Efficient Container Scheduling with Hybrid Deep Learning Model for Improved Service Reliability in Cloud Computing

Abstract

In a cloud computing environment, the container scheduling technique ensures reliability for containerized applications by selecting nodes that satisfy various resource requirements and then deploying containers. If the initial resources of a container are over-allocated, resources may be wasted, or other containers that are waiting in a scheduling queue may not be allocated. However, if resources are under-allocated, service disruptions may occur due to node overbooking, and service reliability cannot be ensured. Therefore, in this study, a forecasted resource-evaluating scheduler (FoRES) is proposed as a container scheduling technique that ensures resource efficiency and service reliability. FoRES predicts future CPU and memory usage by using a time-series decomposition-based hybrid forecasting (DeHyFo) model that combines multiple linear regressions with the LightTS model. FoRES then calculates the optimal scheduling decisions that minimize idle resources and node overload by applying an efficient resource utilization (SERU) scoring function to the predicted resource usage. Evaluating the performance of FoRES based on various scenarios improved resource efficiency and service reliability by up to 2.07 and 2.32 times, respectively, compared with existing scheduling techniques, even if the initial resources of the container were inefficiently allocated. Authors