【作者】 李林；

【导师】 陈咸彰；

【作者基本信息】 重庆大学 ， 工程（专业学位）， 2024， 硕士

【摘要】 自动驾驶和机器人等嵌入式终端设备上数据量的迅速增长,为数据处理的效率和能效提出了更高的要求和挑战。近年来,许多研究提出了以数据为中心的近数据处理(Near-Data Processing,简称NDP)架构,以减轻主机所面临的计算压力以及同存储设备之间数据传输的巨大开销。通常,这些工作将应用的部分数据处理操作卸载到可计算存储(Computational Storage)内的嵌入式处理器或硬件加速器上执行,从而减少数据搬移,提高系统整体的数据处理性能和效率。然而,现有架构在数据通路和访问同步上仍存在一些问题。首先,在任务发射前,NDP应用通常依赖于主机内核的I/O堆栈以定位所请求文件的数据块地址,从而引入额外的数据传输开销。此外,现有工作缺乏高效的访问同步机制,难以应对主机与可计算存储之间潜在的数据访问冲突。针对前者,本文提出了一种文件语义提取器(File Semantics Retriever,简称FSR),直接在存内实现对文件语义的识别及提取,以优化近NDP操作的数据通路。随后,本文设计了一种设备文件锁(Device File Lock,简称DFLock)并基于该锁建立对应同步原语,以实现主机进程与存内模块之间的访问同步,保障数据一致性。本文所做主要工作概况如下:(1)针对现有NDP架构在数据定位时存在的中间数据搬移开销,本文提出了FSR以实现对数据通路的优化。FSR在不影响主机端文件系统运行的同时直接识别解析存储内的文件命名空间布局和元数据结构,且脱离文件系统环境对文件语义进行提取。此外,本文针对现有方案及所提方案进行了工作流程抽象和理论性能建模,深入分析了二者对不同场景的适应性。(2)针对主机进程与存内模块之间的访问同步问题,本文提出一种轻量级文件锁(DFLock),并结合存内软硬件特性建立对应的同步原语。此外,设计了基于队列的锁等待机制,在追求并发的同时保证公平性,避免饥饿现象。本文基于Cosmos Plus Open SSD平台实现了所提出的FSR及DFLock算法,并在该平台上实现并部署了字符串匹配任务。结合主机端应用程序构建了原型系统,采取真实应用负载对原型系统开展了验证测试。实验表明所提方案均行之有效,相较于现有方案,FSR在文件定位上的时延降低了73.6%,并且能够在真实NDP操作上取得54.0%的性能提升。更多还原

【Abstract】 The rapid growth of data in embedded terminal devices,such as autonomous vehicles and robots,has posed higher demands and challenges for data processing efficiency and energy consumption.In recent years,many studies have proposed data-centric Near-Data Processing(NDP)architectures to alleviate the computational pressure faced by the host and the enormous overhead of data transfer between the host and storage devices.Typically,these works offload some of the application’s data processing operations to embedded processors or hardware accelerators within Computational Storage to reduce data movement and improve the overall data processing performance and efficiency of the system.However,existing architectures still have some issues with data paths and access synchronization.First,before the task launch,NDP applications usually rely on the host kernel’s I/O stack to locate the data block addresses of the requested files,introducing additional data transfer overhead.Moreover,existing works lack efficient access synchronization mechanisms,making it difficult to cope with potential data access conflicts between the host and computational storage.To address the former,this thesis proposes a File Semantics Retriever(FSR)to directly realize the identification and extraction of file semantics in-storage,optimizing the data path of NDP operations.Subsequently,this thesis designs a Device File Lock(DFLock)and establishes corresponding synchronization primitives based on the lock to achieve access synchronization between host processes and in-storage modules,ensuring data consistency.The main work of this thesis is outlined as follows:(1)To address the intermediate data movement overhead of existing NDP architectures during file locating,this thesis proposes FSR to optimize the data path.FSR directly parses the in-storage file namespace layout and metadata structures without affecting the operation of the host-side file system,and retrieves file semantics independent of the file system environment.In addition,this thesis abstracts the workflows of existing solutions and the proposed solution,and conducts theoretical performance models to deeply analyze their adaptability to different scenarios.(2)To address the access synchronization issue between host processes and in-storage modules,this thesis proposes a lightweight file lock(DFLock)and establishes corresponding synchronization primitives combining the characteristics of in-storage software and hardware.Furthermore,a queue-based lock waiting mechanism is designed to ensure fairness while pursuing concurrency,avoiding starvation.This thesis implements the proposed FSR and DFLock algorithms on the Cosmos Plus Open SSD platform and deploys a string search task on the platform.A prototype system is built in conjunction with the host-side application,and verification tests are conducted on the prototype system using real application workloads.Experimental results show that the proposed solutions are effective.Compared with existing solutions,FSR reduces the latency of file locating by 73.6%and achieves a 54.0%performance improvement in real NDP operations.更多还原