Re: [Question] About XFS random buffer write performance

[Matthew Wilcox <willy@infradead.org>]

On Fri, Jul 31, 2020 at 12:05:58PM +1000, Dave Chinner wrote:
> On Fri, Jul 31, 2020 at 12:45:17AM +0100, Matthew Wilcox wrote:
> > Essentially, we do as you thought
> > it worked, we read the entire page (or at least the portion of it that
> > isn't going to be overwritten.  Once all the bytes have been transferred,
> > we can mark the page Uptodate.  We'll need to wait for the transfer to
> > happen if the write overlaps a block boundary, but we do that right now.
> 
> Right, we can do that, but it would be an entire page read, I think,
> because I see little point int doing two small IOs with a seek in
> between them when a single IO will do the entire thing much faster
> that two small IOs and put less IOP load on the disk. We still have
> to think about impact of IOs on spinning disks, unfortunately...

Heh, maybe don't read the existing code because we actually do that today
if, say, you have a write that spans bytes 800-3000 of a 4kB page.  Worse,
we wait for each one individually before submitting the next, so the
drive doesn't even get the chance to see that we're doing read-seek-read.

I think we can profitably skip reading portions of the page if the write
overlaps either the beginning or end of the page, but it's not worth
breaking up an I/O for skipping reading 2-3kB.

The readahead window expands up to 256kB, so clearly we are comfortable
with doing potentially-unnecessary reads of at least that much.  I start
to wonder about whether it might be worth skipping part of the page if
you do a 1MB write to the middle of a 2MB page, but the THP patchset
doesn't even try to allocate large pages in the write path yet, so the
question remains moot today.