Hidden Markov Model Based Spot Price Prediction for Cloud Computing

Cloud computing accelerates the processing of big data applications by providing elastic virtual machines. Spot instances are virtual machines sold by auctions which offer lower prices (called Spot prices) than the same virtual machines with fixed prices. Spot prices change stochastically according to the real-time user demands and supplies. Accurate Spot price prediction helps users choose appropriate resources and set optimal bids to reduce resource rental cost. However, great fluctuations of recent Spot prices decrease the prediction accuracy of existing autoregression based prediction methods. Therefore, in this paper, a Hidden Markov Model and Expectation based prediction method (HMM-E) is proposed which uses Markov processes to describe the different demand markets of Cloud computing (hidden states) and the fluctuation degrees of Spot prices (observations) separately. Experimental results demonstrate that HMM-E predicts Spot prices more accurately than the existing autoregression based forecasting methods when the forecast period is shorter than about five hours.