International Journal of Computing and Mathematics

The Log analysis of High-Performance Computing (HPC) is an active research field. The task is how to extract the useful data from the HPC log file as the information from the analysis can be used as new knowledge to re- configure the HPC system to improve its efficiency. The traditional way of analyzing HPC log is considered inefficient in the sense that it takes time and requires the system administrator's specific knowledge and skills. In this research, we study empirically the application of techniques of machine learning to perform a task of analysis of the HPC log. To analyze and predict the work status on the HPC scheme, we apply machine learning methods that are distinct in their learning systems including C5.0, Support Vector Machine (SVM), and Artificial Neuron Network (ANN).We are also proposing a new technique called "Grouping and Combining." Grouping means reducing the target variable class labels. Doing so will reduce the time it takes to analyze. Then the target variable class labels are combined with another variable so that the interpretability efficiency can be increased. The dataset used in our experiment is the real-world data from the National Electronics and Computer Technology Center's HPC system, or NECTEC, Thailand. The C5.0 model has the highest predictive accuracy at 88.74%, according to the experimental results. The ANN model, by contrast, shows the best robustness. Furthermore, the experimental results show that the proposed grouping & combining technique can be used efficiently to handle the classification of multi-label as it helps to increase the accuracy, consume less time, and improve the interpretability of the learned model.