Evaluation of Machine Learning Techniques for Predicting Run Times of Scientific Workflow Jobs

Abstract

Predicting execution time of computational jobs helps improve resource management, reduce execution cost, and optimize energy consumption. In this paper, we evaluate machine learning techniques for the purpose of predicting execution times of scientific workflow jobs. Various aspects of applying these techniques are evaluated in terms of their impact on prediction performance. These include (1) Comparison of performance of different regressors; (2) using a single-stage prediction pipeline vs. two-stage one; (3) impact of categorization granularity in the first stage of the two-stage pipeline; (4) training one global model for all jobs vs. using separate models for individual job types. We also propose a novel prediction model based on symbolic regression and evaluate its performance. Interpretability of prediction models and usage of proper performance metrics are also discussed. Experimental evaluation has led to a number of interesting findings that provide valuable insight on how to apply machine learning techniques to prediction of execution time of computational jobs.