DeML OS - 2026-04-28

DeML OS Daily DeML OS 最新前沿分析 DeML OS デイリー

Explore Frontier

04.28

2026

Tue

📄

Paper

Cloudless-Training: A Framework to Improve Efficiency of Geo-Distributed ML Training https://arxiv.org/abs/2303.05330

Wenting Tan Geo-Distributed ML

Tan et al. proposed Cloudless-Training in their paper, using a two-layer architecture, elastic scheduling, and two new sync strategies (ASGD-GA and MA) to improve geo-distributed ML training efficiency, achieving up to 1.7x speedup and 24% cost reduction on Tencent Cloud.

Notes

Geo-distributed ML training faces challenges of missing elastic scheduling and WAN communication bottlenecks.
Cloudless-Training uses a two-layer architecture separating control and physical training planes.
Elastic scheduling adaptively deploys workflows based on cloud resource heterogeneity and dataset distribution.
Proposes ASGD-GA (async SGD with gradient accumulation) and MA (model averaging) sync strategies.
Implemented with OpenFaaS and validated on Tencent Cloud.
Reduces training cost by 9.2%-24.0% and achieves up to 1.7x speedup over baseline.

Collected by @icerdesign

DeML OS Q & A 问答

Deep Dive 💬

04.28

2026

Tue

😇

What problem does Cloudless-Training solve?

It addresses the lack of elastic scheduling and low WAN communication efficiency in geo-distributed ML training.

😎

😊

How does Cloudless-Training achieve elastic scheduling?

Via a two-layer architecture separating control and physical planes, adaptively deploying workflows based on cloud resource heterogeneity and dataset distribution.

😎

🤓

How does Cloudless-Training guarantee model correctness?

Through ASGD-GA and MA sync strategies, introducing gradient accumulation and model averaging in async training to ensure convergence and model quality.

😎

Prompted by @icerdesign