I／O Forwarding架构下的I／O优化技术研究  

Research on I/O optimizations of the I/O Forwarding Framework

作　　者：刘建[1] 周恩强[1] 

机构地区：[1]国防科学技术大学计算机学院,长沙410073

出　　处：《高性能计算技术》2013年第6期58-61,共4页

摘　　要：摘要磁盘I／O的发展速度永远赶不上CPU性能的提升，I/O常常成为数据处理的瓶颈。本文研究大规模超级并行计算系统基于I/O Forwarding架构的I／O优化技术，在I/O中转节点设计并实现了一个基于主存＋SSD的异构缓冲区，解决I/O中转节点入口网络带宽与出口网络带宽速率不匹配的矛盾。针对I/O中转节点多核多线程的特点对数据进行异步缓存、重组转发、并行压缩等，提高I/O中转节点到磁盘这段I/O路径的有效网络带宽。本文在天河平台实现I/O架构设计和测试，实验显示，通过I/O优化可以加速数据在网络和磁盘上的传输。The development of disk I/O speed never catches up with the improvement of CPU performance, the I/O regularly becomes the bottleneck of data processing. This paper researches on I/O optimizations of the I/O Forwarding framework in the massively parallel computing system, on the I/O Forwarding nodes we designed and implemented a hybrid buffer which makes use of both internal memory and SSD, to solve the contradiction on the I/O Forwarding node that network bandwidth rate does not match between the entrance and export. To better utilize the multi-core and multi-thread on the I/O Forwarding nodes, by using asynchronous cache, recombinant forwarding, parallel compression of data to improve the effective network bandwidth between the I/O Forwarding nodes and the disk. The realization of I/O architecture designed and tested in Tianhe platform, the experiment shows： the I/O optimizations can accelerate the data transmission both in the network and disk.

关 键 词：I/O优化 SSD 异步缓存 并行压缩 

分 类 号：TP315[自动化与计算机技术—计算机软件与理论]