Forecasting Data Center Resource Usage: An Experimental Comparison with Time-series Methods

Capacity planning via resource monitoring in data centers (DCs) is a challenge. The computing resource usage pattern changes over the time which makes its management a complex process. Maintaining system’s performance (or service) goals together with infrastructure management cost in a limit, are very critical. Despite various control measures, still, resources suffer from under utilisation. To alleviate such problems, we need a proactive resource management framework which combines the prediction of the future workload together with the efficient resource allocation strategy. In this paper, we focus on the prediction problem and compare the forecasting efficacy of the well-known time series prediction models by performing an exhaustive experimental campaign on the real workload traces collected from Wikimedia Grid. In this work, first, we compare popular variants of autoregressive moving average (ARMA) models (such as integrated, fractional, seasonal) with two different approaches namely the exponential smoothing model (Holt-Winters) and spectral estimation method (singular spectrum analysis (SSA)). Next, to increase the efficiency of the forecasting capability of the ARMA variants, we add Kalman filter and also Wavelet decomposition. Finally, we report all the results together with detailed explanations and execution time. In experiments, we observe that Kalman filter based ARIMA and its seasonal variants outperform others. These hybrid models also achieve a very high quality of forecasting accuracy (mean absolute percentage error (MAPE) such as it is less than 0.03% for RAM, under 0.8% for CPU and less than 4% for Network), while the complete system knowledge is not known to us. We also found that for one-hour out-sample prediction input size of fifteen and thirty days are enough to get very good results.