Enabling AI Network Cross-Layer Design and Operations with Arcadia: A Simulation Platform at Scale
Abstract
The rapid evolution of Artificial Intelligence (AI) technology is fueling significant investments by hyperscalers, making AI networks crucial for large-scale training. Understanding the design impacts on AI training requires systematic, cross-layer evaluation. Production experience highlights the need for a robust simulation platform to guide network design and operations. This paper defines the platform requirements, addresses complex design challenges, and shares our experience building Arcadia, a scalable, high-fidelity simulation platform for AI Networks. It operates at the cluster level, focusing on overall cluster performance rather than individual job performance. By using our fast-forwarding, lock free, and synchronization-cost reduction mechanisms, Arcadia achieves scalability and speed, allowing us to faithfully simulate real-world-scale training clusters and plays an important role in guiding Meta's AI network evolution.