tree: d77c499209bd747efa664e25ff4b2450d3f45d89 [path history] [tgz]
  1. benchmarks/
  2. cli/
  3. cmd/
  4. common/
  5. generators/
  6. harnesses/
  7. source-assets/
  8. assets.hash

Sweet: Benchmarking Suite for Go Implementations

Sweet is a set of benchmarks derived from the Go community which are intended to represent a breadth of real-world applications. The primary use-case of this suite is to perform an evaluation of the difference in CPU and memory performance between two Go implementations.

If you use this benchmarking suite for any measurements, please ensure you use a versioned release and note the version in the release.


Supported Platforms

  • Linux
  • TODO(mknyszek): Support more.


The sweet tool only depends on having a stable version of Go and git. If you‘re testing out a development version of Go with Sweet, don’t build sweet itself with it! Use what's installed on your system instead, or some known-stable version.

Some benchmarks, however, have various requirements for building. Notably they are:

  • make (tile38)
  • bash (tile38)
  • binutils (tile38)

Please ensure your system has these tools installed and available in your system's PATH.

Furthermore, some benchmarks are able to produce additional information on some platforms. For instance, running on platforms where systemd is available adds an average RSS measurement for the go-build benchmark.


The gVisor benchmark has additional requirements:

  • The target platform must be linux/amd64. Nothing else is supported or ever will be.
  • The ptrace API must be enabled on your system. Set /proc/sys/kernel/yama/ptrace_scope appropriately (0 and 1 work, 2 might, 3 will not).


$ go build ./cmd/sweet

Getting Assets

$ ./sweet get -auth=app-default

Note: make sure you run gcloud auth application-default login before you run the above command!

TODO(mknyszek): Make it so authentication isn't necessary in the future.

Running the benchmarks

Create a configuration file called config.toml with the following contents:

  name = "myconfig"
  goroot = "<insert some GOROOT here>"

Run the benchmarks by running:

$ ./sweet run -shell config.toml

Benchmark results will appear in the results directory.

-shell will cause the tool to print each action it performs as a shell command. Note that while the shell commands are valid for many systems, they may depend on tools being available on your system that sweet does not require (e.g. git).

Tips and Rules of Thumb

  • You can expect the benchmarks to take a few hours to run with the default settings on a somewhat sizable Linux box.
  • If a benchmark fails to build, run with -shell and copy and re-run the command to get output. TODO(mknyszek): Dump the output to the terminal.
  • If a benchmark fails to run, the output should have been captured in the corresponding results file (e.g. if biogo-igor failed, check /path/to/results/biogo-igor/myconfig.results) which is really just the stderr (and usually stdout too) of the benchmark. You can also try to re-run it yourself with the output of -shell.

Memory Requirements

These benchmarks generally try to stress the Go runtime in interesting ways, and some may end up with very large heaps. Therefore, it's recommended to run the suite on a system with at least 30 GiB of RAM available to minimize the chance that results are lost due to an out-of-memory error.

Configuration Format

The configuration is TOML-based and a more detailed description of fields may be found in the help docs for the run subcommand:

$ ./sweet help run

Results Format

Results are produced into a single directory containing each benchmark as a sub-directory. Within each sub-directory is one file per configuration containing the stderr (and usually combined stdout) of the benchmark run, which also doubles as the benchmark output format.

All results are reported in the standard Go testing package format, such that results may be compared using the benchstat tool.

Results then may also be composed together for easy viewing. For example, if one runs sweet with two configurations named config1 and config2, then to quickly compare all results, do:

$ cat results/*/config1.results > config1.results
$ cat results/*/config2.results > config2.results
$ benchstat config1.results config2.results


This benchmark suite tries to keep noise low in measurements where possible.

  • Each measurement is taken against a fresh OS process.
  • Benchmarks have been modified to reduce noise from the input.
    • All inputs are deterministic, including implicit inputs, such as querying an existing database.
    • Inputs are loaded into memory when possible instead of streamed from disk.
  • The suite mitigates external effects we can control (e.g. the suite is aware of its co-tenancy with the benchmark and throttles itself when the benchmarks are running).

Tips for Reducing Noise

  • Sweet should be run on a system where a perflock daemon is running (to avoid noise due to CPU throttling).
  • Avoid running these benchmarks in cloud environments if possible. Generally the noise inherent to those environments can skew A/B tests and hide small changes in performance. See this paper for more details. Try to use dedicated hardware instead.

Do not compare results produced by separate invocations of the sweet tool.


With the current release there are a few notable caveats when it comes to the results and the noisiness of the benchmarks that will addressed with future releases. Notably:

  • Time for the tile38 benchmarks is quite noisy, but at least consistent.
  • RSS numbers for the gopher-lua and biogo-igor benchmarks are quite noisy.
  • Peak RSS is currently not very reliable.
  • The gVisor startup benchmarks are very noisy.