mirror of https://github.com/luxfi/bench.git synced 2026-10-30 17:01:57 +00:00

Files

T

zeekay 65a56203c3 chore: initial commit — tree had no git at all

github.com/luxfi/bench existed only as loose files: no .git, no history, no
remote. An rm -rf would have destroyed it with nothing to recover from, and it
was invisible to every repo sweep that looks for a .git directory.

Includes the dependency repair that makes it build again. It previously failed
with:

    reading github.com/luxfi/geth/go.mod at revision v1.16.98: unknown revision

The v1.16.x line was DELETED from github.com/luxfi/geth. The requirement came
from github.com/luxfi/crypto, which pins the dead tag AND excludes the
alternative in every version through v1.20.0; v1.20.2 is the first that moved
to geth v1.20.1, so bumping crypto is what escapes it. geth then drops out of
go.mod entirely.

The compiled luxbench binary is ignored rather than committed.

Verified: go build ./... exits 0.

NOTE: no remote — luxfi/bench does not exist on GitHub. This is still one copy
in one place.

2026-07-26 04:23:44 -07:00

2.0 KiB

Raw Permalink Blame History

bench — luxbench unified harness

Single bench driver for the luxfi stack. Replaces the ad-hoc /tmp/<name>-bench/ directories.

Layout

bench/
├── cmd/luxbench/        # CLI entry: flag parsing + mode dispatch
├── mode.go              # Mode interface (Name/Description/RegisterFlags/Run)
├── modes/
│   ├── zap_vs_codec/    # CANARY: 8 head-to-head linearcodec vs zap claims
│   ├── gpu/             # scaffold (TODO: port /tmp/spark-gpu-bench/)
│   ├── multi_validator/ # scaffold (TODO: port /tmp/multi-validator/)
│   ├── parity/          # scaffold (TODO: port lux/threshold/protocols/parity/)
│   └── matrix/          # scaffold (TODO: 7-profile sig-pack matrix)
├── result/schema.go     # ResultRecord + File envelope + JSON IO + quantiles
└── README.md            # user-facing docs

Invariants

One harness. No new /tmp/<thing>-bench/ directories — every new bench is a mode under modes/.
One result schema (result.ResultRecord). Mode-specific metrics go in Extra (snake_case keys).
Modes implement bench.Mode; CLI dispatches by --mode=<name>.
The CLI owns JSON serialization. Modes return []ResultRecord, never write files directly.

Building

cd ~/work/lux/bench && GOWORK=off go build ./...
GOWORK=off go test ./...

GOWORK=off because the harness has its own go.mod with local replace directives for luxfi/codec and luxfi/zap. Don't add it to the workspace until the upstream tags catch up.

Canary verification

zap-vs-codec mode produces the same numbers as the original /tmp/zap-vs-codec/ would have, within run-to-run noise. Deterministic metrics (B/op, allocs/op) are identical:

GOWORK=off /tmp/luxbench --mode=zap-vs-codec \
  --bench='^BenchmarkRejectBadMagic_(Codec|ZAP)$' \
  --benchtime=2s --count=3 --output=/tmp/x.json

vs. running go test -bench=... directly in modes/zap_vs_codec/ — the bench source is the same files, the harness only wraps execution and parses output.