Re: Optimizing small reads

[Linus Torvalds <torvalds@linux-foundation.org>]

On Mon, 13 Oct 2025 at 08:35, Kiryl Shutsemau <kirill@shutemov.name> wrote:
>
> The patch is below. Can I use your Signed-off-by for it?

Sure. But you can also just take ownership.

> Here's some numbers. I cannot explain some of this. Like, there should
> be zero changes for block size above 256, but...

I'd assume it's the usual "random cache/code placement" thing. The
numbers seem reasonably consistent, and guessing from your table you
did basically a "random file / random offset" thing.

And even if you go "but with random files and random offsets, cache
placement should be random too", that's not true for the I$ side.

So you could have some unlucky placement of the helper routines (like
the iov stuff) that hits one case but not the other.

> 16 threads, reads from 4k file(s), MiB/s
>
>  ---------------------------------------------------------
> | Block | Baseline  | Baseline   | Patched   |  Patched   |
> | size  | same file | diff files | same file | diff files |
>  ---------------------------------------------------------
> |     1 |      11.6 |       27.6 |      33.5 |       33.4 |
> |    32 |     375   |      880   |    1027   |     1028   |
> |   256 |    2940   |     6932   |    7872   |     7884   |
> |  1024 |   11500   |    26900   |   11400   |    28300   |
> |  4096 |   46500   |   103000   |   45700   |   108000   |

The high-level pattern is fairly clear, with the big improvement being
in the "small reads hitting same file", so at least the numbers don't
look crazy.

                    Linus