bench: the PC-sampling overhead harness (numbers outstanding)

[dpsoft]

Task 12 of the GPU PC sampling plan. The harness, not the verdict — CapEff: 0 and no GPU here, and the plan is explicit that "this task cannot be completed without hardware, and no part of the tier decision may be made without it." Every number is outstanding; no arm of this benchmark has ever executed.

A second workload, not an extension

shim/nvidia/testdata/cuda_concurrent.cu (new, make -C shim nvidia-concurrent). cuda_workload.cu stays as it is: it is the serial fixture the adapter, the join, source resolution and the hardware gate's source-line assertions are all proven against, and reshaping it would change what those gates measure. It is also the wrong instrument — a serial loop gives serialization nothing to destroy.

The new one runs N non-blocking streams whose kernel duration comes from a dependency chain rather than a large grid, so they occupy a fraction of the device and genuinely co-reside. Its computation is the exact identity, so max_abs_err must be 0. Compiled here (nvcc 13.3 is on this box) and verified with cuobjdump -sass that the loops survive: exactly two FFMAs, each inside its own backward branch — nothing folded or unrolled.

Realism is measured, not claimed. Concurrency (Σ exec duration / span) and mean kernel duration are computed from the profile, and the run fails if the baseline arm is near-serial or its kernels are microseconds long.

Every arm proves its own mode, from two ends that can disagree

The adapter's stderr report line and the consumer's counters, both recorded in the JSON. Off ⇒ zero PC samples, windows and bursts. Tier B ⇒ samples and no window. Tier A ⇒ bursts ≥ floor, windows reconciling as 2N/2N−1, no start/stop failures, no graph execs, achieved duty inside a bound derived from the burst timer's tick, and at least one execution actually marked serialized. Cross-arm: a lower duty must open strictly fewer bursts.

Any failure withholds the decision object and exits 3. Tier A gaps are pinned from both sides so the clamp interval collapses to a point and an arm cannot drift to a duty nobody configured.

A benchmark arm that silently did not enable the tier it claims would be this project's signature defect wearing a stopwatch.

A flaw in the thresholds, found without any data

cost ÷ duty > 2 is the same statement as cost > 200 × duty percent — so at 2.5% duty it is exactly the 5% wall-clock bar. TIER_A_DEEP_DIVE_ONLY requires ratio > 2 at every duty tested, which includes 2.5%, so it strictly implies TIER_A_UNSHIPPABLE: on the plan's three duties the deep-dive outcome is unreachable whatever the numbers.

No threshold was adjusted — deciding thresholds after seeing their structure is the same error as deciding them after seeing data. The harness applies them as written and prints the coincidence whenever both fire. Adding a 1%-duty arm would make the distinction real; that is a decision for whoever runs it.

The reassuring half is pinned too: on the plan's duties the decision is total — opt-in, smaller-duty, or unshippable — never indeterminate.

Verification

45 offline unit tests: the parsers, medians, the ratio, all four clauses fired and not-fired including exactly-at-the-bar, each arm assertion shown red, the concurrency measurement, and all four skip branches. Without caps: BENCH_SKIPPED: missing required capabilities (CAP_BPF, CAP_PERFMON, CAP_CHECKPOINT_RESTORE)…, exit 0, no output file — and the message names gpuprobe's own set, deliberately not CAP_SYS_ADMIN.

go build, go vet, go test ./... -count=1, golangci-lint all clean; make -C shim test unaffected.

To run it on the 3090

make -C shim nvidia nvidia-concurrent && make bench-build
sudo setcap cap_bpf,cap_perfmon,cap_checkpoint_restore+ep ./bench/cmd/scenario/scenario
make bench-gpu-pc-overhead

~26 fixed-work runs, 15–25 min. Exit 0 means the measurement completed — an honest TIER_A_UNSHIPPABLE is a successful run. Exit 3 means an arm could not prove what it measured, and the numbers are not a decision.