Research And Implementation Of Algorithm For Autoscaling Overall Cloud Services

With the widespread adoption of various cloud computing platforms,distributed services running on it is getting more and more attention,and has becoming the dominated way for large and medium sized Internet Co and cloud service providers to build their systems.In order to meet the needs of load balancing,resource optimization and service quality assurance,it is imperative for the cloud platform management system to achieve automatic scalability.Scalability is a comprehensive consideration of many factors,such as high performance,low cost and maintainability.Although scalability is an important feature of cloud computing,efficient management and utilization of various resources are still one of the challenges of cloud platform management systems.Traditional reactive scaling mechanisms scale individual services based on performance counters and metrics of certain resources.The de-centralized SOA micro-service can invoke each other's API and its calling chain is complex.So that the scaling of a single service often results in the latency for overall scaling of the entire system,and serious overall performance jitter of the system.In the face of uncertain user access and the differences of workloads between heterogeneous systems,the key to the problem lies in how to predict the load of the system and automatically take corresponding actions.This thesis is focused on the research of prediction algorithm based on SVM and neural network model,the main contents are as follows: 1)Based on the existing prediction algorighms designing load pattern recognition and adaptive selection of forecasting model;2)Designing the computational method of quantitative decision-making for seriver scaling.One of the innovations of this thesis is to dynamically apply the most suitable prediction algorithm according to the pattern workload time series.This thesis presents an adaptive prediction model including the workload pattern distinction module,model selection module and prediction module based on neural network and SVM regression.The predictor selection module based on strategy model and template method design pattern can be extended to deal with more types of workload in the future and select the corresponding customized prediction algorithm automatically.According to the nature of distributed cloud services,this thesis designs the quantification scaling decision-making module,which innovatively combines the call chain data from distributed tracking system to get weighted value for transitive workload caused by upstream services calling downstream serviers,forecast the downstream service workload by summing the weighted value of load prediction value of top upstream services,and adjust resources distribution or deployment according to the future load changes and performance test reports.Finally,based on the built Kubernetes platform and its deployment of the e-Commerce cloud service,the implementation of auto-scaling system and verification analysis is carried out.The experimental results show that the auto-scaling method proposed in this thesis can select the correct prediction model according to the workload pattern and have more accurate results.Compared to the reactive dynamic scaling based method,it can use less computing resources under the condition of assurance of stable performance and improve the response rate of about 15% and save about 23% of the computing resources.The adaptive method of autoscaling cloud services has been researched and tested in the company,and it is likely to be applied directly to the production environment in the future.