基于Myrinet/GM的多通道通信

Multi-Networking Communication Based on Myrinet/GM

通信子系统对并行系统的计算效率有重要影响,大规模应用对并行平台的通信性能和可用性提出了挑战性的要求.多通道通信技术通过并行采用多路网络链路互连来提高并行系统通信性能和可用性.首先分析了多进程复用网络对通信性能的影响,然后以Myrinet/GM网络平台为基础,提出了基于网络接口层的通信链路动态选择与分配策略,设计和实现了支持多路Myrinet网络并行通信的协议层MNC.MNC支持通信进程平等,充分地利用多路Myrinet网络链路资源.在使用2路Myrinet互连的PC机群平台上,MNC进程间通信带宽相对于单链路提高了约34%,有效地提高了应用层通信性能.

Communication subsystem is crucial for cluster computing, which affects its efficiency, adaptability and scalability. Large-Scale applications require challenging communication performance and availability from cluster systems. Multi-Networking communication is a novel approach to improve the communication performance and availability by using multiple network links in parallel. In this paper, the effect of multi-process multiplexing one network link is analyzed, a dynamic link dispatch scheme is proposed, and the design and implementation of a multi-networking communication layer, MNC is introduced, which extends GM messaging layer, and supports multi-Myrinet parallel communication. MNC provides multi-process effectively exploiting the raw performance of multi-Myrinet, and improves the communication performance of application layer significantly. Compared with one-way Myrinet/GM environment, the communication bandwidth between MNC processes has increased by 34% on the PC cluster interconnected with 2-way Myrinet.