{
  "schema": "stable-upstream-long-context-admission-benchmark-v1",
  "date": "2026-08-10",
  "topology": {
    "endpoint_picker_replicas": 1,
    "model_replicas": 2,
    "gpus_per_model_replica": 1
  },
  "hardware": {
    "gpu_per_model_replica": "NVIDIA H100"
  },
  "model_service": {
    "model": "GPT-OSS 20B",
    "runtime": "Red Hat AI Inference Server vLLM",
    "tensor_parallel_size": 1,
    "max_model_len": 32768,
    "max_num_batched_tokens": 8192,
    "max_num_seqs": 128,
    "gpu_memory_utilization": 0.9,
    "prefix_cache": "disabled"
  },
  "endpoint_picker": {
    "version": "llm-d Endpoint Picker v0.9.0",
    "flow_control_gate": "enabled",
    "picker": "random",
    "priority_bands": [
      100,
      0
    ],
    "fairness": "round-robin within each priority band",
    "admission_arms": [
      {
        "name": "request-count admission",
        "concurrency_mode": "requests",
        "max_concurrency_per_model_replica": 128,
        "headroom": 0.1
      },
      {
        "name": "exact-token admission",
        "concurrency_mode": "tokens",
        "max_input_token_concurrency_per_model_replica": 20000,
        "headroom": 0.25,
        "tokenizer": "vLLM model tokenizer",
        "estimated_output_tokens_included": false
      }
    ]
  },
  "traffic": {
    "name": "realtime-with-long-context-burst",
    "duration_s": 270,
    "analysis_windows": [
      {
        "name": "baseline",
        "start_s": 20,
        "end_s": 55
      },
      {
        "name": "burst",
        "start_s": 100,
        "end_s": 190
      },
      {
        "name": "recovery",
        "start_s": 230,
        "end_s": 260
      }
    ],
    "tenants": [
      {
        "fairness_id": "realtime-chat",
        "priority": 100,
        "input_tokens": 1024,
        "output_tokens": 128,
        "phases": [
          {
            "start_s": 0,
            "duration_s": 270,
            "rate_pattern": "noisy_sinusoidal",
            "rate_center": 8,
            "rate_amplitude": 1.5,
            "period_s": 37,
            "phase_offset": 0.0,
            "rate_noise": 0.4,
            "rate_spike_probability": 0.02
          }
        ]
      },
      {
        "fairness_id": "standard-long-context",
        "priority": 0,
        "input_tokens": 20000,
        "output_tokens": 128,
        "phases": [
          {
            "start_s": 60,
            "duration_s": 150,
            "rate_pattern": "noisy_sinusoidal",
            "rate_center": 1.0,
            "rate_amplitude": 0.25,
            "period_s": 29,
            "phase_offset": 0.4,
            "rate_noise": 0.05,
            "rate_spike_probability": 0.02,
            "ramp_s": 8
          }
        ]
      }
    ]
  },
  "execution": {
    "arrival": "open-loop Poisson replay with noisy sinusoidal phases",
    "duration_seconds": 270,
    "paired_holdout_seeds": [
      101,
      102,
      103,
      104,
      105,
      106,
      107,
      108
    ],
    "matched_trace_per_seed": true,
    "load_generator": "GuideLLM",
    "cache_mode": "off"
  },
  "runner": {
    "scenario_source": "pipeline/benchmark.py",
    "sha256": "3811ec26c46bf3a26fa643698ec54bf569bb4bc99c3ea22ca18f805cb077b8e0",
    "traffic_driver": "GuideLLM",
    "guidellm_version": "0.7.0",
    "trace_compiler": "pipeline/guidellm_trace.py",
    "launcher": "pipeline/run_guidellm_scenario.py",
    "scenario_file": "scenario.json"
  }
}