Qwen-Bailian Anonymous Dataset Overview
Four anonymized production-derived Qwen serving traces from Aliyun Bailian. Each record has trace-relative arrival timestamp, input/output token counts, request type, session turn metadata, and 16-token KV block hash IDs. Request durations are not included.
Use for benchmarking: AIPerf supports these JSONL files with
--custom-dataset-type bailian_trace and --fixed-schedule. Since durations are absent, concurrency should be measured during replay rather than inferred directly from the raw trace.Scenario Comparison
| Trace | Scenario | Requests | Hours | Mean RPS | Peak 1s RPS | Input min | Input max | Input p50 | Input p99 | Output p50 | Output p99 | Hash reuse |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Qwen Bailian Trace A | To-C chat-style interactive services | 43,058 | 2.000 | 5.981 | 18.00 | 56 | 89,286 | 1,046.00 | 14,364.59 | 375.50 | 1,641.43 | 59.8% |
| Qwen Bailian Trace B | To-B API-driven task automation | 172,800 | 2.000 | 24.000 | 52.00 | 19 | 66,446 | 574.00 | 6,294.01 | 39.00 | 1,006.02 | 61.5% |
| Qwen Bailian Thinking | Reasoning-intensive chat workload | 10,812 | 1.998 | 1.503 | 40.00 | 11 | 51,622 | 3,680.00 | 24,972.68 | 1,665.50 | 34,686.53 | 46.2% |
| Qwen Bailian Coder | Code generation workload | 43,011 | 2.000 | 5.975 | 19.00 | 9 | 25,777 | 4,540.00 | 14,406.90 | 469.00 | 5,842.20 | 66.4% |
Request Type Composition
For Trace A and Trace B, the text-only windows used for benchmark selection are lower RPS than the aggregate traces because non-text request types are excluded.
| Trace | Per-type load | Interpretation |
|---|---|---|
| Qwen Bailian Trace A | text: 31,744 (73.72%, 4.409 RPS); search: 8,187 (19.01%, 1.137 RPS); image: 1,617 (3.76%, 0.225 RPS); file: 1,510 (3.51%, 0.210 RPS) | Text-only RPS is close to the full trace, but still lower than aggregate. |
| Qwen Bailian Trace B | api: 150,936 (87.35%, 20.963 RPS); text: 21,864 (12.65%, 3.037 RPS) | Aggregate 24 RPS is mostly api traffic; text-only load is ~3 RPS. |
| Qwen Bailian Thinking | thinking: 10,812 (100.00%, 1.503 RPS) | Single native workload type. |
| Qwen Bailian Coder | coder: 43,011 (100.00%, 5.975 RPS) | Single native workload type. |
Comparison Plots
