AMPRO-HPCC: A Machine-Learning Tool for Predicting Resources on Slurm HPC Clusters.

Area of Science:

• High Performance Computing (HPC)
• Machine Learning (ML)
• Resource Management

Background:

• Accurate resource allocation in HPC systems is challenging due to application variety and system complexity.
• Overestimation of resources leads to wasted HPC capacity, reduced utilization, and longer job wait times.
• Existing methods lack automated, accurate prediction for memory and time requirements.

Purpose of the Study:

• To develop and evaluate a novel, fully-automated, open-source ML tool for predicting HPC job resource requirements.
• To assist users in accurately estimating memory and time needs, thereby optimizing job submission and resource utilization.
• To mitigate issues of resource overestimation and its negative impacts on HPC system performance.

Main Methods:

• Implementation of six ML discriminative models, including scikit-learn and LightGBM.
• Application of models to historical Simple Linux Utility for Resource Management (Slurm) sacct data from January 2019 to March 2021.
• Testing and validation using a dataset of approximately 17.6 million jobs from Kansas State University's Beocat HPC resources.

Main Results:

• Achieved high predictive accuracy: R² of 0.72 for memory prediction (LightGBM) and 0.74 for time prediction (Random Forest).
• Demonstrated significant reduction in average job waiting and turnaround times.
• Showcased increased HPC system utilization and decreased power consumption.

Conclusions:

• The developed ML tool offers an effective solution for accurate HPC resource prediction.
• The tool optimizes HPC resource allocation, leading to improved efficiency and reduced operational costs.
• This approach enhances the overall performance and sustainability of High Performance Computing environments.