一种面向不可靠网络的快速RDMA通信方法

发布时间：2018-04-19 06:21

本文选题：远程内存访问 + RDMA　；参考：《湖南大学学报(自然科学版)》2015年08期

【摘要】：大数据量的远程内存访问(RDMA)传输是并行计算机中最基本的通信模式之一,对系统整体性能的影响很大.随着并行计算机系统的规模扩大,系统的容错性设计面临着很大的挑战,互连网络具有链路不可靠、自适应路由等特点,如何面向不可靠网络实现可靠的端到端RDMA传输是并行系统体系结构设计的一大难题.提出一种面向不可靠网络下的快速RDMA传输方法,方法能够在节点控制器芯片上高效实现,对上层驱动软件和应用提供可靠的端到端RDMA传输服务.与传统的建立连接的方法相比,方法的硬件设计复杂度大大降低;方法另一优点是实现了按需重传,避免了传统方法中一次RDMA传输出现错误时,需要重传整个RDMA数据的开销,在相同的错误概率下,新方法的传输效率得到了很大的提升.
[Abstract]:Remote memory access (RDMA) transmission of large amount of data is one of the most basic communication modes in parallel computers, which has a great impact on the overall performance of the system.With the expansion of the scale of parallel computer system, the fault-tolerant design of the system is facing a great challenge. The interconnection network has the characteristics of unreliable links, adaptive routing and so on.How to realize reliable end-to-end RDMA transmission for unreliable networks is a difficult problem in the architecture design of parallel systems.A fast RDMA transmission method for unreliable networks is proposed, which can be efficiently implemented on the node controller chip and provides reliable end-to-end RDMA transmission services for upper layer driver software and applications.Compared with the traditional method, the hardware design complexity of the method is greatly reduced. Another advantage of the method is that it realizes on-demand retransmission and avoids the overhead of retransmitting the whole RDMA data when there is an error in one RDMA transmission in the traditional method.Under the same error probability, the transmission efficiency of the new method is greatly improved.
【作者单位】：国防科学技术大学计算机学院;
【基金】：国家自然科学基金资助项目(61202024,61202126)~~
【分类号】：TP338.6

【相似文献】