors intervene (f.e. servers failure or user movement). We also analyze a potential real-world application of the system to evaluate changes in Edge server technology (from LTE to 5G) and the impact of objective function modifications. These modifications can introduce complex behavior (such as a cooldown mechanism) simply by adding parameters to the objective function.