From Data to Decisions: ML-Based Job Failure Analysis for the ALICE experiment

Abstract

This paper researches the possibility of utilizing machine learning to locate factors contributing to a failing job at the ALICE grid, focusing on the grid site at the University of Bergen. A prototype system has been developed for data collection, management, analysis, and machine learning. The anal- ysis data originates from the ALICE monitoring system, MonaLISA, and its grid middleware JAliEn. A custom transformer model is utilized in the re- search, which addresses memory constraints in the project test environment by processing subsets of the complete input related to a job execution at a time.