虚拟机环境下MPI/RDMA库的通信优化  

Optimizing MPI/RDMA Library in Virtual Machine Environments

作　　者：谢尚威 华蓓[1] XIE Shang-wei;HUA Bei(Department of Computer Science and Technology,University of Science and Technology of China,Hefei 230027,China)

机构地区：[1]中国科学技术大学计算机科学与技术学院,合肥230027

出　　处：《小型微型计算机系统》2021年第7期1505-1510,共6页Journal of Chinese Computer Systems

摘　　要：RDMA是在高性能计算中应用最广泛的网络通信技术,以高吞吐、低延迟、低CPU占用而著称.随着高性能计算向云环境迁移,如何在虚拟机环境下高效地利用RDMA通信成为一个研究热点.本文针对高性能应用最常使用的MPI/RDMA通信库,提出在虚拟机环境下的通信优化方案VMPI.VMPI采用连接虚拟化、控制路径与数据路径分离等技术来满足云环境的要求,通过在数据路径上实现对大消息的零拷贝传输来降低端到端通信延迟.实验表明,相比于传统的RDMA虚拟化实现,VMPI可将MPI的点对点通信延迟减少40%左右,并可降低应用基准测试程序Graph500的通信时间占比.RDMA is the most widely used networking technology in high-performance computing,known for its high throughput,low latency,and low CPU usage.As high-performance computing migrates to cloud environments,how to transfer data efficiently with RDMA in virtual machines has become a research hotspot.For the most commonly used MPI/RDMA communication library in high-performance computing,this paper proposes VMPI,an optimization scheme of MPI/RDMA library for its working in virtual machine environments.VMPI adopts technologies such as connection virtualization,separation of control path and data path to enable its use in the cloud environment,and reduce the end-to-end communication delay by realizing zero-copy transmission of large messages on the data path.Experiments show that compared with the traditional RDMA virtualization implementation,VMPI can reduce the communication delay of a point-to-oint MPI API by about 40%,and reduce the proportion of communication time to a total time on the Graph500 benchmark.

关 键 词：网络I/O虚拟化 远程直接内存访问 MPI通信优化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]