level-bench

Benchmark abstract-leveldown and levelup stores.
Currently only suitable for use in Node.js.

Example

npm i level-bench leveldown rocksdb
npx level-bench run put leveldown
npx level-bench run put rocksdb
npx level-bench plot put

Yields (outdated):

Highlights

• Target the current working directory or something npm-installed
• Compare benchmarks of different targets or options
• Derives plot labels from benchmark metadata (package, platform, ..)
• Uses unique temporary directories for every db
• Can optionally wrap the db in encoding-down and/or levelup
• Also takes level or something else that's already a levelup interface
• Also takes ioredis and sqlite3 (see third-party/).

Usage

level-bench run <benchmark> [target]

Run a benchmark. The benchmark argument must be one of the named benchmarks listed below.

The target argument should be a path or an npm package name that is installed nearby (for example level-bench run put leveldown). It defaults to the current working directory. A package.json must exist alongside the resolved target.

To wrap target with encoding-down or levelup (you must install these dependencies yourself) pass --encode and/or --levelup (or -el for short). Alternatively target can be something that exports a levelup interface, for example level-bench run put level.

If target does not create persistent databases (like memdown or level-mem) you must pass --mem.

Options for the db can be provided via --db <subargs>. For example --db [--cacheSize 16mb] or --db [--valueEncoding json].

Benchmark-specific options can be provided via -b <subargs>. For example -b [-n 1e6 --concurrency 1]. These options are listed below.

Results are by default written to .benchmarks/<benchmark>.<time>.csv and an accompanying JSON file for metadata. To write results to a custom file specify --out example.csv (-o for short). The metadata is used to derive a distinct benchmark name. When this doesn't suffice (for example because you're benchmarking a spinning disk versus an SSD, a fact that isn't included in the metadata) or when labels in the plot become too long, you can specify a custom name with --name example.

Examples

We can compare the performance of two git branches:

git checkout master && npm i
level-bench run put

git checkout wip && npm i
level-bench run put

Or check the overhead of encoding-down:

level-bench run put memdown --mem
level-bench run put memdown --mem --encode

Or a specific encoding:

level-bench run put level --db [--valueEncoding utf8]
level-bench run put level --db [--valueEncoding json]

Or compare the effect of options:

level-bench run put leveldown
level-bench run put leveldown --db [ --no-compression ]

Then plot both (or more) runs with:

level-bench plot put

Options

Yet to document.

level-bench plot <benchmark> [files]

Plot the results using gnuplot (which must be installed and available in PATH). The files argument should be (glob patterns resolving to) CSV files as generated by level-bench run. If not provided, defaults to .benchmarks/<benchmark>.*.csv.

The plot is written to .benchmarks/<benchmark>.<time>.png by default. This can be overridden with --out <filename> (-o for short).

Options

Yet to document.

Benchmarks

put

Perform concurrent put() operations. Records the Simple Moving Average (SMA) of the duration of the last 1000 writes, as well as the Cumulative Moving Average (CMA) of the throughput in MB/s. Options:

• -n: amount of operations, default 1e6
• --concurrency: default 4
• --keys (string): one of:
  • random (default): generate pseudo-random numeric keys (0-N) with a certain probability distribution
  • seq: non-random, sequential numeric keys (0-N)
  • seqReverse: same keys but in reverse (N-0)
• --values (string): one of:
  • random (default): generate pseudo-random values
  • empty: zero-length values or zero-filled if valueSize is set
• --seed (string): seed to use for random numbers, defaults to 'seed'
• --distribution (string): one of zipfian, uniform (default)
• --skew (floating-point number): Zipfian skew (default 0)
• --offset (number): offset keys (for example to simulate timestamps)
• --valueSize: size of value, as a number in bytes or string with unit (e.g. --valueSize 1kb)
• --keyAsBuffer, --valueAsBuffer (boolean): if not set, keys and values are written as strings (hex encoded).

Tips:

• To benchmark writing sorted data, use --keys seq or seqReverse
• Be mindful of --concurrency when using --keys seq or seqReverse: a high concurrency can counter the performance benefits of writing keys sequentially
• To use the zipfian distribution with a negative skew, specify it as --skew=-1 rather than --skew -1 (which would be interpreted as a flag).

batch-put

Perform concurrent batch() operations. Same as put, but in batches rather than singular puts. Options:

• --batchSize: default 1000, must be a multiple of 10, maximum 1000
• --chained: boolean flag, default false, use chained batch
• --concurrency: default 1
• Other options are the same as of the put benchmark, see above.

self-distribution

Not a benchmark, but a temporary cheat to reuse the tooling we have here to test (and visualize) some of the internals. Needs a valid target argument, same as real benchmarks, although that argument is not actually used.

Generate keys with a certain order and probability distribution. Options:

• -n: amount of keys to generate, default 5e3
• Other options are passed to keyspace

Example:

level-bench run self-distribution memdown -b [--distribution zipfian --skew 1]
level-bench run self-distribution memdown -b [--distribution zipfian --skew=-1]
level-bench run self-distribution memdown -b [--keys seq]
level-bench plot self-distribution

Limitations

The target abstract-leveldown implementation must take a location as its first argument (if persistent) or ignore that argument (if transient). Options are passed to both the constructor with the signature (location, options) and to db.open(options, callback).

Contributing

Level/bench is an OPEN Open Source Project. This means that:

Individuals making significant and valuable contributions are given commit-access to the project to contribute as they see fit. This project is more like an open wiki than a standard guarded open source project.

See the Contribution Guide for more details.

Donate

Support us with a monthly donation on Open Collective and help us continue our work.

License

MIT