Predikce spotřeby elektrické energie výpočetního clusteru

Abstract

This thesis focuses on predicting the runtime and energy consumption of computational jobs on the Karolina HPC cluster operated by IT4Innovations. The objective is to develop neural network models capable of forecasting job walltime and average energy usage based on historical data. A multi-layer perceptron (MLP) architecture was implemented using the Keras framework. The models were trained separately for different node types (CN and ACN) and user groups. A custom loss function was designed to penalize underestimated runtimes, addressing common inefficiencies in job scheduling. The results show that specialized models significantly outperform general-purpose models in prediction accuracy. The thesis also outlines potential future enhancements, including the use of natural language processing for analyzing job scripts.