Host load prediction in a Google compute cloud with a Bayesian model

摘要: Prediction of host load in Cloud systems is critical for achieving service-level agreements. However, accurate prediction Clouds extremely challenging because it fluctuates drastically at small timescales. We design a method based on Bayes model to predict the mean over long-term time interval, as well consecutive future intervals. identify novel predictive features that capture expectation, predictability, trends and patterns load. also determine most effective combinations these prediction. evaluate our using detailed one-month trace Google data center with thousands machines. Experiments show achieves high accuracy squared error 0.0014. Moreover, improves by 5.6 -- 50% compared other state-of-the-art methods moving averages, auto-regression, and/or noise filters.