Presentation Information
[B-14-28]Proposal of a Burst-Weighted Learning Method for Server CPU Usage Prediction
〇Yuta Isono1, Tomoya Kosugi1, Takashi Nakanishi1, Katsuya Minami1 (1. NTT, Inc.)
Keywords:
Server resource prediction,Burst-Weighted Learning Method,Machine learning
Estimating future server resource demand in virtualized infrastructures is important for efficient capacity planning and resource management. However, CPU utilization often includes sudden bursts and rapid fluctuations, making it difficult for conventional methods to accurately estimate peak resource usage. In this paper, we propose a weighted loss function for server resource estimation using a supervised variational autoencoder (SVAE), aiming to improve estimation accuracy during burst periods. The proposed method assigns larger weights to high-load samples during training, thereby encouraging the model to better capture burst behavior. Evaluation using real-world data demonstrates that the proposed method improves estimation performance under high-load conditions and reduces the estimation error for peak CPU utilization.