Arcadia: End-to-end AI System Performance Simulator
summary
Arcadia simulates compute, memory and network together to predict AI cluster behavior and outputs training and inference metrics, utilization and reliability numbers. It models jobs with a job generator, failure generator and a workload synthesizer trained by clustering and generative models. Three fidelity modes trade accuracy for speed: packet mode (about a day to simulate a day), flow mode (≈1000× faster), and job mode (full-year runs in minutes). It scales to tens of thousands of GPUs and millions of flows in seconds and shows fragmentation rises with larger job sizes while bandwidth and flow entropy give diminishing performance returns.
AI-written summary · about
chapters (4)
related talks
19mEnabling AI Network Cross-Layer Design and Operations with Arcadia...
19mScaling AI Network with DSF
09mAI for NI
21mDC Networks for Generative AI
21mDesigning Scalable Networks for Large AI Clusters: Challenges and Key Insights
22mResearch SuperCluster(RSC): Accelerating AI Research at Scale
26mAI Traffic Is Different: Rethinking Proxy Arc...
11mAI-Native Network Operation