~/talk

Enabling AI Network Cross-Layer Design and Operations with Arcadia...

Talkat NSDI '26 · 2026-05-04..05-06· [machine-learning, networking]
links:none on file
name:Xu Zhang
links:none on file
links:none on file
links:none on file
links:none on file
links:none on file
name:Xin Liu
linkedin:fbxin
links:none on file
links:none on file
links:none on file
name:Yan Cai
links:none on file
links:none on file
links:none on file
links:none on file
slides:none on file

Arcadia is a calibrated, multi‑fidelity discrete‑event simulator that replays production traces to answer counterfactuals about schedulers, transports, hardware, and operations. It supports packet/flow/packing models, runs a 4,000‑host, eight‑H100‑per‑host evaluation, and reports about 5% error (packing) and ~58.8x speedup vs. AstraSim. It fast‑forwards stable iterations, lets teams plug in custom modules, models PFC/ECN/queues for buffer and QP experiments, and has guided 50+ production decisions over three years.

AI-written summary · about

  1. 0:00Arcadia AI network simulator
  2. 1:01AI infrastructure reasoning
  3. 3:46AI cluster design challenges
  4. 5:32Production testing limits
  5. 6:55Simulator requirements
  6. 8:39Arcadia meets requirements
  7. 10:03Design principles
  8. 11:49Multi-fidelity modeling
  9. 13:15Continuous calibration
  10. 14:20Accuracy and performance
  11. 15:25Hardware what-if analysis
  12. 16:32QP scaling evaluation
  13. 17:39Future directions
  14. 18:20Key takeaways
~/search
type to search 6,249 conferences, 138,987 speakers and 185,172 talks
~/help
cmdctrl k
search every conference, speaker and series
/
focus the filter search
h
home
c
conferences
s
speakers
e
series
t
light / dark — the status bar button does the same
?
this help
esc
close this, then clear every filter — or just the search, while you are typing in it

Ignored while you are typing. Pages without a filter rail have no search to focus.