本项目拟通过研究具有多级层次体系结构的并行存储系统的I/O优化支撑技术，为高性能计算机集群提供具有高性能、可扩展和低延迟等特性的存储基础设施。具体而言，拟研究的并行存储系统优化技术考虑到高性能计算机集群往往采用多级层次体系结构的事实，通过定义与之相匹配的层次结构I/O软件协议栈，并在协议栈不同层次引入相应的I/O功能与属性，构建易于维护和扩展的高性能存储系统。而且，通过研究实现部署在计算机集群的存储系统服务器端的自底而上的I/O优化机制（包括基于磁盘块访问模式的推送式数据预取机制和基于磁盘块访问信息的计算任务动态迁移机制等），减少计算结点参与I/O优化操作的力度进而加速应用程序执行并大幅提升存储系统输入/输出性能，最终解决或者缓解I/O性能与计算性能不匹配的问题。

This project intends to study a parallel file system, which enables a multi-layer I/O stack for providing high performance, scalable and low latency I/O service. To be specific, considering the fact of multi-level hierarchy in high performance clusters, we work on designing and implementing such a parallel file system that has a suited multi-layer I/O stack in which each layer has relevant I/O functionalities. Consequently, it is able to afford extendable, low latency and high performance I/O service. Furthermore, the project aims to research on server-pushing I/O optimization strategies including storage node-side prefetching and computing migration on the basis of block I/O access patterns, and then the compute nodes can be significantly relieved from optimization-related tasks. As a result, not only the execution time of applications running on the compute nodes can be cut down, but also high performance I/O services can be offered by the storage system.