An Auto-Scaling Approach using Load Prediction for IoT-based Cloud Application

Abstract

Cloud Computing enables a dynamic platform for deploying numerous internet applications by offering services such as software, and infrastructure. The Internet of Things (IoT) offers a variety of potential opportunities and applications, including smart grids, smart cities, intelligent transportation systems, e-health, and more, to fulfill various Cloud services. Thus, to address the current and future needs of end users, IoT and cloud computing are merged for the overall growth and organization of such applications. In addition to the advantages, complex cloud scenarios provide several challenges for IoT applications, such as scalability, reliability, heterogeneity, security, and privacy. One of the emerging issues for different IoT-based Cloud applications is auto-scaling. These days, many applications benefit from the auto-scaling capability, which allows them to scale up and down resources automatically. For efficient and autonomic scaling, predicting the future load on the host is recommended. This is because the load fluctuations may occur due to the dynamic resource usage among Cloud tenants. The over-utilized or underutilized status of the host can be monitored based on the predicted resource utilization, and the migration process can be performed to maintain the load on that host. In this research work, an auto-scaling technique has been proposed using load prediction for IoT-based Cloud application. First, an extensive literature survey of existing Cloud-based IoT applications has been done to achieve the objectives. Furthermore, state-of-the-art auto-scaling techniques have been surveyed, and necessary Quality of Service (QoS) parameters have been keyed out. From the literature, it can be inferred that load prediction-based auto-scaling is a challenging issue that must be handled intelligently. For dynamic load prediction, an Ensemble Time-Series Approach for Load Prediction (ETSA-LP) has been proposed which integrates five time-series analysis techniques (ARIMA,ANN, SVM, LSTM & ES) for predicting CPU and memory utilization. To evaluate the efficiency of the proposed approach, a series of experiments on Google and Planet- Lab traces have been conducted in a real Cloud environment. The proposed ensemble approach gives the best performance over the existing models by showing remarkable accuracy improvement and reducing the error rate and execution time. Further, an efficient auto-scaling approach for predicting host load through Virtual Machine (VM) migration has been proposed. Different algorithms have been devised to detect over-utilized and under-utilized hosts based on the predicted resource utilization. Also, a VM migration algorithm has been deployed that helps to choose the appropriate VM to be migrated. The designed approach has been validated by experimentation on a real-time Google cluster dataset. The proposed technique significantly improves average CPU utilization and reduces over-utilization and under-utilization. It also minimizes response time, SLA violations, and the slighter number of migrations and scaling overhead. For testing and validating the proposed approach, actual cloud platforms, GCP (Google Cloud Platform) and AWS (AmazonWeb Services) have been used. In this research work, heterogeneous VMs (Virtual Machines) are created for parallel execution and validating the performance of the proposed approach. The average load variations and status of VMs at different time intervals have been monitored. A case study based on a fog platform for IoT applications has also been considered for further enhancement. A Fog-Enabled Auto-Scaling Technique (FEAST) has been proposed to predict resource usage concerning CPU, memory, disk, and network utilization using the deep learning model. Several experiments have been conducted to evaluate performance and validate several indicators related to resource usage, scaling overhead, delay, and SLA violations, etc.