Vela: A Virtualized LLM Training System with GPU Direct RoCE

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.

Authors

Institutions

Publication Details

Published
2025-03-27
DOI
https://doi.org/10.1145/3676641.3716280
Citations
2
Primary Topic
Cloud Computing and Resource Management
Type
article
Field-Weighted Citation Impact
6.49
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
article

Vela: A Virtualized LLM Training System with GPU Direct RoCE

Sandhya Koteshwara, Talia Gershon, M.B. Dombrowa, Liran Schour et al.
2 citations
Cloud Computing and Resource Management
6.49
article

Vela: A Virtualized LLM Training System with GPU Direct RoCE

Sandhya Koteshwara, Talia Gershon, M.B. Dombrowa, Liran Schour, Seetharami Seelam, Sophia Wen, Shweta Salaria, Abdullah Kayi, Vasily Tarasov, Rei Odaira, Pavlos Maniotis, R. E. Walkup, Md. Abdul Alim, Laurent Schares, Apoorve Mohan, Ming-Hung Chen, Constantinos Evangelinos, Eran Gampel, Ali Sydney, Drew Thorstensen, Joel Belog, Lixiang Luo, I‐Hsin Chung, Bengi Karaçali, Brent Tang
article en
2 citations

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.

University of Massachusetts Lowell (US), IBM (United States) (US), IBM Research - Almaden (US), Cambridge Scientific (United States) (US), Cloud Pharmaceuticals (United States) (US), IBM Research - Haifa (IL)
Affordable and clean energy
Openalex Percentile: Top 4%
Cloud Computing and Resource Management
6.49
DOISourceOpenAlex
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.