Treni

Routing Comparison

Internal vs external routing benchmark results from Plan v2 Track B.

Overview

Plan v2 Track B measures whether in-process routing (monolith runtime) is faster than external routing (controller + remote tool/model endpoints) on the same hardware.

Date: 2026-02-17 to 2026-02-19 (UTC) Hardware: AWS g5.xlarge (NVIDIA A10G)

Headline

MetricInternalExternalExternal/Internal
Mean latency94.849 ms97.927 ms1.032x

Internal routing is faster. Ratio > 1 means internal wins.

Stage Timing

StageValue
Internal route mean23.380 ms
Internal infer mean68.286 ms
Internal TTFT mean53.425 ms
External controller route mean0.003 ms
External tool hop mean2.206 ms
External model hop mean94.859 ms

Per-Task Breakdown

TaskInternalExternal
general_short150.767 ms152.274 ms
receipt_extract80.732 ms81.270 ms
search_grounded46.945 ms57.237 ms
summarize_short100.950 ms100.928 ms

Integrity

  • Errors: top-level 0, warmup 0, internal 0, external 0
  • Warmup ordering bias from earlier run was corrected in this final comparison.

Matrix Expansion (2026-02-19)

Track B was expanded to a 6-profile matrix (baseline + escalating timeout/failure stress):

ProfileExt/Int Latency RatioInt ErrorExt Error
p00 baseline1.042x0.00000.0000
p01 fail mild1.048x0.00000.0000
p02 timeout mild1.142x0.00000.0000
p03 mixed moderate1.164x0.00000.0417
p04 mixed aggressive1.436x0.00000.0833
p05 mixed aggressive + retry21.416x0.00000.0833

Matrix interpretation:

  • Internal path remains faster in all profiles (ratio > 1).
  • External path degrades progressively under timeout/failure pressure.
  • Internal error rate stayed zero across matrix profiles.

Raw Artifacts

On this page