Predicting running time of aerodynamic jobs in HPC system by combining supervised and unsupervised learning method

Abstract Improving resource utilization is an important goal of high-performance computing systems supercomputing centers. To meet this goal, the job scheduler often uses backfilling scheduling to fill short-time jobs into gaps at front queue. Backfilling needs obtain running time job. In past, usually given by users and far exceeded actual job, which leads inaccurate a waste resources. particular, when predicted lower than time, damage caused system’s resources becomes more serious. Therefore, prediction accuracy crucial system The use machine learning methods can make accurate predictions time. Aiming parallel application aerodynamics, we propose framework SU combining supervised unsupervised verify it on real historical data China Aerodynamics Research Development Center (CARDC). experimental results show that has high (80.46%) low underestimation rate (24.85%).