面向E级计算的高突发带宽I/O缓冲存储系统关键技术研究

With the rapid development of high-performance computers， I/O storage system has been the major performance bottleneck of exascale computing systems. How to effectively utilize non-volatile memory to improve I/O performance has become a key issue in high performance computing. Accordingly, this project is to carry out researches on key technologies of high burst-bandwidth I/O buffer storage system for exascale computing. By breaking through a series of key techniques, such as novel parallel file system design，the integration and optimization on storage and network, and novel management for shared storage pool based on NVM, we arm to meet exascale computing requirements for the storage system on performance, reliability and extensibility, and ensure the storage system is not a performance bottleneck for the next-generation high-performance computers. Technologies proposed in this research are promising to be applied to the emerging exascale computing systems.

随着高性能计算机系统计算能力的迅速提升，存储系统已经成为E级计算机系统设计的一个关键环节，I/O性能将成为制约整个系统性能的一大关键因素。如何有效利用新型存储介质来提高I/O存储性能，已经成为当前高性能计算领域的重点关注问题。本项目拟开展面向E级计算的高突发带宽I/O缓冲存储系统关键技术研究，突破新型并行文件系统设计、存储与网络的融合和协同优化，以及新型NVM共享存储池一体化管理等关键技术，旨在满足E级计算对I/O存储系统在性能、可靠性、可扩展性等方面的需求，确保I/O存储系统不成为下一代高性能计算机的性能瓶颈。研究的成果将应用到未来的E级高性能计算系统中。

针对高性能计算和大数据处理在I/O访问延迟、并发I/O吞吐率和Burst带宽等方面的需求，本课题研究了面向E级计算的高突发带宽IO缓冲存储系统关键技术，主要包括：（1）针对当前大规模HPC存储系统高带宽低延迟的需求，提出了面向E级计算的高突发带宽I/O缓冲存储架构，通过在传统计算结点和全系统共享大容量存储池两个层次之间加入高突发带宽I/O缓冲存储层，实现了存储资源的柔性配置和适配多种应用场景；（2）提出了存储与网络的融合和协同优化技术，通过硬件的方式加速DMA和NVMeoF协议的执行，减少CPU处理开销，进一步提高存储网络性能；（3）提出了新型NVM共享存储池一体化管理技术，通过将所有NVMe SSD存储资源映射到一个存储池中，对存储设备进行高效管理和优化配置，实现I/O负载均衡、性能高可扩展；（4）构建了一个基于RDMA的NVMeoF网络存储协议卸载验证平台。