DERP AttachGroup activation and benchmark record

Beginning with v11.3.2, normal builds advertise experimental-derp-attach-group-v1 whenever DERP is available. Two advertising peers automatically use the fixed eight-stream/four-path profile; a peer that does not advertise the feature continues to use legacy single-stream Attach. There is no user-facing tuning control.

The normal-build setting is provided by derpAttachGroupBuildEnabled in src/croc/derp_attach_group_config_default.go. The benchmark-only build tag retains internal topology and fallback controls for reproducible testing.

Activation decision

The original performance gate below found faster Mac-to-server transfers but a normalized server-to-Mac regression. Automatic selection was subsequently enabled for v11.3.2 as an explicit product decision accepting that tradeoff. The historical measurements and gate criteria remain unchanged.

Hermetic framing check

The Go benchmark compares legacy Attach with an eight-stream local AttachGroup using croc's 32 KiB framing, range offsets, deterministic incompressible data, and production AES-GCM encryption:

bash

GOWORK=off go test -tags croc_attach_bench ./src/croc -run '^$' \
  -bench '^BenchmarkDERPAttachFramedTransfer$' -benchtime=8192x -count=3

This is a CPU and allocation regression check, not network promotion evidence.

Streamlined live gate

Build only the selected 8/4 profile and its one-stream force-relay fallback:

bash

GOWORK=off BENCH_UPLOAD=1 \
  benchmarks/derp-attach-group/build_variants.sh
BENCH_FIXTURE_MIB=256 BENCH_FIXTURE_SUFFIX=256m \
  benchmarks/derp-attach-group/prepare_fixtures.sh

Run three alternating legacy/candidate pairs in each direction. Every transfer uses strict DERP, disabled compression, production encryption/framing, and an adjacent ten-second same-direction iperf3 control:

bash

BENCH_PAIRS=3 \
BENCH_LOCAL_SOURCE=/tmp/croc-attach-bench/data/fixture-local-256m.bin \
BENCH_REMOTE_SOURCE=/tmp/croc-attach-bench/data/fixture-remote-256m.bin \
BENCH_RESULT_DIR=/absolute/path/to/results \
  benchmarks/derp-attach-group/live_pair_benchmark.sh
python3 benchmarks/derp-attach-group/analyze_results.py \
  /absolute/path/to/results --candidate group-s8-p4 >analysis.json

The performance portion passes only when all twelve transfers match byte count and SHA-256, all six candidate runs prove mode=raw-direct, streams=8, and raw_paths=4 on both peers, and the candidate's paired normalized-goodput geometric mean exceeds one in both directions. CPU, fixed peak RSS, and setup timings are recorded but do not block the selected throughput profile.

Only after the performance portion passes, run one compatibility transfer in each direction against the latest stable croc and require legacy Attach; run a force-relay transfer and require one manager stream; then run ten small cancellation/reconnect cycles and require bounded shutdown without descriptor, socket, or process growth. A failure in any required stage leaves automatic selection disabled and stops the gate without running later stages.

Keep only the scripts, aggregate analysis, CSV rows, environment summary, and decision in Git. Preserve raw process logs and iperf JSON outside the repository while the result is being reviewed.