基于Init-Less模式的Serverless冷启动优化方法研究

Research on Serverless Cold Start Optimization Methods Based on Init-Less Mode

摘要: 无服务器计算(Serverless)的兴起正为开发人员提供了更加有效的成本节约和弹性计算能力，该模式极大地提高了计算资源分配的灵活性，同时为真正按需租赁计算资源提供了可能。使得用户不必参与资源调度，自动实现无任何计算需求时不占用算力，同时在计算量攀升时及时扩容保证响应的及时性。然而目前基于容器虚拟化技术的无服务计算额外带来了冷启动问题，容器启动以及用户代码初始化会造成数秒的响应延迟，云计算厂商通常建议通过预留实例缓解此问题，但是这并不能消除容器扩容时的冷启动。同时，预留实例的方案也减弱了无服务计算相较于微服务节省成本的优势。现阶段多数研究目标集中于通过改造底层容器技术以尽量减少冷启动耗时，但是相对于广泛采用的Docker方案，难以在实际应用场景中对其进行取代。另一方面，仅仅对底层容器进行改造，无法减少用户代码初始化的额外耗时。本文分析了在无服务器计算场景下常规Docker容器应用的开发模式，提出了基于init-less和lazy-restore策略的Docker容器冷启动优化方法。该方法对用户代码的网络、文件、内存等计算资源的使用模式进行约定，基于CRIU技术捕获已初始化完毕的Docker容器的快照，并通过两阶段lazy-restore对常规Docker容器的启动流程进行替代。根据上述方法实现了docker-initless，绕过了容器应用的启动时间瓶颈，极大降低了Docker容器的冷启动延迟。实验从内存、文件资源等方面对比了Docker和docker-initless，验证了docker-initless在不额外预留计算资源的条件下对容器冷启动优化的有效性，同时可以保证对现有Serverless方案的兼容性。

Abstract: The rise of Serverless computing (Serverless) is providing developers with more effective cost savings and flexible computing capabilities. This model greatly improves the flexibility of computing resource allocation and at the same time makes it possible to rent computing resources on demand. So that users do not need to participate in resource scheduling, automatically realize that no computing power is occupied when there is no computing demand, and at the same time, the capacity in time is expanded to ensure the timeliness of the response when the computing volume rises. However, the current Serverless computing based on container virtualization technology has addition-ally brought about cold start problems. Container startup and user code initialization will cause a response delay of several seconds. Cloud computing vendors usually recommend using reserved in-stances to alleviate this problem, but cold start cannot be eliminated during container expansion. At the same time, the reserved instance solution also reduces the cost-saving advantage of Serverless compared to microservice. At this stage, most research goals are focused on reducing the time-consuming on cold start by modifying the underlying container technology. But compared to the widely used Docker solution, it is difficult to replace it in actual application scenarios. On the other hand, merely modifying the underlying container cannot reduce the additional time-consuming initialize of user code. This article analyzes the development mode of conventional Docker container applications in Serverless computing scenarios, and proposes a cold start optimization method for Docker containers based on init-less and lazy-restore strategies. This method stipulates the usage patterns of computing resources such as network, files, and memory of user code, captures a snapshot of the initialized Docker container based on CRIU technology, and replaces the startup process of the conventional Docker container through a two-stage lazy-restore. According to the above method, docker-initless is realized, which bypasses the bottleneck of the startup time of container applications, and greatly reduces the cold start time of Docker containers. The experiment compares Docker and docker-initless in terms of memory, file resources, etc., and verifies the effectiveness of docker-initless in optimizing the cold start of containers without additional computing resources, while ensuring compatibility with existing Serverless solutions.