PREDICTING SCHEDULING DELAYS IN CLOUD COMPUTING SYSTEMS USING MACHINE LEARNING

Abstract

Properly allocating and scheduling resources in cloud computing systems helps achieve both performance and cost goals. Poor scheduling can cause significant disruptions, which affect throughput, utilization of resources, and customer satisfaction. This study introduces a machine learning framework to predict scheduling delays for cloud instances.To predict scheduling delays for cloud instances, a complete dataset includes all requests and three types of supervised learning models (Random Forest, XGBoost, and Logistic Regression) have been evaluated. The dataset underwent many pre-processing steps, such as the elimination of leaks and the development of new features as well as the establishment of a classification target to identify which instance types had high-delays or low-delays when scheduled.

Our results demonstrate that the ensemble-based models (Random Forest and XGBoost) performed better than the linear models because XGBoost correctly predicted scheduling arrangements 74.5% of the time, while also producing reasonable results in cases of class-imbalance. The combination of feature importance analysis and SHAP interpretation indicates that the total requested resource demand for CPUs, memory limit, and instance termination time are significant contributors to delays in scheduling.

As such, the approach proposed in this paper provides cloud service providers with the means to efficiently manage scheduling delays and thereby enhance the management of resources. Overall, this research illustrates that machine learning can accurately predict scheduling outcomes for cloud computing systems, thus contributing to the transition towards a more resource-efficient cloud computing environment.