摘要
服务器无感知计算凭借其弹性、高效和低成本的优势已成为下一代云计算的重要组成部分.随着服务器无感知应用的复杂度和多样性的增加,现有服务器无感知计算平台通常将不同应用的函数实例混合部署在统一服务器主机中,并使用统一的CPU调度策略(如Linux完全公平调度)进行调度.由于不同应用对性能、延迟以及调度公平性等的要求都不尽相同,使得单一的全局调度策略,难以满足多样化的调度需求.此外,实验发现,函数实例的执行时间跨度较大,短任务占比高,且冷启动对任务完成时间影响显著.这使得现有调度策略难以兼顾任务的效率与公平性,尤其会增加了短任务的调度延迟.为解决上述问题,我们提出并实现了一种调度隔离机制,允许在同一服务器上为单个或多个应用使用独立的CPU调度策略.此外,我们设计了FaaSchedule,一种新的可定制调度策略,综合考虑函数的执行时间、启动时间和等待时间,并支持根据不同应用需求动态定制调度策略.在OpenWhisk平台上的实验表明, FaaSchedule显著降低了短任务的平均完成时间,并有效支持调度隔离与策略定制.
Serverless computing,with its advantages of elasticity,efficiency,and low cost,has become a key component of next-generation cloud computing.As the complexity and diversity of serverless applications increase,existing serverless platforms typically deploy function instances from different applications on the same server and apply a unified CPU scheduling policy,such as Linux’s completely fair scheduler(CFS).However,due to the varying requirements of different applications in terms of performance,latency,and scheduling fairness,a single global scheduling policy struggles to meet diverse scheduling demands.Moreover,our research reveals that function instances exhibit a wide range of execution times,with a high proportion of short tasks,and that cold starts significantly impact task completion times.These characteristics make current scheduling policies ineffective in balancing efficiency and fairness,especially increasing the scheduling latency for short tasks.To address these issues,we propose and implement a scheduling isolation mechanism that allows single or multiple applications with similar requirements to adopt independent CPU scheduling policies on the same server.Additionally,we design FaaSchedule,a novel customizable scheduling policy that takes into account function execution time,startup time,and waiting time,and supports dynamic customization of scheduling strategies based on the needs of different applications.Experiments on the OpenWhisk platform show that FaaSchedule significantly reduces the average completion time of short tasks with minimal impact on the performance of long tasks,while effectively supporting scheduling isolation and policy customization.
作者
张信民
李星儒
樊浩
黄卓
吴松
姚德中
金海
余辰
Xinmin ZHANG;Xingru LI;Hao FAN;Zhuo HUANG;Song WU;Dezhong YAO;Hai JIN;&Chen YU(National Engineering Research Center for Big Data Technology and System,Huazhong University of Science and Technology,Wuhan 430074,China;Key Laboratory of Services Computing Technology and System,Ministry of Education,Huazhong University of Science and Technology,Wuhan 430074,China;Key Laboratory of Cluster and Grid Computing,Hubei Province,Huazhong University of Science and Technology,Wuhan 430074,China;School of Computer Science and Technology,Huazhong University of Science and Technology,Wuhan 430074,China)
出处
《中国科学:信息科学》
北大核心
2025年第3期481-499,共19页
Scientia Sinica(Informationis)
基金
国家自然科学基金(批准号:2022YFB4502001,62032008,62232012)资助项目。
