Research On Intelligent Multi-Resource Collaborative Scheduling For Latency-Sensitive Microservices

User-facing services are gradually evolving from traditional monolithic architecture to loosely-coupled combinations of microservices.These services are typically interactive and latency-sensitive applications that require strict end-to-end tail latency Service Level Objective(SLO)to ensure user satisfaction.Therefore,the microservices that make up user-facing services are also latency-sensitive.While microservice architecture greatly improves the efficiency of development and operation,it also complicates resource allocation and performance guarantee due to complex dependencies across different latency-sensitive microservices.Existing resource scheduling systems mainly leverage auto-scaling to dynamically allocate resources for each latency-sensitive microservice.However,due to ignoring the impact of the dependencies across different latencysensitive microservices on the end-to-end tail latency,the resource utilization rate is relatively low.Recent works employ the Bayesian optimization methods to capture the performance impact of dependencies across different latency-sensitive microservices.However,as the resource allocation space increases,the efficiency of this method will be greatly reduced.Therefore,how to maximize the system’s resource utilization while guaranteeing end-to-end tail latency SLO for user-facing services has become a hot research topic of microservice resource scheduling.To deal with these problems,this study proposes an intelligent multi-resource collaborative scheduling for latency-sensitive microservices.Firstly,we design a multiresource collaborative allocation model based on multi-agent deep reinforcement learning.By capturing the complex relationship between workload,resource configuration,dependencies among different microservices,and end-to-end tail latency SLO,the model conducts optimal resource allocation for each microservice.On this basis,we propose a microservice performance anomaly handling method based on partial SLO.By proactively identifying the critical microservices that cause performance anomalies and dynamically reallocating resources,potential violations of end-to-end tail latency SLO can be eliminated.Finally,we design a lightweight distributed performance monitor.By continuously obtaining and providing feedback on resource usage,latency,and other information of microservice instances in real time,the monitor supports the optimization of resource allocations and performance anomaly handling.Testbed experiments show that the proposed method can significantly improve the utilization of CPU and memory resources while guaranteeing end-to-end tail latency SLOs for user-facing services.