FineFlow:FaaS工作流部署优化与执行系统  

FineFlow:FaaS Workflow Deployment Optimization and Execution System

作　　者：刘璐 高浩城 陈伟[1,2,3,4] 吴国全 魏峻[1,2,3,4] LIU Lu;GAO Hao-Cheng;CHEN Wei;WU Guo-Quan;WEI Jun(Institute of Software,Chinese Academy of Sciences,Beijing 100190,China;University of Chinese Academy of Sciences,Beijing 100049,China;State Key Laboratory of Computer Science(Institute of Software,Chinese Academy of Sciences),Beijing 100190,China;University of Chinese Academy of Sciences,Nanjing,Nanjing 211135,China)

机构地区：[1]中国科学院软件研究所,北京100190 [2]中国科学院大学,北京100049 [3]计算机科学国家重点实验室(中国科学院软件研究所),北京100190 [4]中国科学院大学南京学院,江苏南京211135

出　　处：《软件学报》2025年第2期488-510,共23页Journal of Software

基　　金：国家重点研发计划(2021YFB2600301);中国科学院软件研究所重大项目(ISCAS-ZD-202302)。

摘　　要：FaaS(function-as-a-service,函数即服务)工作流由多个函数服务编排而成,通过对多个函数的协调控制来实现复杂的业务应用.当前FaaS工作流系统主要基于集中式的数据存储实现函数间的数据传递,导致FaaS函数间的数据传输开销大,显著影响应用性能.在高并发情况下,频繁的数据传输还会产生严重的网络带宽资源争用,导致应用性能下降.针对上述问题,基于函数服务间的细粒度数据依赖分析,提出一种基于关键路径的函数部署优化方法,设计了依赖敏感的数据存取与管理机制,有效减少函数间数据传输,从而降低FaaS工作流应用执行的数据传输时延和端到端时延.设计实现了FaaS工作流系统FineFlow,并基于5个真实FaaS工作流应用开展实验评估.实验结果表明,相比于基于集中式数据存储函数交互机制的FaaS工作流平台,FineFlow能够有效降低FaaS工作流应用的数据传输时延:最高降低74.6%,平均降低63.8%;平均降低应用端到端执行时延19.6%.特别地,对于具有明显细粒度数据依赖的FaaS工作流应用,相比于现有的基于数据本地性的优化方法,FineFlow能够使数据传输时延和端到端时延进一步分别降低28.4%和13.8%.此外,FineFlow通过减少跨节点的数据传输,能够有效缓解网络带宽波动对FaaS工作流执行性能的影响,提升应用性能受网络带宽影响的鲁棒性.A function-as-a-service(FaaS)workflow,composed of multiple function services,can realize a complex business application by orchestrating and controlling the function services.The current FaaS workflow execution systems achieve data transfer among function services mainly based on centralized data storages,resulting in heavy data transmission overhead and affecting application performance significantly.In the cases of high concurrency,frequent data transmission will also cause serious contention for network bandwidth resources,resulting in application performance degradation.To address the above problems,this study analyzes the fine-grained data dependency between function services and proposes a critical path-based FaaS workflow deployment optimization method.In addition,the study designs a dependency-sensitive data access and management mechanism to effectively reduce the data transmission between function services,thereby reducing the data transmission latency and end-to-end execution latency of FaaS workflow applications.The study implements a FaaS workflow system,FineFlow,and conducts experiments based on five real-world FaaS workflow applications.The experimental results show that FineFlow can effectively reduce the data transmission latency(the highest reduction and the average reduction are 74.6%and 63.8%,respectively)compared with the FaaS workflow platform with the centralized data storing-based function interaction mechanism.On average,FineFlow reduces the latency of the end-to-end FaaS workflow executions by 19.6%.In particular,for the FaaS workflow application with fine-grained data dependencies,FineFlow can further reduce its data transmission latency and the end-to-end execution latency by 28.4%and 13.8%respectively compared with the state-of-the-art work.In addition,FineFlow can effectively alleviate the impact of network bandwidth fluctuations on application performance by reducing cross-node data transmission,improving the robustness of application performance influenced by the network bandwidth

关 键 词：FaaS工作流 函数即服务 服务器无感知计算 数据本地性 有向无环图 关键路径 部署优化 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]