Qwen-Bailian Stable A/B/C Window Selection

Stable 10-minute A/B/C windows for the Qwen-Bailian trace families, exported as AIPerf replay JSONL files with 5-minute split halves.

Trace families

Exported windows

Window duration

5/10 min

Min source gap

600s

Selection rule: Trace A and Trace B exclude image/file requests but keep text-like/API records such as text, search, and api. Thinking and Coder use native workload labels. Selected 10-minute A/B/C windows require at least 600 seconds between selected source windows.

How These Windows Were Created

Each trace is scanned with 10-minute candidate windows. Candidates are grouped into low, mid, and high RPS regimes using per-trace RPS quantiles. The selected A/B/C combination minimizes the sum of stability scores while enforcing the source-time distance constraint.

half_ratio measures balance between the first and second 5-minute halves, bucket10_cv measures 10-second RPS variability, spike_ratio measures peak short-window RPS divided by average RPS, and stability_score combines imbalance and 10-second variability. Lower is steadier.

Dataset Selection Summary

Dataset	A/B/C windows	A/B/C RPS	Source gaps	Selection rule
Qwen Bailian Trace A	A 1200-1800, B 3600-4200, C 6000-6600	A 5.107 / B 5.265 / C 6.855	1800s, 1800s	Filtered to text-like/API request types before window selection and export: text, search, api when present. Image/file requests are excluded.
Qwen Bailian Trace B	A 600-1200, B 3000-3600, C 4800-5400	A 22.937 / B 24.843 / C 25.567	1800s, 1200s	Filtered to text-like/API request types before window selection and export: text, search, api when present. Image/file requests are excluded.
Qwen Bailian Thinking	A 6000-6600, B 2400-3000, C 4200-4800	A 0.993 / B 1.463 / C 2.197	1200s, 1200s	Uses native trace type labels; no additional filter applied.
Qwen Bailian Coder	A 2400-3000, B 4800-5400, C 6000-6600	A 5.885 / B 6.135 / C 6.610	1800s, 600s	Uses native trace type labels; no additional filter applied.

Qwen Bailian Trace A

Filtered to text-like/API request types before window selection and export: text, search, api when present. Image/file requests are excluded.

A RPS

5.107

B RPS

5.265

C RPS

6.855

Selected windows

1200-1800, 3600-4200, 6000-6600

Regime Mapping

regime	window	start_s	end_s	rps	use
low	A	1,200	1,800	5.107	low-load stable traffic
mid	B	3,600	4,200	5.265	mid-load stable traffic
high	C	6,000	6,600	6.855	high-load stable traffic

Selected 10-Minute Windows

window_id	rps	stability_score	half_ratio	bucket10_cv	spike_ratio
2	5.107	0.171	0.967	0.137	1.312
6	5.265	0.186	0.968	0.153	1.311
10	6.855	0.143	0.997	0.139	1.284

Segment Statistics

segment	start_s	end_s	n	rps	input_mean	input_p50	input_p95	output_mean	output_p50	output_p95	total_tps	rps_10s_cv	rps_10s_max	types
A	1,200	1,800	3,064	5.107	2,112.896	973.500	8,342.800	415.167	361.000	978.000	12,909.972	0.137	6.700	text:2463, search:601
A1	1,200	1,500	1,558	5.193	2,075.010	961.000	8,334.950	408.850	357.000	974.300	12,899.517	0.136	6.600	text:1258, search:300
A2	1,500	1,800	1,506	5.020	2,152.089	991.000	8,342.000	421.701	364.000	992.000	12,920.427	0.135	6.700	search:301, text:1205
B	3,600	4,200	3,159	5.265	1,981.393	908.000	8,109.200	427.009	372.000	982.800	12,680.233	0.153	6.900	text:2521, search:638
B1	3,600	3,900	1,554	5.180	2,010.293	882.000	8,087.450	428.813	372.500	996.350	12,634.573	0.162	6.900	text:1229, search:325
B2	3,900	4,200	1,605	5.350	1,953.411	927.000	8,120.600	425.261	372.000	977.200	12,725.893	0.143	6.900	text:1292, search:313
C	6,000	6,600	4,113	6.855	2,634.764	1,244.000	8,962.200	459.587	403.000	1,070.400	21,211.778	0.139	8.800	search:885, text:3228
C1	6,000	6,300	2,060	6.867	2,540.293	1,189.500	8,804.000	452.957	401.500	1,068.100	20,553.647	0.153	8.600	search:423, text:1637
C2	6,300	6,600	2,053	6.843	2,729.558	1,306.000	9,009.000	466.240	403.000	1,076.800	21,869.910	0.124	8.800	text:1591, search:462

Suggested Combinations

ID transfer: A1 -> A2, B1 -> B2, C1 -> C2
OOD transfer: A1 -> B2, B1 -> C2, C1 -> A2
Three-regime robustness: optimize on A1+B1+C1, test on A2+B2+C2.

Plots

Full trace 60s RPS with colored selected-window bands

Selected-window 10s RPS shape

A/B/C and split-half RPS

Token distributions by selected window

Candidate RPS vs stability

Qwen Bailian Trace B

Filtered to text-like/API request types before window selection and export: text, search, api when present. Image/file requests are excluded.

A RPS

22.937

B RPS

24.843

C RPS

25.567

Selected windows

600-1200, 3000-3600, 4800-5400

Regime Mapping

regime	window	start_s	end_s	rps	use
low	A	600	1,200	22.937	low-load stable traffic
mid	B	3,000	3,600	24.843	mid-load stable traffic
high	C	4,800	5,400	25.567	high-load stable traffic

Selected 10-Minute Windows

window_id	rps	stability_score	half_ratio	bucket10_cv	spike_ratio
1	22.937	0.318	0.868	0.176	1.360
5	24.843	0.282	0.867	0.139	1.373
8	25.567	0.209	0.983	0.191	1.447

Segment Statistics

segment	start_s	end_s	n	rps	input_mean	input_p50	input_p95	output_mean	output_p50	output_p95	total_tps	rps_10s_cv	rps_10s_max	types
A	600	1,200	13,762	22.937	939.866	672.500	2,403.000	83.730	40.000	271.950	23,477.882	0.176	31.200	api:12164, text:1598
A1	600	900	7,368	24.560	965.198	719.000	2,601.950	80.499	42.000	248.650	25,682.320	0.133	31.200	api:6555, text:813
A2	900	1,200	6,394	21.313	910.675	627.000	2,128.000	87.453	39.000	292.700	21,273.443	0.191	26.700	api:5609, text:785
B	3,000	3,600	14,906	24.843	916.032	615.000	2,414.750	93.261	44.000	327.000	25,074.195	0.139	34.100	api:12983, text:1923
B1	3,000	3,300	7,985	26.617	932.767	737.000	2,251.200	94.068	55.000	323.000	27,330.907	0.100	34.100	api:7017, text:968
B2	3,300	3,600	6,921	23.070	896.725	505.000	2,756.000	92.330	34.000	335.000	22,817.483	0.140	29.600	api:5966, text:955
C	4,800	5,400	15,340	25.567	947.655	552.500	2,805.000	97.562	39.000	375.050	26,722.723	0.191	37.000	api:13396, text:1944
C1	4,800	5,100	7,737	25.790	988.033	523.000	2,958.800	97.669	37.000	386.200	28,000.237	0.205	37.000	api:6736, text:1001
C2	5,100	5,400	7,603	25.343	906.566	573.000	2,361.200	97.454	41.000	362.900	25,445.210	0.176	36.200	api:6660, text:943

Suggested Combinations

ID transfer: A1 -> A2, B1 -> B2, C1 -> C2
OOD transfer: A1 -> B2, B1 -> C2, C1 -> A2
Three-regime robustness: optimize on A1+B1+C1, test on A2+B2+C2.

Plots

Full trace 60s RPS with colored selected-window bands

Selected-window 10s RPS shape

A/B/C and split-half RPS

Token distributions by selected window

Candidate RPS vs stability

Qwen Bailian Thinking

Uses native trace type labels; no additional filter applied.

A RPS

0.993

B RPS

1.463

C RPS

2.197

Selected windows

6000-6600, 2400-3000, 4200-4800

Regime Mapping

regime	window	start_s	end_s	rps	use
low	A	6,000	6,600	0.993	low-load stable traffic
mid	B	2,400	3,000	1.463	mid-load stable traffic
high	C	4,200	4,800	2.197	high-load stable traffic

Selected 10-Minute Windows

window_id	rps	stability_score	half_ratio	bucket10_cv	spike_ratio
10	0.993	0.383	0.954	0.336	2.114
4	1.463	0.356	0.960	0.315	1.913
7	2.197	0.329	0.930	0.256	1.730

Segment Statistics

segment	start_s	end_s	n	rps	input_mean	input_p50	input_p95	output_mean	output_p50	output_p95	total_tps	rps_10s_cv	rps_10s_max	types
A	6,000	6,600	596	0.993	4,127.362	1,610.000	21,103.250	3,306.544	1,939.500	11,025.250	7,384.347	0.336	2.100	thinking:596
A1	6,000	6,300	291	0.970	4,192.601	1,652.000	19,375.000	3,390.498	1,875.000	10,822.500	7,355.607	0.292	1.800	thinking:291
A2	6,300	6,600	305	1.017	4,065.118	1,607.000	21,737.800	3,226.443	2,042.000	11,127.000	7,413.087	0.371	2.100	thinking:305
B	2,400	3,000	878	1.463	4,511.036	3,683.000	17,115.800	3,427.131	1,729.500	10,723.850	11,616.185	0.315	2.800	thinking:878
B1	2,400	2,700	430	1.433	4,813.114	3,683.000	20,287.300	3,534.984	1,738.500	11,891.650	11,965.607	0.349	2.800	thinking:430
B2	2,700	3,000	448	1.493	4,221.096	3,683.000	12,517.450	3,323.612	1,726.500	9,850.700	11,266.763	0.278	2.400	thinking:448
C	4,200	4,800	1,318	2.197	5,525.735	3,684.000	14,296.000	3,342.765	1,617.000	11,903.150	19,481.138	0.256	3.800	thinking:1318
C1	4,200	4,500	635	2.117	5,239.302	3,683.000	14,297.300	3,260.630	1,659.000	11,903.300	17,991.523	0.234	3.400	thinking:635
C2	4,500	4,800	683	2.277	5,792.038	3,685.000	14,295.000	3,419.127	1,596.000	11,771.600	20,970.753	0.269	3.800	thinking:683

Suggested Combinations

ID transfer: A1 -> A2, B1 -> B2, C1 -> C2
OOD transfer: A1 -> B2, B1 -> C2, C1 -> A2
Three-regime robustness: optimize on A1+B1+C1, test on A2+B2+C2.

Plots

Full trace 60s RPS with colored selected-window bands

Selected-window 10s RPS shape

A/B/C and split-half RPS

Token distributions by selected window

Candidate RPS vs stability

Qwen Bailian Coder

Uses native trace type labels; no additional filter applied.

A RPS

5.885

B RPS

6.135

C RPS

6.610

Selected windows

2400-3000, 4800-5400, 6000-6600

Regime Mapping

regime	window	start_s	end_s	rps	use
low	A	2,400	3,000	5.885	low-load stable traffic
mid	B	4,800	5,400	6.135	mid-load stable traffic
high	C	6,000	6,600	6.610	high-load stable traffic

Selected 10-Minute Windows

window_id	rps	stability_score	half_ratio	bucket10_cv	spike_ratio
4	5.885	0.133	0.995	0.128	1.359
8	6.135	0.162	0.972	0.133	1.467
10	6.610	0.152	0.991	0.143	1.316

Segment Statistics

segment	start_s	end_s	n	rps	input_mean	input_p50	input_p95	output_mean	output_p50	output_p95	total_tps	rps_10s_cv	rps_10s_max	types
A	2,400	3,000	3,531	5.885	6,114.450	5,127.000	13,514.500	771.109	445.000	2,472.500	40,521.518	0.128	8.000	coder:3531
A1	2,400	2,700	1,770	5.900	6,180.868	5,421.500	13,513.550	800.294	454.500	2,556.050	41,188.857	0.129	7.500	coder:1770
A2	2,700	3,000	1,761	5.870	6,047.693	4,851.000	13,515.000	741.776	437.000	2,329.000	39,854.180	0.128	8.000	coder:1761
B	4,800	5,400	3,681	6.135	5,773.722	4,549.000	13,379.000	851.403	498.000	2,770.000	40,645.137	0.133	9.000	coder:3681
B1	4,800	5,100	1,814	6.047	5,684.976	4,408.500	13,348.050	863.555	515.000	2,762.200	39,596.783	0.138	9.000	coder:1814
B2	5,100	5,400	1,867	6.223	5,859.948	4,698.000	13,419.700	839.595	487.000	2,779.500	41,693.490	0.126	7.600	coder:1867
C	6,000	6,600	3,966	6.610	5,439.353	3,998.000	13,340.500	841.112	520.000	2,642.000	41,513.872	0.143	8.700	coder:3966
C1	6,000	6,300	1,974	6.580	5,641.838	4,228.000	13,376.100	827.749	524.000	2,614.750	42,569.883	0.134	8.700	coder:1974
C2	6,300	6,600	1,992	6.640	5,238.697	3,797.500	13,266.650	854.354	514.500	2,674.400	40,457.860	0.151	8.300	coder:1992

Suggested Combinations

ID transfer: A1 -> A2, B1 -> B2, C1 -> C2
OOD transfer: A1 -> B2, B1 -> C2, C1 -> A2
Three-regime robustness: optimize on A1+B1+C1, test on A2+B2+C2.

Plots

Full trace 60s RPS with colored selected-window bands

Selected-window 10s RPS shape

A/B/C and split-half RPS

Token distributions by selected window

Candidate RPS vs stability