~/talk

Supercharging Packet-level Network Simulation of Large Model Training via Memoization...

Talkat NSDI '26 · 2026-05-04..05-06· [networking, machine-learning]
name:Fei Long
links:none on file
links:none on file
name:Li Chen
links:none on file
name:Dan Li
links:none on file
links:none on file
name:Fei Gui
speakerdeck:shawnshanksgui
links:none on file
links:none on file
links:none on file
slides:none on file

Wormhole implements a C++ ns-3 kernel that uses network partitioning, state memoization and rate-based steady-state detection to skip repetitive packet events and replay cached unsteady transitions via a Flow Contention Graph. It reports 227–745× speedups for GPT workloads (510× for MoE), >1,000× peak when paired with Unison, average FCT error <1%, steady-state skip >99% of events, and database storage below megabytes, demonstrating large practical acceleration with bounded error.

AI-written summary · about

  1. 0:00Wormhole introduction
  2. 3:30Limitations of simulations
  3. 4:35Steady states and flow patterns
  4. 6:43Wormhole architecture
  5. 8:11Simulation memoization
  6. 9:36Steady state fast-forwarding
  7. 11:21Wormhole acceleration example
  8. 13:10Implementation details
  9. 13:30Evaluation results
  10. 15:36Summary and future work
  11. 16:18Q&A
~/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.