sightglass

A benchmarking suite and tooling for Wasmtime and Cranelift

A Bytecode Alliance project

Contributing | Chat

About

This repository contains benchmarking infrastructure for Wasmtime and Cranelift, as described in this RFC. In particular, it has

• a benchmark suite of Wasm applications in benchmarks/*, and

• a benchmark runner CLI tool to record, analyze, and display benchmark results in crates/cli/*.

We plan to implement a server that periodically runs benchmarks as new commits are pushed to Wasmtime and display the history of those benchmark results, similar to Firefox's Are We Fast Yet?. However, this work is not completed yet. See issue 93 for details.

Results are always broken down by phase — compilation vs. instantiation vs. execution — for each program in the suite. This allows us to reason about, for example, compiler performance separately from its generated code quality. How all this works together:

• each benchmark is compiled to a benchmark.wasm module that calls two host functions, bench.start and bench.end, to notify Sightglass of the portion of the execution to measure (see the benchmarks README)
• we build an engine (e.g., Wasmtime) as a shared library that implements the [bench API]; the Sightglass infrastructure uses this to measure each phase (see an engine README)
• the sightglass-cli tool runs benchmarks using the engines and emits measurements for each phase; this is configurable, e.g., by various measurement mechanisms, various output formats, different aggregations, etc.

This is NOT a General-Purpose WebAssembly Benchmark Suite

This benchmark suite and tooling is specifically designed for Wasmtime and Cranelift, as explained in the benchmarking suite RFC:

It is also worth mentioning this explicit non-goal: we do not intend to develop a general-purpose WebAssembly benchmark suite, used to compare between different WebAssembly compilers and runtimes. We don't intend to trigger a WebAssembly benchmarking war, reminiscent of JavaScript benchmarking wars in Web browsers. Doing so would make the benchmark suite's design high stakes, because engineers would be incentivized to game the benchmarks, and would additionally impose cross-engine portability constraints on the benchmark runner. We only intend to compare the performance of various versions of Wasmtime and Cranelift, where we don't need the cross-engine portability in the benchmark runner, and where gaming the benchmarks isn't incentivized.

Furthermore, general-purpose WebAssembly benchmarking must include WebAssembly on the Web. Doing that well requires including interactions with the rest of the Web browser: JavaScript, rendering, and the DOM. Building and integrating a full Web browser is overkill for our purposes, and represents significant additional complexity that we would prefer to avoid.

Even if someone did manage to get other Wasm engines hooked into this benchmarking infrastructure, comparing results across engines would likely be invalid. The wasmtime-bench-api intentionally does things that will likely hurt its absolute performance numbers but which help us more easily get statistically meaningful results, like randomizing the locations of heap allocations. Without taking great care to level the playing field with respect to these sorts of tweaks, as well as keeping an eye on all engine specific configuration options, you'll end up comparing apples and oranges.

Usage

You can always see all subcommands and options via

cargo run -- help

There are flags to control how many different processes we spawn and take measurements from, how many iterations we perform for each process, etc...

That said, here are a couple of typical usage scenarios.

Building the Runtime Engine for Wasmtime

$ cd engines/wasmtime && rustc build.rs && ./build && cd ../../

Running the Full Benchmark Suite

$ cargo run -- benchmark --engine engines/wasmtime/libengine.so -- benchmarks/*/benchmark.wasm

The output will be a summary of each benchmark program's compilation, instantiation, and execution times.

Running a Single Wasm Benchmark

$ cargo run -- benchmark --engine engines/wasmtime/libengine.so -- path/to/benchmark.wasm

Comparing a Feature Branch to Main

First, build libwasmtime_bench_api.so (or .dylib or .dll depending on your OS) for the latest main branch:

$ cd ~/wasmtime
$ git checkout main
$ cargo build --release -p wasmtime-bench-api
$ cp target/release/libwasmtime_bench_api.so /tmp/wasmtime_main.so

Then, checkout your feature branch and build its libwasmtime_bench_api.so:

$ git checkout my-feature
$ cargo build --release -p wasmtime-bench-api

Finally, run the benchmarks and supply both versions of libwasmtime_bench_api.so via repeated use of the --engine flag:

$ cd ~/sightglass
$ cargo run -- \
    benchmark \
    --engine /tmp/wasmtime_main.so \
    --engine ~/wasmtime/target/release/libwasmtime_bench_api.so \
    -- \
    benchmarks/*/benchmark.wasm

The output will show a comparison between the main branch's results and your feature branch's results, giving you an effect size and confidence interval (i.e. "we are 99% confident that my-feature is 1.32x to 1.37x faster than main" or "there is no statistically significant difference in performance between my-feature and main") for each benchmark Wasm program in the suite.

As you make further changes to your my-feature branch, you can execute this command whenever you want new, updated benchmark results:

$ cargo build --manifest-path ~/wasmtime/Cargo.toml --release -p wasmtime-bench-api && \
    cargo run --manifest-path ~/sightglass/Cargo.toml -- \
      benchmark \
      --engine /tmp/wasmtime_main.so \
      --engine ~/wasmtime/target/release/libwasmtime_bench_api.so \
      -- \
      benchmarks/*/benchmark.wasm

Collecting Different Kinds of Results

Sightglass comes enabled with several different kinds of measurement mechanisms — a measure. The default measure is cycles, which simply measures the elapsed duration of CPU cycles for each phase (e.g., using RDTSC). The accuracy of this measure is documented here but note that measuring using CPU cycles alone can be problematic (e.g., CPU frequency changes, context switches, etc.).

Several measures can be configured using the --measure option:

• cycles: the number of CPU cycles elapsed
• perf-counters: a selection of common perf counters (CPU cycles, instructions retired, cache accesses, cache misses); only available on Linux
• vtune: record each phase as a VTune task for analysis; see this help documentation for more details
• noop: no measurement is performed

For example, run:

$ cargo run -- benchmark --measure perf-counters ...

Getting Raw JSON or CSV Results

If you don't want the results to be summarized and displayed in a human-readable format, you can get raw JSON or CSV via the --raw flag:

$ cargo run -- benchmark --raw --output-format csv -- benchmarks/*/benchmark.wasm

Then you can use your own R/Python/spreadsheets/etc. to analyze and visualize the benchmark results.

Adding a New Benchmark

Add a Dockerfile under benchmarks/<your benchmark> building a Wasm file that brackets the work to measure with the bench.start and bench.end host calls. See the benchmarks README for a fuller set of requirements and the build.sh script for building this file.