| Cloud computing resources are gradually becoming an important infrastructure in the field of information technology,among which Serverless architecture is a popular cloud computing technology.This technology has the characteristics of on-demand charging,automatic expansion and contraction,and low cost.The most important thing is that the serverless architecture shields developers from the underlying physical machine conditions and improves development efficiency.However,due to the characteristics of the architecture design,the function,as the smallest execution unit of the serverless platform,needs to be executed in the container.Before the function provides external services,it needs to go through multiple startup steps,which greatly affects the performance of the function operation.Therefore,solving the cold-start problem has become an important challenge for the serverless platform.At present,most cold-start optimization methods propose resource scheduling optimization methods from a system perspective.Related methods lack consideration of the heterogeneity of serverless functions,resulting in insufficient flexibility of the algorithm when facing functions in different fields,resulting in a high cold-start rate.In addition,the resource scheduling method tends to ignore the system revenue and memory consumption,two important indicators for cloud computing service providers,when optimizing the cold-start problem of the serverless platform,which eventually leads to the loss of system revenue of the cloud computing service provider.The resource scheduling method plays a key role in Serverless architecture.This paper studies the task resource allocation optimization strategy for system revenue and cold-start rate.The main contributions of this paper are:(1)Based on the execution and calling characteristics of serverless functions,this paper proposes a fine-grained cold-start optimization method: a job-level histogram keep-alive adjustment algorithm.The algorithm provides parameter options for adjusting algorithm performance according to actual business needs.Experiments prove that the cold-start optimization algorithm in this paper reduces the average cold-start of the system by 92% and15% respectively compared with the fixed keep-alive(default 10 minutes)method and the system-level histogram keep-alive method.(2)After in-depth research on the core billing models of mainstream cloud computing service providers,combined with the operating principle of the serverless platform and the job-level histogram keep-alive adjustment algorithm,this paper proposes a serverless system revenue model and a function memory occupation model;the model has relatively good versatility,it can provide quantitative indicators of system benefits and memory usage in the serverless platform task allocation strategy.(3)Aiming at the problem that the task scheduling strategy related to the cold-start problem affects the system revenue problem that cloud computing service providers care about,this paper proposes an allocation algorithm for the system maximum revenue problem,which solves the system maximum revenue and its task allocation and use the job-level histogram keep-alive adjustment algorithm to ensure the cold-start rate of each function.Experiments show that: Compared with the system histogram method,the allocation algorithm of maximizing system revenue can reduce the average cold-start rate by 16.5% and increase the system revenue by 395%.(4)To solve the problem of further reducing the average cold-start rate of the serverless platform and taking into account important parameters such as system revenue,this paper proposes a resource allocation algorithm for the minimum cold-start rate of the system.The algorithm solves the minimum average cold-start rate of the system and the task allocation under this condition when the system income is higher than the threshold set by the user and the memory consumption is lower than the limit condition.Experiments show that the minimum cold-start problem allocation algorithm can further reduce the average cold-start rate of the system by 29.8% compared with the original job-level histogram keep-alive adjustment algorithm. |
