Research On Reinforcement Learning Based Microservice Scheduling Algorithm

With the rapid development of mobile Internet and Io T,the number of Internet terminals is increasing rapidly,and computer systems need to process more and more requests at the same time.The big Internet enterprises have turned to using microservice architecture which has more flexible system resource scheduling to build their own applications.With the rapid growth of Internet traffic,how to build an efficient scheduling strategy has become a challenge for microservice systems.This paper researched the micro-service scheduling by dividing the existing computing resources of the system into two aspects: load balancing and auto scaling:1.The paper has proposed a formalized representation of resources and load for the micro-service system containing heterogeneous computing resources and a unified load representation method including heterogeneous computing resources was established for the establishment of environmental observation values in load balancing and elastic expansion scenarios.2.In the load balancing scenario where the resource capacity of the target system is different,the environment definition,action space and reward function for reinforcement learning are established based on the formalized resource and load representation methods.By constructing a load balancing framework based on reinforcement learning,this paper has solved the problem that the load does not match with the system resources when the traditional load balancing algorithm is confronted with target service systems with different resource capacity,and leads to higher average response time.This paper also removed human intervention to enact the right weights.3.The environment definition,action space and reward function for reinforcement learning were established based on the formal representation of resources and load for the elastic expansion scenario.In this paper,an automated container elastic expansion decision component based on reinforcement learning is built.Compared with the auto scaling method based on load prediction or trend learning,the decision component realized in this paper fills the shortcoming that the strategy learning cannot be carried out immediately in the case of lack of data in cold start.In the component experiment,Deep Reinforcement-learning based Load Balancer(DRLB)achieves a better average response time than the traditional method in the case of the target services with unbalanced computing resources.With the control from the Deep Reinforcementlearning based Auto Scaler(DRAS),Docker Swarm realizes a faster elastic expansion response from the static threshold strategy,which effectively improves the response speed and the stability of the response time of request processing when the load changes.In the experiment of microservice scheduling framework Deep Reinforcement-learning based Microservice Scheduler(DRMS),DRLB and DRAS realize microservice scheduling based on reinforcement learning together,which improves the response speed of peak load and optimizes the overall system response time of microservice system.