首页|期刊导航|天地一体化信息网络|面向星上大语言模型训练的语义驱动RDMA协议

面向星上大语言模型训练的语义驱动RDMA协议OA

Semantic-driven RDMA protocol for on-orbit LLM training

中文摘要英文摘要

随着大语言模型训练规模持续增长,业界开始探索将部分训练任务部署到轨道空间,依托空间环境中可持续的太阳能供给发展绿色计算,缓解地面训练的能耗压力.然而,多星分布式训练要求星间网络为节点间频繁的同步通信提供高吞吐量、低尾时延以及稳定的传输服务,而星间链路易受环境扰动、丢包现象频发,且星间拓扑随轨道运动周期性动态变化,使传统星间传输协议和地面远程直接内存访问(RDMA)协议均难以直接适配在轨训练场景.为此,提出一种面向星上大语言模型训练的语义驱动RDMA协议——StarRDMA协议.StarRDMA协议充分利用训练通信中各类消息可靠性需求与时效性要求存在显著差异的特征,针对性设计了两大核心机制:一是语义感知的可靠传输服务,协议依据不同训练消息在可靠性需求上的差异,融合选择性纠删编码、选择性确认(SACK)丢包恢复和截止时间感知重传机制,降低高往返时延(RTT)环境下的丢包恢复开销;二是差异化的拥塞控制,结合轨道RTT预测与语义优先级感知的加权加性增大乘性减小(AIMD),使关键路径流量获得更多的带宽分配优势,缓解长RTT流的竞争劣势.基于ns-3仿真平台对StarRDMA协议的性能展开初步评估.仿真结果表明,StarRDMA在高丢包场景下可将AlltoAll、AllReduce集合通信完成时间最高降低72.83%;在后台大流混合共存场景下,StarRDMA差异化拥塞控制可将集合通信完成时间最高降低35.71%.

As large language model(LLM)training scales up continuously,there is growing research interest in offloading partial train-ing workloads to orbital space,this paradigm leverages the sustainable solar energy supply in space to develop green computing and al-leviate the energy consumption pressure of ground-based training.However,multi-satellite distributed training imposes stringent de-mands on inter-satellite networks,which must provide high throughput,low tail latency and stable transmission services for frequent synchronization communication among satellite nodes.Nevertheless,inter-satellite links are susceptible to environmental disturbances and prone to packet loss.Meanwhile,the inter-satellite topology changes periodically with orbital motion,rendering both traditional inter-satellite transport protocols and terrestrial RDMA protocols incapable of directly adapting to on-orbit training scenarios.To this end,we propose StarRDMA protocol,a semantic-driven RDMA protocol for on-orbit LLM training.StarRDMA fully exploits the sig-nificant differences in reliability and timeliness requirements among various training communication messages,and accordingly designs two core mechanisms.Firstly,it provides a semantic-aware reliable transmission service,according to the differentiated reliability de-mands of different training messages,the protocol integrates selective erasure coding,SACK-based packet loss recovery,and deadline-aware retransmission,so as to reduce packet loss recovery overhead in high-RTT environments.Secondly,it introduces a differentiated congestion control mechanism.Combining orbit-based RTT prediction with semantic-priority-aware weighted AIMD,the mechanism enables critical-path traffic to gain superior bandwidth allocation,and mitigates the inherent competitive disadvantage of long-RTT flows.This paper presents a preliminary evaluation of StarRDMA's performance using the ns-3 simulation platform.Simulation results show that StarRDMA reduces the completion time of AlltoAll and AllReduce collective communications by up to 72.83%under high packet loss scenarios.In the presence of coexist background with large flows,StarRDMA's differentiated congestion control reduces the collective communication completion time by up to 35.71%.

黄俊凯;潘恬;杨祎凌;张娇;黄韬

北京邮电大学网络与交换技术全国重点实验室,北京 100876北京邮电大学网络与交换技术全国重点实验室,北京 100876北京邮电大学网络与交换技术全国重点实验室,北京 100876北京邮电大学网络与交换技术全国重点实验室,北京 100876北京邮电大学网络与交换技术全国重点实验室,北京 100876

信息技术与安全科学

太空计算大语言模型训练传输层协议RDMA协议

space computingLLM trainingtransport layer protocolRDMA Protocol

《天地一体化信息网络》 2026 (3)

50-66,17

国家自然科学基金面上项目(No.62372053) The National Natural Science Foundation of China(No.62372053)

10.11959/j.issn.2096-8930.tdyt2026028

评论