Design And Implementation Of Service Collaboration And Load Balancing Strategy For Cloud Services

Nowadays the popularity of the Internet and the extension of the scope of Internet applications pose huge challenges to the extensibility and performance of cloud service systems.On one hand,the service architecture is the basis for solving extensibility problems.Service-oriented architecture and microservices architecture have become mainstream cloud service system architectures for their more detailed service splitting,lower service coupling,and stronger system scalability compared with the monolithic architecture.Therefore,this thesis designs a service collaboration system for the service collaboration problems that may occur when applying these two architectures.On the other hand,load balancing technologies play an important role in improving the performance of cloud service systems.However,various conventional load balancing strategies have various problems in practical applications.Therefore,this thesis proposes a linear regression based dynamic load balancing strategy.Firstly,this thesis designs main components including the service registry,service configuration center,gateway load balancer in the service collaboration system for solving service collaboration problems,such as service discovery and communication,service configuration management,and load balancing.Storage structure,access model,operations and some other aspects of the service registry and the service configuration center are designed based on ZooKeeper,which is a distributed coordination service.Based on the service discovery and configuration management functions provided by the service registry and the service configuration center,the gateway load balancer supports dynamic service and configuration for load balancing.Secondly,a client-side cache for service discovery and service configuration queries based on the watcher mechanism of Zookeeper is designed in this thesis.For the cache expiration feature,a timeout queue is designed in this thesis,which guarantees that at most one timer is started at the same time.In tests,the applying of the client-side cache greatly improved the access performance of the service registry and the service configuration center.Finally,a dynamic load balancing strategy is proposed in this thesis,which uses linear regression to learn the relationship between RPS(Requests Per Second)and load value from the data of running state of every service node for real-time load value prediction.In addition,this strategy designs an optimization scheme in the initial parameter setting of the model and the iteration stopping condition to improve the training speed.The three-sigma rule is adopted in the using condition of the models,so that the models are used more reasonably.Tests show that this strategy has a capability of load adjusting,and behaves well in balancing and RPS performance.