| { |
| "dataset_name": "Aeala/ShareGPT_Vicuna_unfiltered", |
| "split": "train", |
| "tokenizer_backend": "tiktoken", |
| "tokenizer_name": "gpt-4", |
| "source_record_count": 50000, |
| "cases_per_category": 8, |
| "categories": [ |
| { |
| "name": "steady_short", |
| "description": "Steady Poisson arrivals dominated by short prompts.", |
| "arrival_mode": "poisson", |
| "requests_per_case": 32, |
| "sampling_policy": "priority_mix", |
| "priority_mode": "normal", |
| "class_weights": { |
| "short": 0.75, |
| "medium": 0.2, |
| "long": 0.05 |
| } |
| }, |
| { |
| "name": "steady_mixed", |
| "description": "Steady mixed load with balanced request lengths.", |
| "arrival_mode": "poisson", |
| "requests_per_case": 64, |
| "sampling_policy": "priority_mix", |
| "priority_mode": "normal", |
| "class_weights": { |
| "short": 0.4, |
| "medium": 0.4, |
| "long": 0.2 |
| } |
| }, |
| { |
| "name": "bursty_chat", |
| "description": "Bursty conversational traffic with mostly short and medium prompts.", |
| "arrival_mode": "burst", |
| "requests_per_case": 96, |
| "sampling_policy": "priority_mix", |
| "priority_mode": "normal", |
| "class_weights": { |
| "short": 0.5, |
| "medium": 0.35, |
| "long": 0.15 |
| } |
| }, |
| { |
| "name": "bursty_long", |
| "description": "Burst-heavy load with a larger share of long prompts.", |
| "arrival_mode": "burst", |
| "requests_per_case": 72, |
| "sampling_policy": "priority_mix", |
| "priority_mode": "normal", |
| "class_weights": { |
| "short": 0.15, |
| "medium": 0.35, |
| "long": 0.5 |
| } |
| }, |
| { |
| "name": "flash_crowd", |
| "description": "Large clustered traffic spikes for stress testing scheduler fairness and queueing.", |
| "arrival_mode": "burst", |
| "requests_per_case": 192, |
| "sampling_policy": "priority_mix", |
| "priority_mode": "normal", |
| "class_weights": { |
| "short": 0.45, |
| "medium": 0.35, |
| "long": 0.2 |
| } |
| }, |
| { |
| "name": "uniform_workload", |
| "description": "Uniform request sizes and steady arrivals for baseline throughput and scheduler overhead.", |
| "arrival_mode": "poisson", |
| "requests_per_case": 64, |
| "sampling_policy": "target_length", |
| "priority_mode": "normal", |
| "class_weights": {} |
| }, |
| { |
| "name": "low_load", |
| "description": "Low-arrival workload with little contention to approximate ideal latency.", |
| "arrival_mode": "poisson", |
| "requests_per_case": 32, |
| "sampling_policy": "priority_mix", |
| "priority_mode": "normal", |
| "class_weights": { |
| "short": 0.45, |
| "medium": 0.4, |
| "long": 0.15 |
| } |
| }, |
| { |
| "name": "short_long_mix", |
| "description": "Majority short requests with a minority of long requests to expose head-of-line blocking.", |
| "arrival_mode": "poisson", |
| "requests_per_case": 80, |
| "sampling_policy": "priority_mix", |
| "priority_mode": "normal", |
| "class_weights": { |
| "short": 0.82, |
| "medium": 0.03, |
| "long": 0.15 |
| } |
| }, |
| { |
| "name": "heavy_tail_distribution", |
| "description": "Long-tailed prompt lengths sampled with a heavy-tail distribution.", |
| "arrival_mode": "poisson", |
| "requests_per_case": 96, |
| "sampling_policy": "heavy_tail", |
| "priority_mode": "normal", |
| "class_weights": {} |
| }, |
| { |
| "name": "poisson_arrivals", |
| "description": "Standard queueing baseline with mixed job sizes and independent random arrivals.", |
| "arrival_mode": "poisson", |
| "requests_per_case": 64, |
| "sampling_policy": "priority_mix", |
| "priority_mode": "normal", |
| "class_weights": { |
| "short": 0.4, |
| "medium": 0.4, |
| "long": 0.2 |
| } |
| }, |
| { |
| "name": "bursty_traffic", |
| "description": "Bursty arrivals with queue buildup followed by recovery periods.", |
| "arrival_mode": "burst", |
| "requests_per_case": 96, |
| "sampling_policy": "priority_mix", |
| "priority_mode": "normal", |
| "class_weights": { |
| "short": 0.45, |
| "medium": 0.35, |
| "long": 0.2 |
| } |
| }, |
| { |
| "name": "adversarial_long_burst", |
| "description": "Many long requests arrive nearly simultaneously to stress worst-case latency and memory pressure.", |
| "arrival_mode": "synchronized", |
| "requests_per_case": 48, |
| "sampling_policy": "top_longest", |
| "priority_mode": "normal", |
| "class_weights": {} |
| }, |
| { |
| "name": "starvation_test", |
| "description": "Continuous short jobs with occasional long jobs to check that long jobs still make progress.", |
| "arrival_mode": "starvation", |
| "requests_per_case": 128, |
| "sampling_policy": "priority_mix", |
| "priority_mode": "normal", |
| "class_weights": { |
| "short": 0.9, |
| "medium": 0.03, |
| "long": 0.07 |
| } |
| }, |
| { |
| "name": "fairness_test", |
| "description": "Heterogeneous request sizes for slowdown and fairness evaluation.", |
| "arrival_mode": "poisson", |
| "requests_per_case": 96, |
| "sampling_policy": "priority_mix", |
| "priority_mode": "normal", |
| "class_weights": { |
| "short": 0.34, |
| "medium": 0.33, |
| "long": 0.33 |
| } |
| }, |
| { |
| "name": "priority_handling", |
| "description": "Mixed request sizes with explicit scheduler priority tiers.", |
| "arrival_mode": "poisson", |
| "requests_per_case": 72, |
| "sampling_policy": "priority_mix", |
| "priority_mode": "tier_mix", |
| "class_weights": { |
| "short": 0.4, |
| "medium": 0.35, |
| "long": 0.25 |
| } |
| }, |
| { |
| "name": "max_context_stress", |
| "description": "Very long prompts near the context cap to stress cache and memory-aware policies.", |
| "arrival_mode": "poisson", |
| "requests_per_case": 40, |
| "sampling_policy": "top_longest", |
| "priority_mode": "normal", |
| "class_weights": {} |
| }, |
| { |
| "name": "fragmentation_scenario", |
| "description": "Many medium and long requests arrive in waves to stress batching and memory layout.", |
| "arrival_mode": "fragmented", |
| "requests_per_case": 120, |
| "sampling_policy": "priority_mix", |
| "priority_mode": "normal", |
| "class_weights": { |
| "short": 0.1, |
| "medium": 0.45, |
| "long": 0.45 |
| } |
| }, |
| { |
| "name": "load_sweep_30", |
| "description": "Mixed workload at approximately 30% target utilization.", |
| "arrival_mode": "poisson", |
| "requests_per_case": 72, |
| "sampling_policy": "priority_mix", |
| "priority_mode": "normal", |
| "class_weights": { |
| "short": 0.45, |
| "medium": 0.35, |
| "long": 0.2 |
| } |
| }, |
| { |
| "name": "load_sweep_50", |
| "description": "Mixed workload at approximately 50% target utilization.", |
| "arrival_mode": "poisson", |
| "requests_per_case": 72, |
| "sampling_policy": "priority_mix", |
| "priority_mode": "normal", |
| "class_weights": { |
| "short": 0.45, |
| "medium": 0.35, |
| "long": 0.2 |
| } |
| }, |
| { |
| "name": "load_sweep_70", |
| "description": "Mixed workload at approximately 70% target utilization.", |
| "arrival_mode": "poisson", |
| "requests_per_case": 72, |
| "sampling_policy": "priority_mix", |
| "priority_mode": "normal", |
| "class_weights": { |
| "short": 0.45, |
| "medium": 0.35, |
| "long": 0.2 |
| } |
| }, |
| { |
| "name": "load_sweep_85", |
| "description": "Mixed workload at approximately 85% target utilization.", |
| "arrival_mode": "poisson", |
| "requests_per_case": 72, |
| "sampling_policy": "priority_mix", |
| "priority_mode": "normal", |
| "class_weights": { |
| "short": 0.45, |
| "medium": 0.35, |
| "long": 0.2 |
| } |
| }, |
| { |
| "name": "load_sweep_100", |
| "description": "Mixed workload at approximately 100% target utilization.", |
| "arrival_mode": "poisson", |
| "requests_per_case": 72, |
| "sampling_policy": "priority_mix", |
| "priority_mode": "normal", |
| "class_weights": { |
| "short": 0.45, |
| "medium": 0.35, |
| "long": 0.2 |
| } |
| } |
| ], |
| "generation_seed": 7 |
| } |
