Different congestion onset in Rust vs Haskell

[bwbush]

Observations

• At 5 IB/s the Haskell and Rust simulations show very similar distributions of the time of an IB's generation to its arrival at all of the nodes.
• However, at 50 IB/s the Haskell simulation shows a much longer tail in arrival time. Do we have an explanation for this?

Evidence

Many other diagnostic graphs in analysis.ipynb confirm this behavior.

Context

• Tag: leios-2025w11
• Network topo-default-100
• Base configurations
  • With CPU timings on: config.with-cpu.yaml
  • With CPU timings off: config.without-cpu.yaml
• Variation on base configuration:

  ouroboros-leios/analysis/sims/2025w11/run-experiment.sh

  Lines 35 to 36 in f78f1dc

  	IB_RATES="5.0 50.0"
  	IB_SIZES="98304"

[Saizan]

@bwbush can you confirm which stage length these numbers are from? The graph says 60 but run-experiment.sh seems to show config.without-cpu.yaml is only run with stage length of 20?

[bwbush]

@bwbush can you confirm which stage length these numbers are from? The graph says 60 but run-experiment.sh seems to show config.without-cpu.yaml is only run with stage length of 20?

Sorry, the stage length was 20.

I don't think the stage length is an issue here because I see the similar congestion patterns in many of the other plots.

[bwbush]

Agreement is better at tag leios-2025w12. Is this close enough?

Note that turning of the TCP optimizations in Haskell results in worse agreement: