Enhancing Cloud Job Failure Prediction with A Novel Multilayer Voting-Based Framework
B Charishma,
U Siva Teja,
P Swathi,
Y Venkata Charu Vikram,
M Pavaneswar
Accurate prediction of job failures in modern cloud data centers is essential for ensuring system
reliability, optimizing resource utilization, and maintaining uninterrupted service delivery. Cloud jobs
frequently fail due to resource insufficiency, misallocation, scheduling anomalies, and system-level
faults, highlighting the need for intelligent predictive models. This paper presents a multilayer votingbased
predictive framework for cloud job failure detection and failure type classification using the
Google Cluster 2019 trace dataset. In the first layer, an ensemble of machine learning and deep learning
models Decision Tree, K-Nearest Neighbours, XG-Boost, AdaBoost, and Artificial Neural Networks
is employed. These models are optimized using Grid Search-based hyperparameter tuning, and their
probabilistic outputs are combined through a weighted voting mechanism to improve prediction accuracy.
The second layer applies a Random Forest classifier to identify specific failure types, including Lost,
Kill, Finish, Evict, and Fail. Additionally, a Cat Boost classifier is incorporated, demonstrating superior
performance due to its efficient gradient boosting and resistance to overfitting. Experimental results
show that Cat Boost achieves 100% accuracy, outperforming all other models and providing a reliable
solution for predicting cloud job failures.
