Files

T

sami7777 674bdc7192 Add chain-DB benchmark harness (contrib/bench/)

bench-chaindb.sh times FastImportBlockFile() under each backend using a
user-supplied blk0001.dat. Wall time comes from the daemon's existing
StartupPerfLog line; peak RSS via ps sampling; datadir size via du.

Output is one CSV row per backend appended to ./bench-results.csv, plus
a stdout summary. Network is disabled during the run (-nolisten -connect=0)
so we measure only DB ingest cost.

Does not yet measure: reorg cost, network IBD speed, raw disk I/O.
LoadSnapshot path is still LevelDB-only; the harness intentionally exercises
the FastImportBlockFile rebuild instead, which works on both backends.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

2026-04-26 17:52:34 -07:00

2.9 KiB

Raw Blame History

Chain DB benchmark harness

Measures FastImportBlockFile() speed under each chain-DB backend (LevelDB vs RocksDB) using a user-supplied blk0001.dat block stream.

Prerequisites

A trianglesd binary built with both backends:

cmake -B build -G Ninja \
    -DCMAKE_BUILD_TYPE=Release \
    -DBUILD_QT=OFF \
    -DBUILD_DAEMON=ON \
    -DBUILD_ROCKSDB=ON
cmake --build build

An blk0001.dat file (old-style block stream). If you have a synced node, copy ~/.triangles/blk0001.dat (Linux) or %APPDATA%\triangles\blk0001.dat (Windows).
Free disk space: ~3× the size of blk0001.dat per backend run (raw blocks + chain DB index + working space).

Usage

contrib/bench/bench-chaindb.sh \
  --binary=$(pwd)/build/bin/trianglesd \
  --bootstrap=/path/to/blk0001.dat

Runs each backend in turn, appends a CSV row to ./bench-results.csv, and prints a summary to stdout. Default --dbcache=2048 (MB).

Options

Flag	Default	Notes
`--binary=PATH`	(required)	Path to `trianglesd`
`--bootstrap=PATH`	(required)	Path to `blk0001.dat`
`--backends=LIST`	`leveldb,rocksdb`	Comma-separated subset
`--workdir=DIR`	`/tmp/triangles-bench-XXXXXX`	Per-backend datadirs go here
`--dbcache=MB`	`2048`	Chain DB cache size
`--results-csv=FILE`	`./bench-results.csv`	Appended to
`--keep-datadirs`	off	Preserve datadirs after run for inspection
`--rpc-port=BASE`	`19112`	Each backend uses `BASE+offset`

What it measures

Column	Source
`wall_ms`	The daemon's own log line: `FastImportBlockFile: indexed N blocks in Mms`
`peak_rss_kb`	`ps -o rss=` sampled once per second
`datadir_bytes`	`du -sb` of the working datadir (includes `blk0001.dat`)
`blocks_indexed`	Parsed from the same log line

What it does not measure

Network IBD (peer fetch, header sync) — this is pure DB ingest.
UTXO snapshot load — LoadSnapshot is currently rocksdb-guarded (see src/utxosnapshot.cpp); will be unblocked when LevelDB is retired.
Reorg cost — separate test, not yet implemented.
Disk I/O bytes (read/written) — could be added with iostat integration.

Interpreting results

A meaningful comparison requires both rows to have run on the same machine with the same blk0001.dat. The host column makes mixing runs across machines visible in the CSV.

Backend-relevant size comparisons should subtract bootstrap_size_bytes from datadir_bytes to isolate the chain DB tree.

One-liners

# LevelDB only
./bench-chaindb.sh --binary=... --bootstrap=... --backends=leveldb

# Compare 2GB vs 4GB cache on RocksDB
./bench-chaindb.sh --binary=... --bootstrap=... --backends=rocksdb --dbcache=2048
./bench-chaindb.sh --binary=... --bootstrap=... --backends=rocksdb --dbcache=4096

# Keep the datadirs for poking around afterwards
./bench-chaindb.sh --binary=... --bootstrap=... --keep-datadirs

2.9 KiB Raw Blame History Unescape Escape