Re: slowest tap tests - split or accelerate?

Hi,

On 2022-01-18 13:40:40 -0800, Andres Freund wrote:
> Maybe we really should do at least the most simplistic caching for initdbs, by
> doing one initdb as part of the creation of temp_install. Then Cluster::init
> would need logic to only use that if $params{extra} is empty.

I hacked this together. And the wins are bigger than I thought. On my
workstation, with plenty cpu and storage bandwidth, according to
/usr/bin/time check-world NO_TEMP_INSTALL=1
things go from

321.56user 74.00system 2:26.22elapsed 270%CPU (0avgtext+0avgdata 93768maxresident)k
24inputs+32781336outputs (2254major+8717121minor)pagefaults 0swaps

to

86.62user 57.10system 1:57.83elapsed 121%CPU (0avgtext+0avgdata 93752maxresident)k
8inputs+32683408outputs (1360major+6672618minor)pagefaults 0swaps

The difference in elapsed and system time is pretty good, but the user time
difference is quite staggering.

This doesn't yet actually address the case of the basebackup tests, because
that specifies a "non-default" option, preventing the use of the template
initdb. But the effects are already big enough that I thought it's worth
sharing.

On CI for windows this reduces the time for the subscription tests from 03:24,
to 2:39. There's some run-to-run variation, but it's a pretty clear signal...

Greetings,

Andres Freund