Software performance measurements

Measured 2026-11-02 with CPython 3.14.9 on Linux x86-64. These are costs of the synthetic Python reference implementation, not codec throughput, USB compatibility or capture-to-display latency. Timings are informational; correctness and resource limits remain verification gates.

The comparison table records the optimization snapshot. Later review added explicit stream finalization, jitter discontinuity diagnostics and nullable unknown latency. Current outcome fingerprints include those additions and therefore intentionally differ from historical fingerprints. Compare performance variants of the same hardened revision when checking semantic equivalence; a cross-release fingerprint difference alone is not a regression.

Optimization results

The stage comparison retained all new validation and bounded assembly logic. Only the performance changes were reversed for the comparison: skipping empty pipeline frontiers, reusing a deterministically reseeded random generator, caching the color-bar body/checksum, using a fixed frame header structure and avoiding unused sink-report allocations. Complete outcome fingerprints and counters matched in both stage workloads. Peak tracked allocation for the counter workload remained approximately 1 MiB.

The jitter comparison replaced only the jitter implementation. Emitted units, events and feedback matched. Releasing due packets from a heap avoids repeatedly scanning the remaining buffer.

The observer measurement includes constructing a new observer, decoding two synthetic sequence values and recording both arrivals. It does not measure an isolated packet callback or network throughput. Verdicts and wire counters matched for this workload and checked recovery, duplication, reordering and rollover cases within retained history. An independent repeat measured 1,275.8 µs versus 5.17 µs.

Observation history is bounded to 4,096 sequence positions. A packet older than that window cannot repair an expired gap or prove duplication; those limits are part of the reported semantics, not performance shortcuts that silently preserve an unlimited-history claim.

Reproduce current workloads

make benchmark-control
make benchmark-media BENCH_FRAMES=10000 REPEAT=7

For more focused media measurements:

.scratch/venv/bin/python benchmarks/media.py --frames 10000 --repeat 7 --memory
.scratch/venv/bin/python benchmarks/media.py --frames 10000 --repeat 7 --pattern bars
.scratch/venv/bin/python benchmarks/media.py --workload jitter --frames 1000 --repeat 5

The default four slices per frame make the last command a 4,000-packet burst. Benchmark JSON includes the workload, interpreter, machine class, wall/CPU samples, counters and output fingerprints. Memory sampling and fingerprinting run separately from the timed samples.

The observer microbenchmark uses an injected two-byte sequence decoder so it isolates gap accounting from RTP parsing:

from statistics import median
from timeit import repeat
from yoke.media.snoop import MediaSnooper

def observe_gap():
    observer = MediaSnooper(lambda raw: {"seq": int.from_bytes(raw, "big")})
    observer.observe_packet(b"\x00\x00", 0.0)
    observer.observe_packet(b"\x7f\xff", 1.0)

samples = repeat(observe_gap, number=200, repeat=5)
print(median(samples) / 200 * 1_000_000)  # microseconds per complete workload

For historical comparisons, use baseline commit 10ab5d4 and the same interpreter, workload and machine. Keep the measured path isolated from unrelated work. A targeted improvement does not establish that the entire release is faster: stronger validation, bounds and recovery add work. An early whole-pipeline comparison was slower and overlapped other lanes, so it does not support an attributable performance claim.

Verification boundary

Use the simulation campaign and make verify before interpreting timings. Benchmark tests verify deterministic results and successful integrity, without imposing a speed threshold on shared CI. The independent code review and final cold review cover the optimized paths alongside their correctness and memory bounds.