Federated Learning based Workload Prediction in Cloud Computing

Abstract

Predicting CPU demand is a major challenge in cloud computing due to the volatile nature of CPU utilization. Moreover, gathering CPU utilization data from multiple virtual machines to develop a prediction method raises concerns about data privacy, transmission costs, and system scalability. To address these challenges, this paper introduces FL-LSTM, a novel workload prediction technique that combines long short-term memory networks (LSTM) with federated learning (FL). In the FL-LSTM approach, each client uses LSTM along with its local CPU utilization data to create a local model. These local models are then aggregated to form a global model using the standard Federated Averaging (FedAvg) algorithm. We conducted a thorough evaluation of FL-LSTM using eight clusters of Google cluster traces. Our results demonstrate that FedAvg outperforms alternative FL strategies, while FL-LSTM matches or exceeds the performance of other state-of-the-art methods for cloud workload prediction. Notably, FL-LSTM achieved a Mean Squared Error of 0.00438, representing improvements of 74.7% and 9.4% compared to ARIMA and HBND, respectively. These findings highlight the potential of FL-LSTM as an effective solution for predicting CPU demand in cloud computing environments.