论文部分内容阅读
在高性能计算系统中,RDMA Read消息是基本的数据交换方式之一,在系统实现中起着重要作用.不同的消息实现机制在设计复杂度、设计开销、通信性能方面有着较大的差别.因此,针对RDMA Read机制的研究具有重要的现实意义.目前,针对RDMA Read实现机制的研究相对较少,而现有的两种RDMA Read实现机制均有着明显的不足.我们分别在三个方面对现有机制进行了深入地对比分析.设计复杂度方面,设计实现了基本的硬件实现流程,并在此基础上进行定性分析;设计开销、通信性能方面,采用基于模型的方法进行理论分析,并设计软件验证平台进行实验验证.根据分析结果,结合已有两种机制的实现方法和特点,提出了一种改进的RDMA Read实现机制.模拟结果表明,新提出的RDMA Read机制在设计复杂度、设计开销、通信性能方面均有着优秀且均衡的表现.
In high-performance computing system, RDMA Read message is one of the basic data exchange methods, and plays an important role in system implementation.Different message implementation mechanisms have great differences in design complexity, design overhead and communication performance. Therefore, research on RDMA Read mechanism is of great practical significance.At present, there are relatively few studies on RDMA Read mechanism and two existing RDMA Read mechanisms have obvious deficiencies.We have three aspects The existing mechanisms are compared and analyzed in depth.For the design complexity, the basic hardware implementation process is designed and implemented, and qualitative analysis is carried out on this basis.And the model-based method is used to carry out theoretical analysis in terms of design overhead and communication performance. A software verification platform is designed to verify the experimental results.An improved RDMA Read implementation mechanism is proposed based on the analysis results and the implementation methods and features of the two existing mechanisms.The simulation results show that the RDMA Read mechanism proposed in this paper is of great importance in design complexity, Design overhead, communication performance have excellent and balanced performance.