command-center/agent/managed-skills/radiance-disk-tier-throughput-census/SKILL.md

name, description

name	description
radiance-disk-tier-throughput-census	Measure a radiance rig's real prefill rate and prove whether the bottleneck is disk, CPU or shape — for calibration taps, KLD recording or score boots; use when a rate looks wrong, an ETA must be quoted, or someone asks about CPU allocation.

Census a radiance rig's throughput without lying to yourself

Ask three questions in order: what is the rate, what is the binding resource, and does the shape explain it. Every number below comes from a kernel or engine counter — never from a gauge.

1. The rate: counters, never prefill_tps

/stats on :7770 exposes prefill_tps, and it is a rolling gauge. Reading 69.5 → 149.1 across 120 s and multiplying by uptime_s produced a bogus "320 tok/s / 10.4 h" verdict that had to be retracted (measured 2026-10-29). Derive the rate from counters over a ≥240 s window:

python3 - <<'PY'
import json, time, urllib.request
def stats(): return json.load(urllib.request.urlopen("http://127.0.0.1:7770/stats", timeout=10))
a = stats(); t0 = time.time(); time.sleep(240); b = stats(); t1 = time.time(); dt = t1 - t0
docs = b["total_requests"] - a["total_requests"]
steps = b["steps_prefill"] - a["steps_prefill"]
MEAN = 4238.0            # tokens/doc, from the corpus: total_tokens / documents
print("docs %d in %.0fs -> %.0f tok/s | steps/doc %.2f" % (docs, dt, docs * MEAN / dt, steps / docs))
PY

steps/doc is the read-amplification signal: a document cut into N prefill steps re-streams each layer's routed expert union N times when the weights are on the disk tier, so N > 1 is leaked time.

2. The binding resource: diskstats and the idle/iowait split

# per-device read bandwidth (parent device, not the partition)
awk '{print $3, $6*512}' /proc/diskstats | grep -E "^nvme"
# host busy = total - idle - iowait (summing /proc/stat's first 8 fields WITHOUT subtracting is
# the classic error: the sum already covers idle, so it reads ~800% on an 8-core box)

Then split CPU by process: sample /proc/<pid>/stat fields 14+15 (utimes + stimes) twice.

Two hard readings that decide the argument:

• radiance compute at ~0.26 cores while the disk delivers ~3.2 GB/s and iowait is ~29 % of the box ⇒ disk-bound. More cores or a --cpuset-cpus pin cannot add bytes; pinning only caps it.
• A container with cpu_quota=0 and an empty CpusetCpus (check podman inspect -f '{{.HostConfig.CpuQuota}} {{.HostConfig.CpusetCpus}}') is already unrestricted — "allocate more CPU" is a no-op, and "allocate 6 of 8" is a ceiling.

Sanity relation: tok/s × (GiB streamed per document ÷ tokens per document) should match the measured device bandwidth. If it does not, the shape, not the hardware, is the story.

3. The shape: what may legally be widened

• kldref and score must keep production's --max-num-batched-tokens and --max-num-seqs: their KL numbers are only comparable across boots at an identical shape (the audit law).
• A calibration tap is exempt — its output is layer Hessians, nothing compares it. Give it a step that covers a whole document (Q38_CALIB_MNBT ≥ the corpus window) and read steps/doc: on a 7,168-token corpus, 2,048 → 8,192 moved 2.5 → 1.00 steps/doc and 94 → 113 tok/s (1.2×, not the 4× the step-count ratio suggests — total bytes are unchanged, so do not promise more).
• The KV pool, not the step, bounds concurrency: when kv_tokens_total ≈ one document, requests queue and preempt; extra slots buy nothing.

Reporting

Quote tok/s, steps/doc, device GB/s and the busy/iowait split together, plus the shape each number was taken at. A rate without its shape is not a measurement; a gauge without its window is not a rate.