# Fleet report `fleet_report_091494c4`

**OFFLOAD HEADROOM EXISTS: 13.00 pp of block hits between GPU-only LRU (3.9%) and the 16.9% ceiling at 65,536 tokens. This is an upper bound on any connector -- the host tier here is unbounded and free -- and it counts block hits, not GPU-hours.**

| | |
|---|---:|
| requests | 600 |
| block references | 16,495 |
| distinct blocks | 13,712 |
| block size (tokens) | 64 |
| capacity (tokens / blocks) | 65,536 / 1,024 |
| working set / capacity | 13.391 |
| **cold-miss floor** | **0.8313** |
| max achievable hit rate | 0.1687 |
| GPU-only LRU hit rate | 0.0387 |
| GPU share of achievable | 0.229249 |
| host-backed (unbounded, free) | 0.1687 |
| **offload headroom** | **13.00 pp** |

## Arrivals

| | |
|---|---:|
| span_s | 200.999 |
| offered_qps_mean | 2.9851 |
| offered_qps_peak_60s | 3.65 |
| peak_to_mean | 1.223 |

## LRU hit rate by capacity (tokens)

| capacity | hit rate |
|---:|---:|
| 16,384 | 0.0363 |
| 32,768 | 0.0363 |
| 65,536 | 0.0387 |
| 131,072 | 0.0433 |
| 262,144 | 0.0809 |
| 524,288 | 0.1584 |
| 1,048,576 | 0.1687 |
| 2,097,152 | 0.1687 |
| 4,194,304 | 0.1687 |
| 8,388,608 | 0.1687 |

## GPU-hour conversion: REFUSED

- Converting block hits to GPU-hours needs the bandwidth-bound fraction of GPU time on YOUR model and hardware (SPRINTS.md C11: never measured by anyone in this program) and a paired knee ratio from your engine.
- The one hardware measurement this program has (S46: vLLM's prefix cache raised the sustainable knee 30.5% on Mooncake at this capacity) does not transfer: the effect is a property of the trace, and yours is not that one.

What would change it:

- run `make sprint53` (the S46 protocol) against your own engine on this trace; it returns a paired knee ratio with its confidence interval
- a measured d(GPU-hours)/d(block-hit) for your model class (S49 in the plan)

## Does not prove

- Block hits are not GPU-hours. A hit on a block that was cheap to recompute saves less than a hit on one that was expensive; this report weights every block equally.
- LRU stands in for the engine's eviction. vLLM's prefix cache is a tree; on prefix-structured streams it can do better than LRU at the same capacity, so gpu_only_lru_hit_rate may UNDERSTATE what the engine already captures.
- The host tier is unbounded and free. A real connector pays PCIe transfer, metadata and scheduler overhead; every headroom number here is an upper bound.
- Ordering is the input's ordering. Without timestamps the stream is replayed as given, which may not be arrival order.