Re: slowest tap tests - split or accelerate?

Hi,

On 2021-12-31 11:25:28 -0800, Andres Freund wrote:
> cfbot now runs most tests on windows, the windows task is by far the slowest,
> and the task limitted most in concurrency [2]. Running tap tests is the
> biggest part of that. This is a bigger issue on windows because we don't have
> infrastructure (yet) to run tests in parallel.
> 
> There's a few tests which stand out in their slowness, which seem worth
> addressing even if we tackle test parallelism on windows at some point. I
> often find them to be the slowest tests on linux too.
> 
> Picking a random successful cfbot run [1] I see the following tap tests taking
> more than 20 seconds:
> 
> 67188 ms pg_basebackup t/010_pg_basebackup.pl
> 25751 ms pg_verifybackup t/002_algorithm.pl

The reason these in particular are slow is that they do a lot of
pg_basebackups without either / one-of -cfast / --no-sync. The lack of -cfast
in particularly is responsible for a significant proportion of the test
time. The only reason this didn't cause the tests to take many minutes is that
spread checkpoints only throttle when writing out a buffer and there aren't
that many dirty buffers...

Attached is a patch changing the parameters in all the instances I
found. Testing on a local instance it about halves the runtime of
t/010_pg_basebackup.pl on linux and windows (but there's still a 2x time
difference between the two), it's less when running the tests concurrently CI.

It might be worth having one explicit use of -cspread. Perhaps combined with
an explicit checkpoint beforehand?

Greetings,

Andres Freund