首页|基于HPC的RoCE网络和InfiniBand网络性能对比

基于HPC的RoCE网络和InfiniBand网络性能对比

扫码查看
高性能计算(High Performance Computing,HPC)系统的通信接口一直在不断发展,然而对新一代的无限带宽(InfiniBand,IB)互联与RoCE 互联在性能上的对比缺乏研究.文章从消息传递接口(Message Passing Interface,MPI)基准程序、HPC基准程序以及HPC应用程序等多个方面,提出了一套实验方法对IB互联和RoCE互联进行性能评估.结果表明,在拥塞条件下RoCE 相较于IB有更好的网络性能;在MPI场景下,RoCE网络的通信性能略高于 IB 网络;在HPC基准程序测试中,RoCE方案相较于IB方案浮点性能略高但综合性能相当;在不同的终端应用场景下,RoCE方案和IB方案的性能则各有高低.
Performance comparison of the RoCE networks and InfiniBand networks based on HPC
The communication interface of high performance computing(HPC)systems is constantly evolving,but there is a lack of research on the performance comparison between the new generation of IB interconnect and RoCE interconnect.This paper presents a set of experimental methods for performance evaluation from MPI benchmark,HPC benchmark and HPC application.The conclusion shows that RoCE has better network performance than IB under congestion conditions,and the communication performance of RoCE network in MPI scenario is slightly higher than that of IB network.In HPC benchmark test,RoCE scheme has slightly higher floating point performance than IB scheme,but the overall performance is comparable.The performance of the RoCE and IB schemes varies according to the application scenarios.

RoCEInfiniBandHPCperformance

徐德发

展开 >

上海超级计算中心,上海 201023

RoCE InfiniBand HPC 性能

2024

无线互联科技
江苏省科学技术情报研究所

无线互联科技

影响因子:0.263
ISSN:1672-6944
年,卷(期):2024.21(9)