On Tue, Jan 21, 2020 at 8:11 AM Andres Freund <andres@anarazel.de> wrote:
> FWIW, I think we should just flat out delete all this logic, and replace
> it with a few explicit PrefetchBuffer() calls. Just by chance I
> literally just now sped up a VACUUM by more than a factor of 10, by
> manually prefetching buffers. At least the linux kernel readahead logic
> doesn't deal well with reading and writing to different locations in the
> same file, and that's what the ringbuffer pretty invariably leads to for
> workloads that aren't cached.
Interesting. Andrew Gierth made a similar observation on FreeBSD, and
showed that by patching his kernel to track sequential writes and
sequential reads separately he could improve performance, and I
reproduced the same speedup in a patch of my own based on his
description (that, erm, I've lost). It's not only VACUUM, it's
anything that is writing to a lot of sequential blocks, since the
writeback trails along behind by some distance (maybe a ring buffer,
maybe all of shared buffers, whatever). The OS sees you flipping back
and forth between single block reads and writes and thinks it's
random. I didn't investigate this much but it seemed that ZFS was
somehow smart enough to understand what was happening at some level
but other filesystems were not.
