{
  "dataset": "cluster-trace-v2026-GenAI (REAL DATA)",
  "analysis": "TP4 NVLink Deadlock Cascade",
  "total_llm_intervals": 157411,
  "original_metric": {
    "definition": "VRAM >90% pod_ceil AND duty <5%",
    "intervals": 7553,
    "capacity_pct": 4.7983,
    "wasted_kwh": 44.06,
    "note": "Only caught GPU 0 deadlock \u2014 NVLink waiters invisible to this filter"
  },
  "gpu0_deadlock": {
    "definition": "VRAM >85% ceil AND duty==0",
    "intervals": 9200,
    "capacity_pct": 5.8446,
    "wasted_kwh": 53.67,
    "containers": 134,
    "mean_vram_pct": 0.939
  },
  "nvlink_waiters": {
    "definition": "VRAM 30-70% ceil AND duty==0 (GPU 1/2/3 stalled on NVLink all-reduce)",
    "intervals": 71678,
    "capacity_pct": 45.5356,
    "wasted_kwh": 418.12,
    "containers": 127,
    "mean_vram_pct": 0.585,
    "note": "Invisible to original filter \u2014 below 90%% of own historical ceiling"
  },
  "full_tp4_blast_radius": {
    "definition": "All duty==0 intervals",
    "intervals": 105226,
    "capacity_pct": 66.8479,
    "wasted_kwh": 613.82,
    "containers": 140
  },
  "vram_locked_by_tp4": {
    "gpu0_vram_gb_intervals": 222368.7,
    "nvlink_waiter_vram_gb_intervals": 1649527.9,
    "total_vram_gb_intervals": 3616626.1,
    "pct_total_vram_locked": 51.76
  },
  "tp4_costal_pattern": {
    "timestamps_showing_pattern": 1440,
    "total_timestamps": 1441,
    "gpu0_containers_per_event": 4,
    "gpu0_mean_vram_pct": 0.934,
    "waiter_containers_per_event_median": 59,
    "waiter_mean_vram_pct": 0.585,
    "confirmed_tp4_events": 1440
  },
  "reinterpretation": "The 4.8%% original metric was measuring only GPU 0. The TP4 all-reduce deadlock holds GPU 1/2/3 at 56.9%% VRAM with zero SM activity \u2014 45.54%% additional capacity destroyed. Total TP4 blast radius: 51.76%% of all fleet VRAM locked at any given moment."
}