Add benchmarks for json string serialize/deserialize

[paradust7]

• Pulls the Catch2 testing and benchmarking library into lib/catch2 (a single header file)
• Adds --run-benchmarks command-line flag
• Adds BUILD_BENCHMARKS flag which is FALSE by default.
• Adds src/benchmark directory
• Adds benchmark_serialize.cpp which contains benchmarks of serializeJsonStringIfNeeded and deSerializeJsonStringIfNeeded
• Adds .github/workflows/perf.yml which runs the benchmarks and stores them to the github pages branch. (*)

(*) I have no way of testing this, so it may be better to use the web interface to create it. It assumes the existence of a gh-pages branch specially created as described here: https://github.com/benchmark-action/github-action-benchmark

[paradust7]

There's an embarassing bug in the workflow alerter. It ignores time units. So "1 ms" compared to "900 us" looks like a performance regression of 900x:

paradust7@28ed9bf#commitcomment-72699948

I'm going to disable alerts for now, and hope this is the only bug.

[sfan5]

Example page with results: https://sfan5.github.io/minetest/dev/bench/

[rubenwardy]

Example page with results: sfan5.github.io/minetest/dev/bench

That's pretty cool

[sfan5]

The graphs are junk though, this is not usable.

[paradust7]

Sad. It's strange that it has no problem parsing the time units, but does nothing with them.

We could fork github-action-benchmark to fix this, but it would be probably be easier to just have catch2 always emit the same unit for now.

EDIT: I filed an issue for them: benchmark-action/github-action-benchmark#122

[paradust7]

Can you give this action write permissions only on gh-pages? I didn't realize it was a third-party action. Maybe forking it is a good idea, or revision pinning.

[sfan5]

How would I do that? It worked out of the box on my personal repo (after creating the branch).

[paradust7]

Unfortunately, I don't see any way to limit it. The workflow gives it write permission for "deployments" and "contents" access groups. These are documented in terms of API access. "Contents" allows it to generate a pull request and merge it. But I don't see any way to limit it to a specific branch. There's a "pages" API group but it seems to be read-only.

https://docs.github.com/en/rest/overview/permissions-required-for-github-apps

If the action is pinned to a specific revision, at least that will guarantee that future changes to upstream will not introduce a sudden surprise. But won't help if there's already a backdoor or exploit in it. (very unlikely, it's only 4529 lines, but still...)

Alternatively, could put the workflow in a completely separate repository, e.g. minetest/benchmarks, which has its own gh-pages, and keeps sync with a cron workflow. I'll try to do this and see if there are any snags.

[paradust7]

@sfan5 I got it working successfully from another repo.

Can you create a "benchmarks" repo (or other appropriate name) under the minetest organization, so I can make a PR against it? Thanks.

The "benchmarks" repo will also need a gh-pages branch to work. One will be created automatically if you setup a theme, as per these instructions: https://docs.github.com/en/pages/getting-started-with-github-pages/adding-a-theme-to-your-github-pages-site-with-the-theme-chooser

The graphs will then be available at https://minetest.github.io/benchmarks/dev/bench

[paradust7]

I removed the github workflow from this PR. It will instead be in the benchmarks repo.
This PR still needs to go in to add catch2, --run-benchmarks, etc.

[sfan5]

Sounds good. How will the other workflow track commits in the main repo? Does Github Actions have something for that?
Or will it just run e.g. every 12 hours?

Can you create a "benchmarks" repo (or other appropriate name) under the minetest organization, so I can make a PR against it?

I can't do this myself but I'll get someone who can to do it.

[rubenwardy]

Here you are: https://github.com/minetest/benchmarks

[sfan5]

Looks good.

[rubenwardy]

Looks good apart from that indentation issue. Haven't tested, will trust sfan5 on that

[paradust7]

Thanks! Here's the PR on the benchmarks repo:

luanti-org/benchmarks#1