Reproduce “ScaleAcross: Designing Multi-Data-Center Infrastructure for Geo-Distributed AI Training” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Naved Inam, Aryan Alpesh Bhavsar, Masabattula Teja Nikhil, Sidharth Sharma
An investigation into EVPN-VXLAN for geo-distributed AI training infrastructure demonstrates improved traffic distribution with a queue-pair-aware ECMP mechanism, achieving a 13.7% better load factor, and accelerated failure recovery to 110 ms using BFD. The work also provides an open emulation framework to characterize AI workload behavior across data centers.
Key insights
- EVPN-VXLAN provides a scalable and robust foundation for geo-distributed AI training, enabling multi-tenancy and logical isolation across data centers.
- Enhancing ECMP with queue-pair-aware source-port allocation can effectively overcome hash collisions in low-entropy AI traffic, leading to better load balancing.
- Rapid fault detection via BFD integration is critical for maintaining distributed AI training continuity, reducing recovery times from minutes to milliseconds.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.