Carbon Friendly Federated Learning Framework for Optimized Distributed Storage Systems
S. Hou, Pavana Prakash, Rolando P. Hong Enriquez, Ningfang Mi, Alex Veprinsky, Dejan S. Milojicic · ACM SIGMETRICS Performance Evaluation Review · 2025
The rapid growth of data and distributed storage systems to store them, are leading to an increased demand for e!- cient storage management system. Existing methods however, often rely on centralized machine learning approaches, which are not well-suited for heterogeneity of access patterns across distributed nodes, may compromise the data privacy and are not optimized for energy consumption and carbon emissions. In this paper, we propose an intelligent storage management strategy named FL-LSTM, based on Federated Learning (FL) with Long Short-Term Memory (LSTM) that e''ectively predicts cache data access patterns in order to make caching decisions and manage resources that enhance the system performance. Through comparative experiments, we demonstrate that our FL-LSTM framework not only predicts request access but also reduces communication energy overhead and carbon emissions. This study validates the potential of FL in optimizing distributed storage systems and provides guidance for further improvements in storage management strategies. By integrating FL with the time-series modeling capabilities of LSTM networks, our FL-LSTM framework can predict cache demands, enabling dynamic cache allocation that improves the system response time and resource utilization. This research o''ers theoretical insights into the prediction of storage patterns in largescale data access scenarios and also serves as a practical reference to improve the e!ciency and performance of future distributed storage systems.