【作者】 刘昕；

【导师】 左德承；

【作者基本信息】 哈尔滨工业大学 ， 电子信息（专业学位）， 2025， 硕士

【摘要】 近年来随着物联网设备的迅速普及,边缘计算成为一项关键技术,提供了低延迟和实时处理能力,与此同时流计算技术的发展经历了多个阶段,与边缘计算的结合逐渐成为一个重要方向,推动了工业物联网、智慧城市、智慧交通等关键应用的落地。然而,边缘环境中的计算资源受限、设备异构性以及复杂的网络条件对传统为云环境设计的流处理系统提出了重大挑战,因此,本研究主要聚焦于如何将流计算更好的与边缘环境结合。动态重配置是将流计算与边缘环境相结合的有效方式,通常通过任务重分区、调度优化及弹性扩展等策略对流计算作业进行重构,但实际上现有研究大多单独探讨单一策略的优化,优化效果有限且仅适用于特定维度,尽管已有一部分研究提出了多层次协同优化的思路,但这些方法未充分考虑各层次之间的协同关系,多数仅停留在独立优化的层面。本研究针对流处理系统在边缘环境中需要适应动态变化的需求,设计了运行时自适应的关键机制,包括分区、调度和弹性管理,构建了流处理系统重配置的数学模型,并提出了一种结合分区、调度和弹性管理策略的多层次优化框架,采用两阶段的重配置方法,第一阶段在资源受限的边缘环境下快速得到有效的重配置方案,第二阶段在此基础上进行调整,在更大的优化空间内进行搜索。本研究引入了一种细粒度状态迁移机制,以在重配置期间最大限度减少延迟和资源开销。通过将状态划分为更小、对用户透明的粒度,系统能够在保持一致性的同时高效完成迁移。此外,通过对状态迁移过程进行建模以计算最佳状态迁移粒度,实验证明了该建模的正确性。本文基于Apache Flink实现了细粒度迁移机制及多层次协同优化算法,并结合负载预测机制及开销收益模型,使系统在边缘环境下能够避免过多的重配置。在边缘设备上的对比实验结果表明,所提出的系统在边缘环境下的表现优于现有方案。更多还原

【Abstract】 In recent years,with the rapid proliferation of Io T devices,edge computing has emerged as a critical technology,offering low latency and real-time processing capa-bilities.At the same time,the development of stream computing has undergone multi-ple stages,and its integration with edge computing has gradually become an important direction,driving the implementation of key applications such as industrial Io T,smart cities,and intelligent transportation.However,the limited computational resources,de-vice heterogeneity,and complex network conditions in edge environments pose signifi-cant challenges to stream processing systems originally designed for cloud environments.Therefore,this study focuses on how to better integrate stream computing with edge en-vironments.Dynamic reconfiguration is an effective approach to integrating stream comput-ing with edge environments.It typically involves strategies such as task repartitioning,scheduling optimization,and elastic scaling to restructure stream processing jobs.How-ever,existing studies mainly focus on optimizing individual strategies in isolation,leading to limited optimization effects and applicability restricted to specific dimensions.Al-though some research has proposed multi-level collaborative optimization approaches,these methods often fail to fully account for the interdependencies between levels,re-maining at the level of independent optimization.To address the need for stream process-ing systems to adapt to dynamic changes in edge environments,this study designs run-time adaptive mechanisms,including partitioning,scheduling,and elasticity management.A mathematical model for stream processing system reconfiguration is constructed,and a multi-level optimization framework combining partitioning,scheduling,and elasticity management strategies is proposed.This framework employs a two-stage reconfigura-tion method:the first stage rapidly generates effective reconfiguration plans in resource-constrained edge environments,while the second stage refines these plans by exploring a larger optimization space.This study introduces a fine-grained state migration mechanism to minimize la-tency and resource overhead during reconfiguration.By dividing states into smaller,user-transparent granularities,the system achieves efficient migration while maintaining con-sistency.Additionally,a model is developed to determine the optimal state migration granularity,and experiments validate the correctness of the modeling.The paper implements a fine-grained migration mechanism and a multi-level collab-orative optimization algorithm based on Apache Flink,combined with a load prediction mechanism and a cost-benefit model,to minimize excessive reconfigurations in edge en-vironments.Comparative experimental results on edge devices show that the proposed system outperforms existing solutions in edge environments.更多还原