Energy and time efficient federated learning

Hwang, Jing-Teng

Permalink

https://hdl.handle.net/2142/132539 Copy

Description

Title

Energy and time efficient federated learning

Author(s)

Hwang, Jing-Teng

Issue Date

2025-12-02

Director of Research (if dissertation) or Advisor (if thesis)

Shanbhag, Naresh R.

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Edge Device
• Federated Learning

Abstract

Over the past decade, the volume of data generated by edge devices has grown exponentially as the number of edge devices surges. Federated learning (FL) enables on-device training while preserving privacy, but edge devices typically operate under tight time and energy budgets, highlighting the need for time- and energy-efficient FL algorithms. While prior work focuses on either compute or communication costs, our hardware characterization shows that, although per-bit communication is far more expensive than per-bit computation, the relative importance of compute and communication costs depends on the communication period. Therefore, both costs must be considered before we find the optimal communication period. In this work, we formulate the optimal time and energy efficiency of FL systems into two separate optimization problems, characterized by four design variables: communication period C, GPU frequency f_GPU, number of clients K, and number of global rounds G. We solve both problems analytically and empirically using our simulation framework featuring a maximum of 10 Jetson clients with IID Cifar-10 dataset. Our empirical results show that time efficiency is optimal at the maximum available GPU frequency f^(max)_GPU, while energy efficiency is optimal at the highest GPU frequency with minimum supply voltage. Ablation studies reveal that optimal K for time efficiency increases with target accuracy; while optimal K for energy efficiency is always the smallest feasible K. The ablation also shows that optimal C for both time and energy efficiency increase with target accuracy. Based on the optimal hyperparameter setting, we find that optimal communication latency dominates only in low-accuracy regions (20%−29%); whereas optimal compute latency dominates in high-accuracy regions (30%−69%). On the other hand, optimal compute energy dominates across all accuracy regions. These results imply that compute costs are the primary concern for system efficiency under homogeneous FL settings.

Graduation Semester

2025-12

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/132539

Copyright and License Information

Copyright 2025 Jing-Teng Hwang

Owning Collections