RE: [PATCH v3 2/5] benchtests: Add memset zero fill benchtest

[naohirot--- via Libc-alpha <libc-alpha@sourceware.org>]

Hi Wilco,

Thank you for the comment.
I understood your concerns about the start size and the end size.

> From: Wilco Dijkstra <Wilco.Dijkstra@arm.com>
> Sent: Tuesday, September 14, 2021 11:03 PM
>
> I had a quick go at running the new benchmark. The main problem is that it doesn't
> give repeatable results - there are huge variations from run to run of about 50% for
> the smaller sizes. This is a fundamental problem due to the timing loop, and the only
> way to reduce it is to increase the time taken by memset, ie. start at a much larger
> size (say at 16KB).

In terms of the start size, 256B is chosen because __memset_generic
(sysdeps/aarch64/memset.S) calls DC ZVA for zero fill from 256B, which
code you committed [1].
And I reported an interesting insight in the mail [2] that DC ZVA is
slower than store instruction from 256B to 16KB on A64FX [3].
So it seems valuable to measure the range from 256B to 16KB to see
each CPU's behavior.
What do you think?

[1] https://sourceware.org/git/?p=glibc.git&h=a8c5a2a9521e105da6e96eaf4029b8e4d595e4f5
[2] https://sourceware.org/pipermail/libc-alpha/2021-August/129805.html
[3] https://drive.google.com/file/d/1fonjDDlF4LPLfZY9-z22DGn-yaSpGN4g/view

> It also takes a long time to run - generally it's best to ensure a benchmark takes less
> than 10 seconds on a typical modern system (remember there will be many that are
> slower!). It should be feasible to reduce the iteration count for large sizes, but you
> could go up to 16MB rather than 64MB.

OK, I'll change the end size to 16MB.

Thanks.
Naohiro