Home

Awesome

BenchFX: a collection of benchmark programs for WasmFX

This repository contains a suite of curated benchmarks for performance testing the WasmFX implementation in wasmtime.

Using the benchmark harness

The benchmark harness is invoked by running harness.py, its user-facing configuration resides in config.py.

The harness has 3 subcommands, determining its mode of operation.

The simplest way of running all the benchmarks is as follows:

./harness setup # required only once
./harness run

Note that both the toplevel harness and the individual subcommand each have --help options:

./harness              --help
./harness setup        --help
./harness run          --help
./harness compare-revs --help

All benchmarks are listed in config.py. Each benchmark is part of a benchmark suite, which is uniquely identified by a subfolder inside the benchfx repository and contains logically related benchmarks (e.g., different implementations of the same program to compare their performance).

The harness is designed such that all dependencies (including wasmtime) are built from source at fixed commits or a specific version is downloaded. In other words, invoking ./harness.py with the same options at the same revision of the benchfx repository (and therefore the same definitions in config.py), should uniquely determine the exact versions of all dependencies as well as the flags they are built and run with.

setup subcommand

The benchmark harness uses and controls the following repositories as subfolders in tools/external:

Before each benchmark run, the harness checks out and builds the requested revision of each tool (determined either in config.py or overridden with a command line flag) before running benchmarks.

When running ./harness.py setup, the harness checks that these repositories exist and if not, clones them from Github appropriately.

The harness also uses the WASI SDK, running ./harness.py setup checks that the version configured in config.py exists in the tools subdirectory and downloads it otherwise.

In addition, the harness requires a few standard tools (hyperfine, cmake, make, dune, ... ) and will report and error if these are not found in $PATH. These must be installed manually by the user.

Using your own development repository of wasmtime

Since the benchmarks will be executed by checking out and building a specific commit of wasmtime, it can be handy not to use a Github repository, but to use commits only available in a local development repository.

The setup subcommand therefore allows that instead of checking out wasmtime from Github, it creates two git worktrees inside tools/external/, which are connected to your development repository elsewhere. This can be achieved as follows:

./harness.py setup \
  --wasmtime-create-worktree-from-development-repo ~/path/to/my-wasmtime-devel-repository

This effectively means that tools/external/wasmtime1 and tools/external/wasmtime2 are not independent git repositories, but share the .git folder with the development repository and can see all commits therein.

run subcommand

This is the main way to perform actual benchmarking. In this mode, for each benchmark suite defined in config.py, the benchmarks of that suite are compared against each other.

The subcommand has multiple options to override for example how wasmtime is built and run, see ./harness.py run --help for a full list of options.

Filtering benchmarks

The --filter option can be used to run only a subset of the benchmarks.

Filters are ordinary glob patterns that are matched against a pseudo-path identifying each benchmark, of the form <suite's path>/<benchmark name>. For example the suite with path c10m contains a benchmark called c10m_wasmfx. It is selected if the glob filter matches the pseudo-path c10m/c10m_wasmfx.

The --filter option can be used multiple times, and the harness will run a benchmark if it matches any of the filters.

compare-revs subcommand

This subcommand can be used to compare two revisions of wasmtime, rev1 and rev2 against each other. Unlike the run subcommand, benchmarks are not compare against the others in the same suite. Instead, for each suite s and each benchmark b in s, we compare b executed by wasmtime rev1 against b executed wasmtime rev2.

The two revisions are provided to the subcommand directly as positional arguments:

./harness.py compare-revs main my-feature-branch

Filtering works the same as for the run subcommand.

Most other options available for the run subcommand are also available, but are prefixed with rev1- and rev2- now.

As a result, compare-revs can actually compare different configurations rather than just revision: By using the same argument for the revision to use, but varying the other options, we can determine their influence.

./harness.py compare-revs --filter="*/*wasmfx*" \
  --rev1-wasmtime-run-args="-W=exceptions,function-references,typed-continuations -Wwasmfx-stack-size=4096 -Wwasmfx-red-zone-size=0" \
  --rev2-wasmtime-run-args="-W=exceptions,function-references,typed-continuations -Wwasmfx-stack-size=8192 -Wwasmfx-red-zone-size=0" \
  my-branch my-branch

Examples

Within each suite, only compare the wasmfx implementations against asyncify:

./harness.py run --filter="*/*wasmfx*" --filter="*/*asyncify*"

Running benchmarks using a particular wasmtime commit:

./harness.py run my-special-feature-branch

Enable verbose mode of harness itself (note -v appearing before subcommand name), disable mimalloc use:

./harness.py -v run --use-mimalloc=n

Gotchas