Job Runtime Prediction: A Two-Stage Framework Beyond PQR2 with Fallback and Enhanced Classification

In this work, we propose a solution to the problem of predicting job runtimes by improving the existing Predicting Query Runtime 2 (PQR2) approach. Our solution relies on two alternatives to PQR2 that include new mechanisms to solve identified PQR2 issues. First, we propose a fallback mechanism that uses a global regression model instead of a categorical regression model when confidence in first-phase categorization is below a given threshold. Second, we propose an extension mechanism to solve what we call the boundary problem of PQR2, where predictions of lower-than-average quality result when jobs are close to category borders. Finally, we propose two new optimization metrics specially adapted to regression problems with time intervals. We demonstrate using experimental results and comparison with PQR2 that our two alternatives can significantly improve the prediction accuracy. In addition, we show that our proposed extension mechanism can increase model performance up to 10%.