BurstGPT Window Selection
This report focuses only on BurstGPT `without_fails` traces and proposes 5-minute windows for optimize/test overfitting checks.
Recommended Windows
| role | window_id | start_s | end_s | n | rps | input_mean | input_p50 | input_p95 | input_p99 | output_mean | output_p50 | output_p95 | output_p99 | total_tps | max_1s_arrivals | api_log_pct | gpt4_pct |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| similar_optimize | 33,970 | 10,191,000 | 10,191,300 | 2,230 | 7.433 | 208.462 | 200.000 | 214.000 | 222.000 | 8.636 | 7.000 | 7.000 | 7.000 | 1,613.8 | 41 | 99.552 | 0.179 |
| similar_test | 34,140 | 10,242,000 | 10,242,300 | 2,230 | 7.433 | 200.388 | 200.000 | 214.000 | 220.000 | 8.423 | 7.000 | 7.000 | 7.000 | 1,552.2 | 40 | 99.910 | 0.359 |
| contrast_a | 10,213 | 3,063,900 | 3,064,200 | 4,765 | 15.883 | 274.629 | 238.000 | 692.000 | 692.000 | 3.817 | 1.000 | 4.000 | 11.000 | 4,422.7 | 29 | 99.454 | 99.035 |
| contrast_b | 6,836 | 2,050,800 | 2,051,100 | 631 | 2.103 | 2,150.3 | 2,372.0 | 2,650.0 | 2,660.0 | 296.810 | 102.000 | 1,861.0 | 2,318.2 | 5,147.1 | 20 | 97.623 | 0.951 |
How To Use These
Use `similar_optimize` for tuning and `similar_test` for transfer validation. These two windows are intentionally very similar: same request count, similar input/output lengths, and similar 1-second burst pressure.
Use `contrast_a` and `contrast_b` as a distribution-shift stress test. `contrast_a` is high-RPS and very short-output; `contrast_b` is lower-RPS but much longer input/output.
Plots
