Task execution Failure Prediction Based on 1DCNN and Transformer in Cloud

• Binbin Huang,
• Weiwei Xu,
• Zewen Lin,
• Zixin Huang,
• Yuyu Yin,

Vol. 19, No. 2, February 28, 2025
10.3837/tiis.2025.02.004, Download Paper (Free):

• Cloud computing
• Deep Learning
• failure prediction
• Transformer

Abstract

With the growth of cloud data centers’ scale, task execution failure becomes an inevitable problem. Task execution failure in cloud can incur huge economic losses. To reduce the impact of task execution failure, we need to predict task execution failure accurately and take proactive measures to improve the service quality of users and avoid huge economic losses. The previous machine learning and deep learning-based approaches are inefficient and inaccurate when dealing with high-dimension features and dependent time series data. To further improve the accuracy of task execution failure prediction in cloud, this paper adopts a joint one-dimensional CNN and transformer approach. First, we use 1DCNN network to capture the local features of the input time series data. Second, global features of time series data are extracted through Transformer with the multi-head attention mechanism as the core. Residuals and layer normalization operations are added between layers to solve gradient vanishing and explosion problems. At last, we use mean pooling and fully connected layers to predict the final binary classification result. Extensive experiments have been conducted on the cloud data center cluster dataset published by Google, and the results show that 1DCNN-Transformer can extract potential dependencies of task more efficiently than previous work. It also has better performance metrics regarding prediction accuracy, precision, recall rate, F1 score, and ROC curve.

Statistics

Cite this article