[docs] add MathOptInterface.jl to the list of packages

[odow]

Hey!

I started using this in https://github.com/jump-dev/MathOptInterface.jl

Some may find our workflow interesting.

There are a lot of tests in MOI, and it is almost entirely bottlenecked by compilation time. (I should probably experiment with Base.Experimental.@optlevel.)

We shard the tests into directories in /test, and then use ParallelTestRunner to parallelise over the files in each directory.

Each file is a module that automatically detects and runs all of the functions named test_.

I find this works much better than @testset, where I kept leaking stuff into setup and global scope and it was hard to maintain.

Now adding a test is: choose a file, add a function.

You can improve parallelisation by splitting tests into a new file

The biggest downside is that it's easy to add two functions with the same name that overwrite each other. So there's a test to avoid that:
https://github.com/jump-dev/MathOptInterface.jl/blob/ea3cf8ca0060d776bb351fe4f7e1698d308f105c/test/General/test_errors.jl#L430-L453

[giordano]

Thanks!

[vchuravy]

What is the reason for using PTR only on 64bit?

https://github.com/jump-dev/MathOptInterface.jl/blob/c5744855c655f1679acb88f1219feaf36887b214/test/runtests.jl#L49-L51

[odow]

On 32-bit it parallelised only over a single thread. Forcing more made it much slower. And it was faster still to not use PTR

[giordano]

On 32-bit it parallelised only over a single thread.

You can always pass --jobs to force however many parallel jobs you want. That's what I usually do on macOS: https://github.com/NumericalEarth/Breeze.jl/blob/63af70c3fbb3934a57467ff03f74e2a4156134bf/.github/workflows/CI.yml#L96-L100. But it's not clear to me why the default would be one with a 32-bit Julia, maybe it's because available_memory() in

ParallelTestRunner.jl/src/ParallelTestRunner.jl

Lines 390 to 394 in 10225a7

	function default_njobs(; cpu_threads = Sys.CPU_THREADS, free_memory = available_memory())
	jobs = cpu_threads
	memory_jobs = Int64(free_memory) ÷ (2 * 2^30)
	return max(1, min(jobs, memory_jobs))
	end

reports 2 GB?

[giordano]

Ah, this is hilarious, the problem is

ParallelTestRunner.jl/src/ParallelTestRunner.jl

Line 392 in 10225a7

memory_jobs = Int64(free_memory) ÷ (2 * 2^30)

julia> 2 * 2^30
-2147483648

This needs to be fixed 😃

[giordano]

Forcing more made it much slower.

Wait, I just noted this: that's surprising: there's some extra compilation latency involved when using separate parallel processes, but why would the overall speedup be substantially different between 32- and 64-bit julia builds? 🤔

[odow]

Nah I meant: given PTR used one thread, then there is overhead to using PTR instead of not. There's a similar overhead in 64-bit, but the parallelism was a win.