Virtual Machine Failure Prediction Using Machine Learning

Abstract

"In recent years, cloud computing has transformed how organizations manage their information systems. With this transformation, they have shifted to deploy their application on cloud platforms. These cloud platforms consist of many servers and Virtual Machines (VMs). VMs are complex systems. As with any other complex system, VMs are also prone to failures. VM failures can lead to service disruptions and data losses. Which negatively impacts the platform and its end users. These failures may occur due to a wide range of reasons. Various indicators signal these failures in the system. Even though VM failures are frequent, their potential consequences can be mitigated by predicting the occurrence of a failure in advance and taking proactive fault tolerance approaches such as VM migration. Since manually monitoring these signals is labour-intensive and impractical, in this project, the author proposes an ML-based failure prediction system that predicts the failure probability of VMs using Physical Machine (PM) log data and resource usage measurements. The proposed approach uses a combination of a Neural Network based autoencoder model and an LSTM model to detect anomalies and predict failures. The proposed system was tested using data collected from a PM, and it was able to detect Hard Disk Drive failures and Out of Memory failures with an accuracy of 97%, verifying the viability of the proposed methodology. "