1 总体互联架构
2 跨机内存之间映射
3 RDMA数据传输机制设计
4 兼容性消息队列设计
表1 内嵌消息大小和偏移量Tab.1 Inline message size and offset |
| 消息大小 | 偏移量 | 成员 | 复制大小 |
|---|---|---|---|
| 1-4 | 112 | data 4 | 16 |
| 5-52 | 64 | data 52 | 64 |
| 53-116 | 0 | data 116 | 128 |
|
张桉齐(1992—),男,工程师,研究方向为指挥控制系统相关设计。 |
|
张利彬(1982—),男,工程师。 |
Copy editor: 许韦韦
收稿日期: 2024-07-11
修回日期: 2024-09-22
网络出版日期: 2025-05-28
A high-speed interconnection mechanism based on multi-hosts memory mapping
Received date: 2024-07-11
Revised date: 2024-09-22
Online published: 2025-05-28
针对指挥控制、仿真推演高性能集群系统对按需数据共享、实时数据流传输等能力的需求,利用原生的PCIe通道,结合提出的单向消息队列,设计了一种基于跨机内存映射的高速互联机制。该互联机制设计了三种不同的协议来传输数据,包括内嵌协议、短协议和长协议。结合PCIe跨主机内存段映射方法,实现了系统节点之间的内存共享、RDMA等功能,具备低开销、低延迟、高带宽等特点,可支持指挥控制、仿真推演系统中批量数据在局域节点之间的共享与流动。通过功能性能实验,对设计的局域高速互联机制进行了测试,测试结果表明提出的互联机制具备高速数据传输能力,PCIe X16情况下RDMA传输速率可达到9 000 MB/s,同时可以有效降低数据传输延迟。
张桉齐 , 张利彬 . 基于跨机内存映射的高速互联机制[J]. 指挥控制与仿真, 2025 , 47(3) : 109 -115 . DOI: 10.3969/j.issn.1673-3819.2025.03.015
In order to meet the requirements of on-demand data sharing and real-time data stream transmission in Command Control and Simulation high-performance cluster systems, a high-speed interconnection mechanism based on Multi-hosts memory mapping is designed by using the native PCIe channel and the proposed one-way message queue. The interconnection mechanism provides three different protocols for transferring data, including inline, short, and long protocols. Combined with the PCIe Multi-hosts memory segment mapping method, it implements the functions of memory sharing, reflection memory, RDMA and other functions between the nodes. The proposal has the characteristics of low overhead, low latency, high bandwidth, etc., and can support the sharing and flow of batch data between local nodes in Command Control and Simulation systems. The final test results show that the proposed interconnection mechanism has high-speed data transmission capability, and the RDMA transmission rate can reach 9 000 MB/s with PCIe X16 lanes, and the data transmission delay can be effectively reduced.
Key words: PCIe; computing cluster; NTB; RDMA
表1 内嵌消息大小和偏移量Tab.1 Inline message size and offset |
| 消息大小 | 偏移量 | 成员 | 复制大小 |
|---|---|---|---|
| 1-4 | 112 | data 4 | 16 |
| 5-52 | 64 | data 52 | 64 |
| 53-116 | 0 | data 116 | 128 |
| [1] |
孙长秋, 杜长斌, 李欣宇, 等. 智算中心关键技术研究[J]. 通信管理与技术, 2024(2): 33-37,52.
|
| [2] |
李仁刚, 王彦伟, 等. Direct xPU:一种新型节点间通信优化的分布式异构计算架构[J]. 计算机研究与发展, 2024, 61(6):1388-1 400.
|
| [3] |
|
| [4] |
|
| [5] |
梁晨, 陈昊, 杨舟, 等. 机载嵌入式超算平台RDMA技术研究及实现[J]. 航空电子技术, 2023, 54(4): 36-42.
|
| [6] |
|
| [7] |
|
| [8] |
|
| [9] |
翁创. 一种面向虚拟化环境的Linux TCP/IP流程优化方法[J]. 软件导刊, 2024, 23(1): 81-89.
|
| [10] |
张众钦. 基于InfiniBand的高性能计算机技术研究[J]. 信息记录材料, 2021, 22(11):231-232.
|
| [11] |
徐德发. 基于HPC的RoCE网络和InfiniBand网络性能对比[J]. 无线互联科技, 2024, 21(9):111-116.
|
/
| 〈 |
|
〉 |