Abstract: Aiming at the problem of underutilization of job locality and  low prediction accuracy in high-performance computing systems, we  proposed a job runtime prediction algorithm considering locality. The algorithm comprehensively utilized the global and local features of job log data, and improved prediction accuracy through a voting mechanism that combined with machine learning prediction and locality-based time-series prediction. Experimental results show that on actual scheduling log datasets such as 
 Unliu Gaia and PIK IPLEX, the JRPL algorithm outperforms or is not inferior to the machine learning algorithms  as the baseline in all  three  metrics: average absolute error, average prediction accuracy, and hit rate. This research result provides an improved prediction model for job scheduling in high-performance computing systems, which hepls to make more accurate execution time forecasting, improve system resource utilization, and reduce computational costs.

Key words: high-performance computing system,  , job runtime prediction, machine learning, localized, prediction accuracy