cargo/benches/README.md
2021-10-12 13:23:56 -07:00

4.9 KiB

Cargo Benchmarking

This directory contains some benchmarks for cargo itself. This uses Criterion for running benchmarks. It is recommended to read the Criterion book to get familiar with how to use it. A basic usage would be:

cd benches/benchsuite
cargo bench

The tests involve downloading the index and benchmarking against some real-world and artificial workspaces located in the workspaces directory.

Beware that the initial download can take a fairly long amount of time (10 minutes minimum on an extremely fast network) and require significant disk space (around 4.5GB). The benchsuite will cache the index and downloaded crates in the target/tmp/bench directory, so subsequent runs should be faster. You can (and probably should) specify individual benchmarks to run to narrow it down to a more reasonable set, for example:

cargo bench -- resolve_ws/rust

This will only download what's necessary for the rust-lang/rust workspace (which is about 330MB) and run the benchmarks against it (which should take about a minute). To get a list of all the benchmarks, run:

cargo bench -- --list

Viewing reports

The benchmarks display some basic information on the command-line while they run. A more complete HTML report can be found at target/criterion/report/index.html which contains links to all the benchmarks and summaries. Check out the Criterion book for more information on the extensive reporting capabilities.

Comparing implementations

Knowing the raw numbers can be useful, but what you're probably most interested in is checking if your changes help or hurt performance. To do that, you need to run the benchmarks multiple times.

First, run the benchmarks from the master branch of cargo without any changes. To make it easier to compare, Criterion supports naming the baseline so that you can iterate on your code and compare against it multiple times.

cargo bench -- --save-baseline master

Now you can switch to your branch with your changes. Re-run the benchmarks compared against the baseline:

cargo bench -- --baseline master

You can repeat the last command as you make changes to re-compare against the master baseline.

Without the baseline arguments, it will compare against the last run, which can be helpful for comparing incremental changes.

Capturing workspaces

The workspaces directory contains several workspaces that provide a variety of different workspaces intended to provide good exercises for benchmarks. Some of these are shadow copies of real-world workspaces. This is done with the tool in the capture directory. The tool will copy Cargo.lock and all of the Cargo.toml files of the workspace members. It also adds an empty lib.rs so Cargo won't error, and sanitizes the Cargo.toml to some degree, removing unwanted elements. Finally, it compresses everything into a tgz.

To run it, do:

cd benches/capture
cargo run -- /path/to/workspace/foo

The resolver benchmarks also support the CARGO_BENCH_WORKSPACES environment variable, which you can point to a Cargo workspace if you want to try different workspaces. For example:

CARGO_BENCH_WORKSPACES=/path/to/some/workspace cargo bench

TODO

This is just a start for establishing a benchmarking suite for Cargo. There's a lot that can be added. Some ideas:

  • Fix the benchmarks so that the resolver setup doesn't run every iteration.
  • Benchmark this section of code which builds the unit graph. The performance there isn't great, and it would be good to keep an eye on it. Unfortunately that would mean doing a bit of work to make generate_targets publicly visible, and there is a bunch of setup code that may need to be duplicated.
  • Benchmark the fingerprinting code.
  • Benchmark running the cargo executable. Running something like cargo build or cargo check with everything "Fresh" would be a good end-to-end exercise to measure the overall overhead of Cargo.
  • Benchmark pathological resolver scenarios. There might be some cases where the resolver can spend a significant amount of time. It would be good to identify if these exist, and create benchmarks for them. This may require creating an artificial index, similar to the resolver-tests. This should also consider scenarios where the resolver ultimately fails.
  • Benchmark without Cargo.lock. I'm not sure if this is particularly valuable, since we are mostly concerned with incremental builds which will always have a lock file.
  • Benchmark just resolve::resolve without anything else. This can help focus on just the resolver.