一种低开销的面向节点内互连的网络接口控制器

发布时间：2018-08-04 19:01

【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.
[Abstract]:The rapid development of high performance computing and cloud computing make the design of high performance interconnection network more and more demanding: besides ensuring high bandwidth, low delay and high reliability, it also faces the challenge of cost and system scale. This paper presents a low overhead supernode network interface controller based on cHPP architecture for these characteristics and challenges. (1) A network communication protocol compatible with PCIe is designed to reduce the overhead of protocol conversion. Reduce the communication delay and enhance the extensibility of the system. (2) using PCIe high-speed communication interface and supporting user-level communication to improve the efficiency of hardware and software interaction, MPI programming model abstracts efficient communication primitives (such as NAP-put and GET) to speed up big data transmission; (3) hardware supports I / O virtualization to implement efficient sharing of network interface controllers within supernodes. In order to verify the function and performance of the design, the prototype of the system is implemented based on FPGA. The experimental results show that the minimum delay is 1.242 渭 s and the effective data bandwidth can reach 3.19 GB / s.
【作者单位】：中国科学院大学计算机与控制工程学院;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;
【基金】：国家自然科学基金(61100014)资助~~
【分类号】：TP332

【参考文献】