Predictive Scaling of Elastic Pod Instances for Modern Applications on Public Cloud through Long Short-Term Memory

In Cloud Computing (CC) environment, container virtualization through Docker engine is bringing the new digital transformation in the enterprise multi-tier application architecture in this modern era. Elastic pod containers attracts and helps the application developers in developing and executing the cloud-native modern applications with key benefits such as light weight, agility to launch, easy deployment through images, consuming less power, minimum cost, less carbon footprints with increased resource utilization and provisioning on public cloud data centers. In existing, reactive auto-scaling mechanism of pod resources is used to add or remove resources manually or rule based for handling static workloads from users and most of the times it may lead to over provision or under provisioning of instances that violates QOS and SLA. In this paper, predictive horizontal scaling of pods utilizing custom metrics with orchestration through Kuberenetes is proposed with LSTM (Long Short-Term Memory) based on Deep Learning (DL) technique. DL is data hungry for right prediction of replicas of a cluster that is needed in advance to run cyclic workloads and to handle the sudden spike of demand by using the cloud large dataset real time traces. Moreover, LSTM model is compared with GRU (Gated Recurrent Unit) and experimental results shows that the results of LSTM prediction has less absolute error rate on comparing with GRU to keep the resource provisioning accuracy better for running the modern workloads seamlessly on public cloud.