Learning from imbalance: Cross-server power prediction in large data centers via domain adaptation regression

Abstract

Machine learning (ML) models currently excel at predicting power consumption for servers in cloud data centers. The heterogeneity of server configurations violates the assumption of independent and identically distributed (i.i.d.) data, resulting in distribution shifts that pose significant challenges for cross-server power prediction. Additionally, the labels of power consumption data collected over a limited time show a natural imbalance, causing the power prediction performance to degrade when encountering missing labels. Therefore, learning meaningful knowledge from imbalanced power consumption data of real servers for cross-server power prediction remains challenging. To address this challenge, we consider imbalanced cross-server power prediction, formulated as a semi-supervised domain adaptation regression problem in scenarios where few labeled data points of target servers are available. Consequently, an imbalanced cross-server power prediction method, named ICSP, is proposed. To prevent learning biased knowledge from imbalanced data, unbalanced optimal transport is employed to align the joint probability distribution of the source and target servers. Moreover, by incorporating the few labels of target servers as a priori constraints, the performance of ICSP in coping with distribution shift is further improved. Extensive experiments on a real-world dataset demonstrate the superiority of ICSP over existing domain adaptation regression methods for imbalanced cross-server power prediction.