The elastic cloud applies autoscaling technology to allow users to automatically provision or deprovision resources on demands, attracting many application providers to migrate their applications to the cloud. However, autoscaling multi-service applications are still challenging due to the complex correlations among services. This paper presents HRA, an intelligent, holistic resource autoscaling framework for multi-service applications, utilizing model-based deep reinforcement learning (DRL), mitigating service-level agreements (SLA) violations while saving costs. HRA (i) leverages historical telemetry data and machine learning methods to build a simulated environment adaptively, modeling relations between resources, workloads and performance, (ii) exploits the environment model to drive up training efficiency of DRL agent, and (iii) uses the agent to automatically take actions to scale resources online based on simple low-level features from a monitor instead of elaborate high-level features that are representing the complex correlations and needing much sophisticated prior knowledge. Experiments (i) evaluated the fidelity of the proposed (simulated) environment modeling method, (ii) evaluated the reliability of the resource allocation policy from the simulation to reality, and (iii) compared related autoscaling methods. The evaluation results demonstrate that HRA realizes a more effective resource allocation policy under the limited number of time-consuming interactions and significantly decreases the 32-92% in SLA violation rate at a lower cost compared to other main methods.

Container-supported microservice technology is widely used in cloud applications. For elastic cloud, it's vital to maintain application response time within service-level agreements (SLA) by auto-scaling technology. For applications composed of multiple services (i.e. multi-service applications), due to complex topologies, there are many factors that reduce auto-scaling algorithm performance, such as correlations among services, untimely decision, oversupply, etc. To resolve this, we propose a holistic auto-scaling algorithm (HAB) based on balanced Jackson queuing network (JQN) to reduce SLA violations rapidly with less resource cost. With the holistic auto-scaling strategy, HAB scales all services quickly and accurately. Keeping the balanced state among services, HAB saves resource cost, reduces auto-scaling decision space and simplifies algorithm parameters. The experimental results demonstrate that HAB has an average decrease of 42.31% in SLA violation rate, an average decrease of 17.88% in resource cost and an average increase of 19.39% in stability, compared with other main methods.