Uncertainty-Aware Workload Prediction in Cloud Computing. (arXiv:2303.13525v1 [cs.DC])

Predicting future resource demand in Cloud Computing is essential for
managing Cloud data centres and guaranteeing customers a minimum Quality of
Service (QoS) level. Modelling the uncertainty of future demand improves the
quality of the prediction and reduces the waste due to overallocation. In this
paper, we propose univariate and bivariate Bayesian deep learning models to
predict the distribution of future resource demand and its uncertainty. We
design different training scenarios to train these models, where each procedure
is a different combination of pretraining and fine-tuning steps on multiple
datasets configurations. We also compare the bivariate model to its univariate
counterpart training with one or more datasets to investigate how different
components affect the accuracy of the prediction and impact the QoS. Finally,
we investigate whether our models have transfer learning capabilities.
Extensive experiments show that pretraining with multiple datasets boosts
performances while fine-tuning does not. Our models generalise well on related
but unseen time series, proving transfer learning capabilities. Runtime
performance analysis shows that the models are deployable in real-world
applications. For this study, we preprocessed twelve datasets from real-world
traces in a consistent and detailed way and made them available to facilitate
the research in this field.