decent-bench : A benchmarking framework for decentralized optimization

Sammanfattning

In distributed optimization, agents cooperate to solve optimization problems in peer-to-peer networks. Each agent’s local data is kept private, addressing critical privacy concerns. Despite real-world distributed systems facing asynchrony, bandwidth limitations, noise, and packet loss, benchmarking tools simulating communication constraints in distributed optimization remain limited. This thesis presents decent-bench, an open-source Python framework developed specifically for benchmarking distributed optimization algorithms under various communication constraints. It addresses limitations of existing benchmarking tools, providing a user-friendly, highly configurable, and metrics-oriented design that facilitates implementing new algorithms, defining benchmark problems, and evaluating algorithm performance. To demonstrate the framework’s capabilities, an empirical study was conducted, comparing four algorithms: distributed gradient descent (DGD), simple gradient tracking (GT1), simple exact diffusion (GT2), and distributed alternating direction method of multipliers (ADMM). The algorithms were evaluated on 12 benchmark problems consisting of linear and logistic regression under various communication constraints. The results reveal distinct performance patterns: ADMM demonstrated superior robustness to all communication constraints but incurred significant computational costs in problems without a closedform proximal solution, DGD exhibited consistent performance across all scenarios but never reached exact convergence, and the gradient tracking algorithms GT1 and GT2 achieved accurate results under ideal conditions but proved fragile to communication constraints. These findings provide practical guidance for algorithm selection based on deployment conditions, computational resources, and optimality requirements. In addition, the results’ consistency with previous research depends on the interpretation of the essential term robustness, revealing the need for more precise nomenclature in the literature.