【作者】 李磊；

【导师】 申兆岩；

【作者基本信息】 山东大学 ， 计算机技术（专业学位）， 2025， 硕士

【摘要】 随着人工智能大模型、云计算以及大数据等技术的快速发展,数据规模呈指数级增长,数据量的激增和对高性能计算的需求日益增加。传统存储与计算分离的架构面临数据传输带宽不足、处理延迟高、能耗大等瓶颈。当前主流的网络存储和存算融合方案(如基于CPU的软件定义存储、分布式存储系统等)存在可扩展性受限、硬件资源利用率低、难以实现低延迟与高带宽协同等问题。另外,在网络、存储、计算三者融合的架构设计领域仍存在研究空白。针对上述挑战,本文首次提出一种基于多核异构FPGA的网络、存储、计算融合架构,通过与华为公司合作研发,实现了网络协议卸载、存储数据管理、计算加速功能的硬件级融合。本文依托Xilinx Zynq UltraScale+MPSoC核心板,基于自主设计的搭载Flash芯片的FPGA平台,主要工作体现在三个功能模块的设计:1)网络模块,采用RoCE v2协议卸载引擎,联合LWIP协议栈建链,通过硬件加速RDMA通信,降低网络协议栈处理开销,实现高吞吐量、低延迟的网络传输;2)存储模块,自主设计Flash驱动控制器,集成FTL(闪存转换层)、FIL(闪存接口层)及ECC校验纠错算法,优化NANDFlash读写性能与可靠性,支持大容量数据的高效存取;3)计算模块,为突出数据安全的重要性,本文以加解密算法为例,实现了 Chacha20流密码算法的并行化硬件加速模块,支持数据加密与解密的高速处理,满足现代数据安全需求。融合网存算功能搭建轻量存储服务器,通过RoCE v2协议与远端Client建立RDMA通信链路,在数据传输阶段触发Chacha20算法实现硬件级数据流加密,最终将数据经存储控制模块持久化至NAND Flash介质,实现“协议卸载-实时加密-持久化存储”的异构硬件加速闭环。实验分为功能完备性测试与系统级性能评估。在功能验证方面:1)网络模块可精准解析RoCE v2协议报文,实验测试结果显示READ/WRITE请求处理速率分别可达560MB/s和1060MB/s;2)存储模块完成页级地址映射与ECC校验纠错功能,读写数据一致性的测试结果验证了其高效读写能力和数据准确性,Flash写可达300MB/s,Flash读可达266MB/s;3)计算模块成功实现数据加解密且前后数据一致,确保了数据传输过程中的安全性,处理性能可达520MB/s。整体性能评估方面,通过联合网络、存储、计算三个模块部署轻量级存储服务器,受存储模块和硬件资源限制,端到端协同工作整体吞吐量呈现存储依赖特性,在数据持久化场景下系统有效带宽为308MB/s,较纯网络传输模式下降70.9%。尽管如此,随着硬件性能的提升,本文的研究设计为云计算、人工智能等领域提供了低延迟、高带宽、高安全的网存算融合解决方案,推动了新型存储架构与计算平台的协同创新,具有重要的工程应用价值与产业化前景。更多还原

【Abstract】 With the rapid development of technologies such as Artificial Intelligence Big Models,Cloud Computing and Big Data,the scale of data is growing exponentially,and the surge in data volume and demand for high-performance computing is increasing.The traditional architecture of storage and compute separation faces bottlenecks such as insufficient data transmission bandwidth,high processing latency,and high power consumption.The prevailing network storage and storage-computing convergence solutions(e.g.,CPU-based software-defined storage,distributed storage systems,etc.)encounter limitations in scalability,inadequate hardware resource utilization,and challenges in achieving low-latency,high-bandwidth synergy.Additionally,there is a notable research gap in the domain of architecture design for the convergence of networking,storage,and computing.To address these challenges,this paper proposes for the first time a network,storage,and compute convergence architecture based on multi-core heterogeneous FPGAs,which realizes hardware-level convergence of network protocol offloading,storage data management,and compute acceleration functions through collaborative research and development with Huawei.Relying on the Xilinx Zynq UltraScale+MPSoC core board,this paper is based on an independently designed FPGA platform equipped with a Flash chip,and the main work is embodied in the design of three functional modules:1)network module,using RoCE v2 protocol offload engine,joint LWIP protocol stack to build the chain,through hardware accelerated RDMA communication,reduce the network protocol stack processing overhead,to achieve high throughput and low latency network transmission;2)Storage module,independently designed Flash driver controller,integrated FTL(Flash Translation Layer),FIL(Flash Interface Layer)and ECC checksum error correction algorithms to optimize NAND Flash read/write performance and reliability,and support efficient access to large-capacity data;3)Computing module,in order to highlight the importance of data security,this paper takes encryption and decryption algorithms as an example,and implements a parallelized hardware acceleration module for Chacha20 stream cipher algorithm,which supports high-speed processing of data encryption and decryption to meet the modern data security requirements.The fusion of network storage and computing function builds a lightweight storage server,establishes RDMA communication link with remote client through RoCE v2 protocol,triggers Chacha20 algorithm to realize hardware-level data stream encryption during data transmission,and finally persists the data to NAND flash media through storage control module,thus realizing the heterogeneous "protocol offload-real-time encryption-persistent storage" of the closed loop of hardware acceleration.The experiment is divided into two stages:a functional completeness test and a system-level performance evaluation.In terms of functional verification,the network module is capable of accurately parsing RoCE v2 protocol messages.The experimental test results demonstrate that the processing rate of READ/WRITE requests can reach 560MB/s and 1060MB/s,respectively.The storage module completes page-level address mapping,ECC check,and error correction.The experimental test results of the read/write data consistency demonstrate the high efficiency and accuracy of the read/write ability,with Flash write up to 300MB/s and Flash read up to 266MB/s.The computing module successfully realizes data encryption and decryption and data consistency before and after data transmission.The Flash write speed is up to 300MB/s,and the Flash read speed is up to 266MB/s.The computing module successfully encrypts and decrypts data,ensuring data consistency before and after transmission.This module also ensures the security of the data transmission process and has a processing performance of up to 520MB/s.However,the deployment of lightweight storage servers through the three modules of the joint network,storage,and computing is constrained by the storage module and the hardware resources.Similarly,the overall end-to-end collaborative work is limited by the storage module and the hardware resources.Consequently,the overall throughput of end-to-end cooperative work exhibits storage-dependent characteristics,and the effective bandwidth of the system in the data persistence scenario is 308MB/s,which is 70.9%lower than that of the pure network transmission mode..Nevertheless,with the improvement of hardware performance,the research design of this paper provides a low-latency,high-bandwidth,high-security network storage and computing fusion solution for cloud computing,artificial intelligence and other fields,and promotes the collaborative innovation of the new storage architecture and computing platform,which has an important value of engineering applications and industrialization prospects.更多还原