🚀 Feature: Make parallel mode smarter

[boneskull]

I have a couple ideas for improvements here now that we're past the MVP:

1. Detect if we're only running a single file, and disable parallel mode automatically. Running a single file in parallel mode will always (AFAIK) be slower than running in serial. Add a command-line option to disable this behavior (a general-purpose, contextual --force might be helpful)
2. Automatic optimization via duration caching
   • Cache per file durations (cache could live in node_modules/.cache/mocha, which is an unofficial convention).
     • This would be timing the run() call in lib/nodejs/worker.js, from beginning to end.
     • Maybe calculate the mean over n most recent runs?
   • On subsequent runs, execute slowest test files first. This will help avoid the case at the end of the run where there's only a single worker process, munching on a meaty test file, and the other workers are idle. TypeScript uses a strategy like this in their custom tooling around Mocha (look at their implementation for ideas).

[boneskull]

(Note new parallel label)

[stoicchild8910]

@boneskull
hello, I read this issue and working on it and came up with two questions.

first, where to change.
I found mocha has two spots to turn on parallel function,

1. parallelMode function (in lib/mocha.js)
2. parallelrun (in lib/run-helpers)

Because parallelMode function comes first, I think it would be better to check and switch in parallelMode function if the file is better with running in single mode. I want to know your opinion.

second, the testing size
would it be enough only with a single file?
maybe two files, for three.. I also want to know your opinion on this!

thank you for your time.

[JoshuaKGoldberg]

This would be really nice for performance - other test runners such as Jest take similar strategies. Since parallel runs aren't very deterministic right now, I don't think it'd likely be received as a big breaking change for users. But slapping the semver-major label on just to be safe.