Hi,
On 2022-03-10 15:43:16 +1300, Thomas Munro wrote:
> Wow, I see the effect on Cirrus -- test_world ran in 8:55 instead of
> 12:43 when I tried (terrible absolute times, but fantastic
> improvement!). Hmm, on my local FreeBSD 13 box I saw 5:07 -> 5:03
> with this change. My working theory had been that there is something
> bad happening in the I/O stack under concurrency making FreeBSD on
> Cirrus/GCP very slow (ie patterns to stall on slow cloud I/O waits,
> something I hope to dig into when post-freeze round tuits present
> themselves), but that doesn't gel with this huge improvement from
> noodling with optimiser details, and I don't know why I don't see
> something similar locally. I'm confused.
The "terrible IO wait" thing was before we reduced the number of CPUs and
concurrent jobs. It makes sense to me that with just two CPUs we're CPU bound,
in which case -Og obviously can make a difference.
> Just BTW it's kinda funny that we say -ggdb for macOS and then we use
> lldb to debug cores in cores_backtrace.sh. I suppose it would be more
> correct to say -glldb, but doubt it matters much...
Yea. I used -ggdb because I didn't know -glldb existed :). And there's also
the advantage that -ggdb works both with gcc and clang, whereas -glldb doesn't
seem to be known to gcc.
Greetings,
Andres Freund
