Performance Monitoring And Resource Scaling For Applications In Cloud

Efficiently allocating resource while keeping the performance objectives of applications in cloud is a crucial problem for both cloud providers and application owners. The research of performance monitoring and resource scaling can help to solve this problem. Auto scaling schemes decide when to scale according to the performance indicator, and then dynamically allocate virtual machines to match with the fluctuant demand.It is challenging to decide right time and proper amount of resources to scale using existing auto scaling methods for cloud web applications. Fortunately, the challenge possibly can be overcome by using a performance model to mapping load/demand to resource amount and a hybrid auto scaling algorithm. This solution uses request tracing techniques to monitor performance metrics and model a Resource-Pressure model which can provide proper amount of resources if the workload is given. The core idea of the hybrid auto scaling algorithm is combining long-term workload prediction-based horizontal scaling and short-term threshold-based vertical scaling. It takes the advantages of the two auto scaling methods which can handle the common fluctuant and bursting workload. The hybrid auto scaling system is implemented basing on OpenStack cloud platform called HybridScaler. And it includes request log collecting and analyzing, auto scaling mechanism, horizontal and vertical resource allocating by using OpenStack Nova.The evaluation result shows that HybridScaler can achieve promotions in effectiveness of performance assurance and efficiency of resource utilization while comparing with other mainstream auto scaling schemes to handle bursting workload. HybridScaler decreases 16-39% average response time and 34-50% SLO(Service Level Objectives) violation rate than both static threshold-based scheme and workload prediction-based scheme. Meanwhile, it uses less instance-hours than static threshold-based scaling method and keeps CPU utilization almost between 60% and 70% which can avoid significant resource waste and SLO violation in the other methods.