On Wed, 2025-09-03 at 10:34 +0200, Tomas Vondra wrote:
> I'll try to reproduce this, but if it's due to the same IPC overhead,
> that would be surprising (for me). In the index case it makes sense,
> because the reads are random enough to prevent I/O combining. But for
> a
> sequential workload I'd expect I/O combining to help. Could it be
> that
> it ends up evicting buffers randomly, which (I guess) might interfere
> with the combining? What's shared_buffers set to?
I left it as the default (128MB).
> Have you watched how
> large the I/O requests are? iostat, iosnoop or strace would tell you.
In my test there's not much device IO, because the data is cached. I
used:
strace -p $io_worker_pid -e trace=preadv
and looked briefly at the output. iov_len seems to range between 8kB
and about 128kB, but still a lot of 8kB. That's from a very brief look,
I can try to get more precise numbers, but there seem to be enough 8kB
ones to support your theory.
Regards,
Jeff Davis
