一种低开销的面向节点内互连的网络接口控制器  

A Low Overhead Intra-Node Interconnection Oriented Network Interface Controller

作　　者：苏勇[1,2] 曹政[2] 刘飞龙[1,2] 王展[1,2] 刘小丽[2] 安学军[2] 孙凝晖[2] 

机构地区：[1]中国科学院大学计算机与控制工程学院,北京100190 [2]中国科学院计算技术研究所计算机体系结构国家重点实验室,北京100190

出　　处：《计算机学报》2015年第5期909-922,共14页Chinese Journal of Computers

基　　金：国家自然科学基金(61100014)资助~~

摘　　要：高性能计算和云计算的飞速发展对高性能互连网络的设计提出了越来越高的要求:除了要保证高带宽、低延迟和高可靠性等特性,还要面临成本和系统规模的挑战.该文针对这些特性和挑战提出了一种低开销的基于cHPP体系结构的超节点网络接口控制器:(1)设计了兼容PCIe的网络通信协议,降低协议转换开销、减少通信延迟并增强系统可扩展性能;(2)采用PCIe高速通信接口并支持用户级通信提高软硬件交互效率,面向MPI编程模型抽象出高效通信原语(如NAP、PUT和GET)加速大数据传输;(3)硬件支持I/O虚拟化实现超节点内对网络接口控制器的高效共享.为了对该文的设计进行功能和性能验证,文章基于FPGA实现了系统原型,实验结果显示最低延迟为1.242μs,有效数据带宽可达3.19GB/s.High performance computing and Cloud Computing are requiring the interconnection network：not only high bandwidth,low latency and high reliability,but also low cost and high scalability.To address these challenges,a low overhead hyper-node network interface controller（NIC）for cHPP architecture is proposed：（1）a PCIe-compliant network communication protocol is proposed to decrease the overhead of protocol conversion and the user-level communication is supported to decrease the software latency;（2）high efficient communication primitives such as NAP,PUT and GET are proposed and implemented to accelerate the MPI programming model and large data exchanging;（3）Hardware-based I/O virtualization is supported to enable the efficient NIC sharing in the hyper-node.The proposed hyper-node network interface controller design has been verified based on a FPGA prototype.Evaluations carried out on the prototype show that the lowest latency is 1.242μs and the highest bandwidth is 3.19GB/s.

关 键 词：互连 网络接口控制器 直接存储器访问 PCI EXPRESS I/O虚拟化 

分 类 号：TP303[自动化与计算机技术—计算机系统结构]