Performance Evaluation And Optimization For Serverless In Cloud Native System

Cloud-native relies on containerized packaging,native application modularity,fine-grained resource control and other features to provide a closed-loop "developdeploy-run" paradigm for cloud computing,which has become the de facto standard in cloud computing after long-term development and iteration.Serverless computing,designed natively for the cloud,inherits and develops the advantages of cloud-native,fundamentally transforming the cloud computing architecture from virtual machines and containers as a service to functions as a service.Serverless improves the efficiency and flexibility of cloud-native platforms,solves the problems of low cloud resource utilization and complex cloud service architecture,and is expected to become the new standard design concept for cloud-native platforms.However,compared to common cloud loads such as long-running microservices and offline task processing,Serverless systems have new application characteristics such as "short lifecycle" and "instance statelessness" as a function of the current cloud-native infrastructure and application controls are still unable to meet their low latency response requirements.Many research efforts have explored lightweight function runtimes and new hardware-based high-speed data centers to address the mismatch between the current cloud architecture and the load characteristics of Serverless systems,effectively improving the function response performance of Serverless systems.Existing research works are still inadequate in characterizing the performance for complex and sensitive Serverless loads and optimizing the platform performance under bursty scenarios.In this thesis,we take representative open source Serverless systems in cloudnative systems and public cloud loads as the research objects,aiming to study and solve the poor effect of existing coarse-grained resource management strategies in dealing with complex and sensitive loads,and explore novel resource management and task scheduling methods suitable for the characteristics of cloud-native applications.The main work of this thesis is as follows.(1)A cross-tier performance-aware benchmark suite-BBServerless for bursty loads of cloud-native Serverless systems is designed to achieve accurate evaluation of cloud-native loads.In this thesis,we extract the application performance characteristics of Serverless systems and their important features in resource management and scheduling policies by studying the functions of cloud-native Serverless systems and their representative loads,and accurately profiling them in terms of complexity,sensitivity and resource interference.Meanwhile,a performance benchmark test suite with end-to-end,OS,and hardware-level multi-layer awareness is designed on this basis.Finally,the multi-layer performance metrics of three representative platforms are comprehensively analyzed using BBServerless,on which the performance bottlenecks in the design of seven platforms are observed and the reasons for the performance bottlenecks are analyzed.(2)A location-aware Serverless system function scheduling algorithm is proposed to address the lack of performance of time-localized components and optimize the platform performance.Based on the performance characteristics characterized by the benchmark suite,this thesis investigates function scheduling and optimization strategies for Serverless systems,and proposes a comprehensive scheduling strategy combining location-and affinity-awareness and Bayesian approximation estimation.In particular,the performance problems caused by the "short life cycle of functions","lightweight" and "dynamic interaction of functions" of Serverless systems are addressed.The performance optimization method based on the locality principle is proposed.The experimental results show that the proposed approach effectively improves the end-to-end performance of Serverless loads.