Combining genetic algorithms and bayesian neural networks for resource usage prediction in multi-tenant container environments.

Traditional cloud architectures struggle to effectively allocate resources to container-based workloads due to fluctuating usage patterns and potential interference among multi-tenants. Conventional scheduling methods, which primarily rely on user-specified resource requests, often lead to over-provisioning and suboptimal resource utilization. Although efforts have been made to predict container resource usage and allocate resources more tightly than the full requests, such approaches typically fall short during sudden demand spikes, thus failing to meet Service Level Objectives (SLOs). In this article, we introduce a novel cloud resource prediction engine specifically designed to differentiate between online and batch jobs. Our engine prioritizes ensuring SLOs for online jobs where immediate responsiveness is crucial. Specifically, our approach employs a combination of genetic algorithms (GA) and Bayesian neural networks (BNN) to enhance the prediction accuracy of CPU and memory resources. Trained on real-world trace data, our model significantly outperforms traditional forecasting methods like ARIMA and exponential smoothing, especially in reducing the risk of under-prediction for online jobs. This not only ensures more efficient resource utilization but also improves adherence to SLOs without compromising performance. [ABSTRACT FROM AUTHOR]