Downscaling

Lower-load variants of Qwen-Bailian Trace A/A1 and Trace B/B1 for measuring TTFT and other serving metrics below the original saturation point. Retained requests keep their original timestamps, ordering, input/output lengths, and payloads.

Source windows

Trace A/A1 · B/B1

Duration

300 seconds

Sampling levels

25/50/75/100%

Sampling seed

Only selection changes. Every retained JSONL record is byte-identical to its source. Samples are nested: 25% ⊂ 50% ⊂ 75% ⊂ 100%. Each 100% file is identical to its original B1 file.

Sampling methods

Stratified requests

Individual requests are sampled across 10-second time buckets, request type, log2 input length, and log2 output length. This keeps the 300-second schedule and approximates the full workload mix.

Complete conversations

Conversation roots are resolved from each full parent trace. All requests from each selected conversation that occur inside the chosen B1 window are retained; traffic outside the window is not added.

Trace A · A1

Original load: 5.183 RPS · 1,555 requests.

Stratified request files

Level	Requests	RPS	Input p50	Input p90	Output p50	Output p90	File
25%	390	1.300	914	6,644	369	786	Download JSONL
50%	778	2.593	931	6,347	366	792	Download JSONL
75%	1,166	3.887	928	6,162	363	776	Download JSONL
100%	1,555	5.183	961	5,817	356	771	Download JSONL

Complete conversation files

Level	Requests	Conversations	RPS	Input p50	Input p90	Output p50	Output p90	File
25%	387	300	1.290	897	5,263	343	805	Download JSONL
50%	775	603	2.583	921	5,427	354	758	Download JSONL
75%	1,166	904	3.887	940	5,626	355	768	Download JSONL
100%	1,555	1,208	5.183	961	5,817	356	771	Download JSONL

Trace A/A1 machine-readable summary

Trace B · B1

Original load: 26.600 RPS · 7,980 requests.

Stratified request files

Level	Requests	RPS	Input p50	Input p90	Output p50	Output p90	File
25%	1,994	6.647	724	1,603	52	153	Download JSONL
50%	3,988	13.293	729	1,606	54	161	Download JSONL
75%	5,985	19.950	736	1,623	56	167	Download JSONL
100%	7,980	26.600	736	1,627	55	166	Download JSONL

Complete conversation files

Level	Requests	Conversations	RPS	Input p50	Input p90	Output p50	Output p90	File
25%	1,994	1,994	6.647	735	1,626	54	165	Download JSONL
50%	3,989	3,989	13.297	734	1,640	55	162	Download JSONL
75%	5,986	5,986	19.953	734	1,632	53	162	Download JSONL
100%	7,980	7,980	26.600	736	1,627	55	166	Download JSONL

Trace B/B1 machine-readable summary

Validation

Generation passed record-identity, nested-subset, full-window identity, and complete-conversation-integrity checks for both source windows.

AIPerf: --custom-dataset-type bailian_trace --fixed-schedule --fixed-schedule-auto-offset