E级计算系统中基于SSD的高性能高可靠IO节点子系统关键技术研究

The emerging Exascale computing systems are mostly likely to be prohibited by the IO subsystem. By considering that IO node is an efficient method improving IO performance in high performance computing systems, this work proposes to deploy the SSD-based IO node into Exascale computing systems. In such a novel IO subsystem, there are three issues should be addressed. First, the reliability of IO node must be considered seriously as SSDs survive limited lifespan. Second, SSDs are able to supply a plethora of bandwidth which cannot be completely exploited by existing software stack. Third, simply introducing IO nodes into high performance computing system fails to improve the backend storage systems. Accordingly, this work aims at building a reliable as well as high performance IO node cluster by taking measures in three aspects. First, we explore novel architecture that enhances the reliability of SSD-based IO node. Based on the reliable platform built above, we optimize the parallel file system to make IO nodes be capable of responding to ultra-concurrent IO requests. Finally, we develop some intelligent mechanism to optimize the interaction between IO nodes and the backend parallel file system, thus improving the backend storage system as well. Technologies proposed in this research are promising to be applied to the emerging Exascale computing systems.

在未来的E级高性能计算系统中，IO子系统将是重要的瓶颈之一，而引入IO节点则是提高IO性能的有效手段。本研究拟在E级高性能计算系统中引入基于固态盘的IO节点系统。在此子系统中存在三方面的问题：（1）固态盘寿命有限，基于SSD的IO节点系统可靠性问题尤其突出；（2）固态盘能够提供空前的性能，但现有软件系统不能充分发挥基于SSD的IO节点系统的优势；（3）简单引入IO节点系统并不能完全消除后端存储系统的性能瓶颈。相应地，本项目从三个方面展开研究：基于固态盘的IO节点系统高可靠体系结构、面向高性能IO节点系统的数据并发访问技术、面向后端存储系统的IO节点系统优化技术。研究的成果将确保基于固态盘的IO节点系统稳定可靠和性能充分发挥，相关技术将应用到未来的E级高性能计算系统中。

E级计算系统将包含十万量级的计算节点和千万量级的处理器核心，相应地，在如此规模的系统上运行的E级应用将派生出千万量级的进程，大量进程的并发IO势必对底层文件系统形成强大的压力。近年来，以闪存为代表的非易失存储介质发展迅速，在计算系统中配备IO节点、并通过非易失存储介质增强IO节点的IO能力，是缓解E级系统中IO瓶颈的有效手段。本项目主要研究基于SSD的IO节点子系统，并延伸到RRAM、STT-RAM等新型非易失存储介质，重点解决IO节点子系统面临的可靠性、并发性问题。具体研究点包括：基于固态盘的IO节点系统高可靠体系结构、面向高性能IO节点系统的数据并发访问技术、面向后端存储系统的IO节点系统优化技术。在三年执行周期内，项目研究团队取得一系列成果：采用重复数据删除技术将SSD写性能提升21%；采用盘内缓存技术克服了读干扰问题；设计了基于Flash和DRAM的Ultra-DIMM混合主存系统；论证了采用纠删码技能能够将SSD寿命延长百倍以上；提出了多种针对SSD的编码方法；提出了基于图模型的文件系统名字空间管理方法；设计了新型的并行文件系统元数据组织结构；采用Key-Value系统构建元数据存储系统；采用代理技术缓解元数据瓶颈问题；利用SSD缓存提高文件系统的并发性；最终汇聚以上技术实现了面向高速IO加速节点的并行文件系统原型，并在其上展开应用优化研究。研究过程中发表论文16篇，其中CCF A类期刊1篇，CCF B类会议2篇，SCI检索4篇，EI检索14篇，两篇国内会议论文被评为优秀论文推荐到一级学报《计算机研究与发展》，申请专利11项。项目研究成果对E级计算系统中的IO子系统构建具有重要指导意义。