Forecasting Duration Intervals of Scientific Workflow Activities Based on Time-Series Patterns

In scientific workflow systems, time related functionalities such as workflow scheduling and temporal verification normally require effective forecasting of activity durations due to the dynamic nature of underlying resources such as Web or grid services. However, most existing strategies cannot handle well the problems of limited sample size and frequent turning points which are typical for the duration series of scientific workflow activities. To address such problems, we propose a novel pattern based time-series forecasting strategy which utilises a periodical sampling plan to build representative duration series, and then conducts time-series segmentation to discover the smallest pattern set and predicts the activity duration intervals with pattern matching results. The simulation experiment demonstrates the excellent performance of our segmentation algorithm and further shows the effectiveness of our strategy in the prediction of activity duration intervals, especially the ability of handling turning points.