Assessing Forecasting Model Performance for Distributed Stream Processing Systems

Load forecasting plays an important role in the load balancing of the distributed stream processing systems. So, the performance and cost of three models: weighted moving average (WMA), exponential smoothing (ES) and grey model (GM(1,1)) are empirically evaluated by running three typical test cases on the load traces of the distributed stream processing systems and their results are reviewed according to three metrics: mean absolute percentage errors (MAPE), root of mean square error (RMSE), processing cost. According to the metrics of MAPE and RMSE, GM(1,1) performs best while WMA and ES perform much better than GM(1,1) according to the processing cost. However, when the load fluctuates dramatically, the prediction precision of the above models is low.