关键词: IT系统, 负荷预测, 时间序列分析, R语言应用, auto ARIMA, 中位数模型

Abstract: In order to find a simple and effective model to predict the workload of an IT system（run upon a JVM）, a press test environment upon an existing IT system is set up, simulating the workload increasing gradually, meanwhile sampling the JVM’s performance parameters data（CPU usage, memory usage, network bandwidth usage）, based on the time series data, eight conventional models are used to predict continuously the next ten second’s probable value, and compute out the mode parameters with R language, finally the one with the most robustness and prediction accuracy is found out. The eight models are: linear regression, linear regression with logarithm, two curve regression, Holt-Winters exponential smoothing, ARIMA, auto ARIMA in R language package（by Hyndman）, mean model, median model. The result shows that different suitable models exist in the three performance parameters predication: the CPU usage changing fast usually, the median model is the optimal one; the memory usage changing slowly usually, the auto ARIMA or the ARIMA model mentioned in the paper is the optimal two; the network bandwidth usage changing slowly with large amplitude, the Holt-Winters model is the optimal one. The conclusion can be applied to production as basis when choosing the optimal quasi-real time predict model.

Key words: IT system, workload forecast, time series, R language, Auto Regressive Integrated Moving Average（ARIMA）, median model