{
  "execution_mode": "CPU algorithm with controlled inputs; any latency/capacity coefficients are assumptions",
  "experiment": "multi-tenant-service-token-drr-and-admission-control",
  "decode_weight": 4,
  "interactive_ttft_slo_ms": 500.0,
  "quotas": [
    {
      "tenant_id": "tenant_chat_pro",
      "weight": 2.0,
      "burst_service_tokens": 48000,
      "sustained_service_tokens_per_sec": 180000,
      "priority_class": "interactive"
    },
    {
      "tenant_id": "tenant_ide_assist",
      "weight": 1.5,
      "burst_service_tokens": 36000,
      "sustained_service_tokens_per_sec": 140000,
      "priority_class": "interactive"
    },
    {
      "tenant_id": "tenant_batch_rag",
      "weight": 1.0,
      "burst_service_tokens": 28000,
      "sustained_service_tokens_per_sec": 60000,
      "priority_class": "batch"
    }
  ],
  "comparison": {
    "request_count_no_admission": {
      "total_requests": 60,
      "admitted_requests": 60,
      "shed_429_requests": 0,
      "interactive_p50_ttft_ms": 1228.31,
      "interactive_p99_ttft_ms": 2093.0,
      "batch_p99_ttft_ms": 2125.56,
      "peak_kv_occupancy_ratio": 0.7275,
      "shed_reasons": {},
      "tenants": {
        "tenant_batch_rag": {
          "submitted": 18,
          "admitted": 18,
          "shed_429": 0,
          "admitted_service_tokens": 82800,
          "p50_ttft_ms": 1035.56,
          "p99_ttft_ms": 2125.56
        },
        "tenant_chat_pro": {
          "submitted": 24,
          "admitted": 24,
          "shed_429": 0,
          "admitted_service_tokens": 15360,
          "p50_ttft_ms": 1382.44,
          "p99_ttft_ms": 2093.0
        },
        "tenant_ide_assist": {
          "submitted": 18,
          "admitted": 18,
          "shed_429": 0,
          "admitted_service_tokens": 11880,
          "p50_ttft_ms": 1006.08,
          "p99_ttft_ms": 2006.08
        }
      }
    },
    "service_token_drr_with_admission": {
      "total_requests": 60,
      "admitted_requests": 46,
      "shed_429_requests": 14,
      "interactive_p50_ttft_ms": 222.08,
      "interactive_p99_ttft_ms": 358.97,
      "batch_p99_ttft_ms": 948.67,
      "peak_kv_occupancy_ratio": 0.65,
      "shed_reasons": {
        "shed_kv_watermark": 14
      },
      "tenants": {
        "tenant_batch_rag": {
          "submitted": 18,
          "admitted": 4,
          "shed_429": 14,
          "admitted_service_tokens": 18400,
          "p50_ttft_ms": 764.22,
          "p99_ttft_ms": 948.67
        },
        "tenant_chat_pro": {
          "submitted": 24,
          "admitted": 24,
          "shed_429": 0,
          "admitted_service_tokens": 15360,
          "p50_ttft_ms": 211.67,
          "p99_ttft_ms": 322.78
        },
        "tenant_ide_assist": {
          "submitted": 18,
          "admitted": 18,
          "shed_429": 0,
          "admitted_service_tokens": 11880,
          "p50_ttft_ms": 268.31,
          "p99_ttft_ms": 358.97
        }
      }
    }
  },
  "sample_decisions": [
    {
      "request_id": "batch-00",
      "tenant_id": "tenant_batch_rag",
      "priority_class": "batch",
      "status_code": 200,
      "reason": "admitted",
      "service_tokens": 4600,
      "ttft_ms": 66.67,
      "queue_wait_ms": 0.0,
      "kv_occupancy_ratio": 0.1388
    },
    {
      "request_id": "batch-04",
      "tenant_id": "tenant_batch_rag",
      "priority_class": "batch",
      "status_code": 429,
      "reason": "shed_kv_watermark",
      "service_tokens": 4600,
      "ttft_ms": 0.0,
      "queue_wait_ms": 0.0,
      "kv_occupancy_ratio": 0.65
    },
    {
      "request_id": "batch-05",
      "tenant_id": "tenant_batch_rag",
      "priority_class": "batch",
      "status_code": 429,
      "reason": "shed_kv_watermark",
      "service_tokens": 4600,
      "ttft_ms": 0.0,
      "queue_wait_ms": 0.0,
      "kv_occupancy_ratio": 0.65
    },
    {
      "request_id": "batch-06",
      "tenant_id": "tenant_batch_rag",
      "priority_class": "batch",
      "status_code": 429,
      "reason": "shed_kv_watermark",
      "service_tokens": 4600,
      "ttft_ms": 0.0,
      "queue_wait_ms": 0.0,
      "kv_occupancy_ratio": 0.65
    },
    {
      "request_id": "batch-07",
      "tenant_id": "tenant_batch_rag",
      "priority_class": "batch",
      "status_code": 429,
      "reason": "shed_kv_watermark",
      "service_tokens": 4600,
      "ttft_ms": 0.0,
      "queue_wait_ms": 0.0,
      "kv_occupancy_ratio": 0.65
    },
    {
      "request_id": "batch-08",
      "tenant_id": "tenant_batch_rag",
      "priority_class": "batch",
      "status_code": 429,
      "reason": "shed_kv_watermark",
      "service_tokens": 4600,
      "ttft_ms": 0.0,
      "queue_wait_ms": 0.0,
      "kv_occupancy_ratio": 0.65
    },
    {
      "request_id": "batch-09",
      "tenant_id": "tenant_batch_rag",
      "priority_class": "batch",
      "status_code": 429,
      "reason": "shed_kv_watermark",
      "service_tokens": 4600,
      "ttft_ms": 0.0,
      "queue_wait_ms": 0.0,
      "kv_occupancy_ratio": 0.65
    },
    {
      "request_id": "batch-10",
      "tenant_id": "tenant_batch_rag",
      "priority_class": "batch",
      "status_code": 429,
      "reason": "shed_kv_watermark",
      "service_tokens": 4600,
      "ttft_ms": 0.0,
      "queue_wait_ms": 0.0,
      "kv_occupancy_ratio": 0.65
    },
    {
      "request_id": "chat-00",
      "tenant_id": "tenant_chat_pro",
      "priority_class": "interactive",
      "status_code": 200,
      "reason": "admitted",
      "service_tokens": 640,
      "ttft_ms": 105.22,
      "queue_wait_ms": 97.22,
      "kv_occupancy_ratio": 0.1537
    },
    {
      "request_id": "batch-11",
      "tenant_id": "tenant_batch_rag",
      "priority_class": "batch",
      "status_code": 429,
      "reason": "shed_kv_watermark",
      "service_tokens": 4600,
      "ttft_ms": 0.0,
      "queue_wait_ms": 0.0,
      "kv_occupancy_ratio": 0.65
    }
  ]
}
