ML-Based System Failure Prediction Using Resource Utilization

Abstract

Digital transactions are growing exponentially on a regular basis without any sign of interruption. Unfortunately, most online service providers were not able to operate 24/7 every day due to the limitation of system components. Online service providers can be highly benefited if an accurate system failure prediction can be obtained. The adverse effect of computer failure might be mitigated if a proper prediction could be made beforehand. In this paper, we propose a simple model training approach to detect the failure that might arise in a system by parsing the log files and conducting a probabilistic analysis of future performance values in advance. We utilize a Recurrent Neural Networks (RNN), namely, Long Short-Term Memory (LSTM) to provide the optimal solution for predicting system failure by reckoning the hardware performance utilization value and utilizing the prediction of log data system for representing the system benchmark. Consequently, the significant constituents that are considered for calculation are all utilization of CPU, MEM, DISK, and NET. Apart from utilization, another essential constituent that needs to examine is System Callout, which is a representative for displaying the alert signal to inform whenever the information system should ignore the incoming transactions to maintain several server systems.