A Workload Characterization Methodology Using Supervised and Unsupervised Deep Learning

Abstract

A recuring decision problem in data centers optimally matches the computational requirements of an arbitrary workload with the capabilities of available servers. As the first step in this decision-making process, it is essential to obtain a set of meaningful and accurate workload classifiers. However, traditional approaches, such as principal component analysis (PCA), cannot capture time-varying workload characteristics. To fill this gap, this study proposes a deep-learning workload analysis tool (DLWAT) to capture complex workload dynamics. The DLWAT tool uses a hybrid deep learning model that includes a convolutional neural network (CNN) and recurrent neural networks (RNN) to study workloads from a supervised learning perspective. The DLWAT tool uses a hybrid deep learning model that includes an autoencoder (AE) and k-means clustering for unsupervised learning studies. Experiments showed that the DLWAT tool could precisely classify workloads for both labelled datasets (supervised use cases) and unlabeled datasets (unsupervised use cases). In the comparison studies, the DLWAT tool yielded good results.