paperbot · PL 论文追踪

全部论文 待读 收藏

P2 Vela: A Virtualized LLM Training System with GPU Direct RoCE

Vela：基于 GPU Direct RoCE 的虚拟化 LLM 训练系统
ASPLOS Proceedings of the 30th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 22025匹配度 1/10 · 论文聚焦基于 KVM 虚拟化、虚拟化 RoCE 与 GPU Direct DMA 的云原生 LLM 训练系统，属于机器学习系统与云基础设施方向；与用户关注的静态分析、Agentic 程序分析、浮点程序分析、贝叶斯程序分析及 LLM 与 PL 的交叉均无直接关联。引用 5 inbox
Apoorve Mohan, Robert Walkup, Bengi Karacali, Ming-hung Chen, Abdullah Kayi, Liran Schour, Shweta Salaria, Sophia Wen, I-hsin Chung, Abdul Alim, Constantinos Evangelinos, Lixiang Luo, Marc Dombrowa, Laurent Schares, Ali Sydney, Pavlos Maniotis, Sandhya Koteshwara, Brent Tang, Joel Belog, Rei Odaira, Vasily Tarasov, Eran Gampel, Drew Thorstensen, Talia Gershon, Seetharami Seelam
机器学习与程序语言运行时与内存管理
crossref: 2026-09-12dblp: 2026-09-12

AI 速览

一句话：提出并评估了基于 KVM 虚拟化、虚拟化 RoCE 与 GPU Direct DMA 的云原生 LLM 训练系统 Vela。

Vela 是一个面向 LLM 训练的云原生系统，目标是在通用硬件、Linux KVM 虚拟化和虚拟化 RoCE 网络之上提供可扩展的训练基础设施。其虚拟机支持 GPU 与基于 SRIOV 的网卡之间的点对点 DMA，从而在虚拟化环境中保持高吞吐通信。论文介绍了在 IBM Cloud 中基于 NVIDIA A100 的部署，并总结了约 2.5 年的设计、构建与运维经验。在约 1500 GPU 规模上训练 50B 参数解码器模型时，系统达到理想吞吐的约 80%，单 GPU FLOPS 约为单虚拟机 HPL 基准的 70%。

问题
LLM 训练需要高性能 GPU 间通信，但云环境通常依赖虚拟化，如何用通用软硬件构建可访问且高效的训练基础设施是核心问题。
方法
Vela 采用 Linux KVM 虚拟机、SRIOV 网卡与虚拟化 RoCE，并支持 GPU 与网卡间的点对点 DMA，在 IBM Cloud 的 A100 集群上部署和运行。
评估
在约 1500 GPU 规模上训练 50B 参数解码器模型，使用模型并行；测得约 80% 的理想吞吐，以及相对单 VM HPL 约 70% 的单 GPU FLOPS。
原文摘要（Abstract）

Vela is a cloud-native system designed for LLM training workloads built using off-the-shelf hardware, Linux KVM-based virtualization, and a virtualized RDMA over Converged Ethernet (RoCE) network. Vela virtual machines (VMs) support peer-to-peer DMA between the GPUs and SRIOV-based network interface. In this paper, we share Vela's key architectural aspects with details from an NVIDIA A100 GPU-based deployment in one of the IBM Cloud data centers. Throughout the paper, we share insights and experiences from designing, building, and operating the system over a ~2.5 year timeframe to highlight the capabilities of readily available software and hardware technologies and the improvement opportunities for future AI systems, thereby making AI infrastructure more accessible to a broader community. As we evaluated the system for performance at ~1500 GPU scale, we achieved ~80% of the ideal throughput while training a 50 billion parameter decoder model using model parallelism, and ~70% per GPU FLOPS compared to a single VM with the High-Performance Linpack benchmark.

链接与引用

DOI 原文 · PDF（开放获取） · DBLP

BibTeX 
@inproceedings{MohanWKCKSSWCAE25,
  title = {Vela: A Virtualized LLM Training System with GPU Direct RoCE},
  author = {Apoorve Mohan and Robert Walkup and Bengi Karacali and Ming-hung Chen and Abdullah Kayi and Liran Schour and Shweta Salaria and Sophia Wen and I-hsin Chung and Abdul Alim and Constantinos Evangelinos and Lixiang Luo and Marc Dombrowa and Laurent Schares and Ali Sydney and Pavlos Maniotis and Sandhya Koteshwara and Brent Tang and Joel Belog and Rei Odaira and Vasily Tarasov and Eran Gampel and Drew Thorstensen and Talia Gershon and Seetharami Seelam},
  booktitle = {ASPLOS},
  year = {2025},
  doi = {10.1145/3676641.3716280}
}