Shelemet’eva Y. V. Study of the remote direct memory access technology in high-performance system architectures

Abstract: The article describes main principles of the remote direct memory access technology (RDMA) in highperformance computing systems. The specific attention is given to the algorithms for data transmission to and from network. The author reviews standard algorithm for data transmission and RDMA algorithm on the example of TCP/IP-based networks. The author studies the most popular computational architectures supporting remote direct memory access, proves the effectiveness of RDMA for high-performance network architectures on the example of 10 Gigabit Ethernet. Using the theory of sets and calculations of processor’s work cycles during the processing of network instructions and data the author compares performance of the nodes of computing system transmitting data via standard algorithm and nodes that implement RDMA technology. The article explains the expedience of RDMA implementation for compute nodes of high-performance architectures. In modern computer architectures computing nodes are interconnected via network, so a considerable amount of data is transmitted through network. One of the methods of improving the efficiency of computing nodes and the network as a whole is to offload processor of network computations with the remote direct memory access technology (RDMA). RDMA-mode is a technology for transmitting network data without CPU, directly from the network adapter buffer to application buffers, thereby freeing the CPU from processing network data. In high-performance computing architectures, systems the amount of incoming and outgoing network traffic is very high and the processing of network data occupies a large amount of CPU time, so the proposed technology of remote direct memory access is highly relevant to offloading processor of a specific computing node and hence to improving the performance of the whole system . Compared with the standard data transmission in which the processor of a compute node system completely handles network data RDMA technology frees the CPU from network load and makes it possible to focus all production capacity to internal tasks increasing the speed and reliability of the whole system.

Keywords: