2021-03-22 12:07:50

by Jesper Dangaard Brouer

[permalink] [raw]
Subject: Re: [PATCH 0/3 v5] Introduce a bulk order-0 page allocator

On Mon, 22 Mar 2021 09:18:42 +0000
Mel Gorman <[email protected]> wrote:

> This series is based on top of Matthew Wilcox's series "Rationalise
> __alloc_pages wrapper" and does not apply to 5.12-rc2. If you want to
> test and are not using Andrew's tree as a baseline, I suggest using the
> following git tree
>
> git://git.kernel.org/pub/scm/linux/kernel/git/mel/linux.git mm-bulk-rebase-v5r9

page_bench04_bulk[1] micro-benchmark on branch: mm-bulk-rebase-v5r9
[1] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/mm/bench/page_bench04_bulk.c

BASELINE
single_page alloc+put: Per elem: 199 cycles(tsc) 55.472 ns

LIST variant: time_bulk_page_alloc_free_list: step=bulk size

Per elem: 206 cycles(tsc) 57.478 ns (step:1)
Per elem: 154 cycles(tsc) 42.861 ns (step:2)
Per elem: 145 cycles(tsc) 40.536 ns (step:3)
Per elem: 142 cycles(tsc) 39.477 ns (step:4)
Per elem: 142 cycles(tsc) 39.610 ns (step:8)
Per elem: 137 cycles(tsc) 38.155 ns (step:16)
Per elem: 135 cycles(tsc) 37.739 ns (step:32)
Per elem: 134 cycles(tsc) 37.282 ns (step:64)
Per elem: 133 cycles(tsc) 36.993 ns (step:128)

ARRAY variant: time_bulk_page_alloc_free_array: step=bulk size

Per elem: 202 cycles(tsc) 56.383 ns (step:1)
Per elem: 144 cycles(tsc) 40.047 ns (step:2)
Per elem: 134 cycles(tsc) 37.339 ns (step:3)
Per elem: 128 cycles(tsc) 35.578 ns (step:4)
Per elem: 120 cycles(tsc) 33.592 ns (step:8)
Per elem: 116 cycles(tsc) 32.362 ns (step:16)
Per elem: 113 cycles(tsc) 31.476 ns (step:32)
Per elem: 110 cycles(tsc) 30.633 ns (step:64)
Per elem: 110 cycles(tsc) 30.596 ns (step:128)

Compared to the previous results (see below) list-variant got faster,
but array-variant is still faster. The array variant lost a little
performance. I think this can be related to the stats counters got
added/moved inside the loop, in this patchset.

Previous results from:
https://lore.kernel.org/netdev/20210319181031.44dd3113@carbon/

On Fri, 19 Mar 2021 18:10:31 +0100 Jesper Dangaard Brouer <[email protected]> wrote:

> BASELINE
> single_page alloc+put: 207 cycles(tsc) 57.773 ns
>
> LIST variant: time_bulk_page_alloc_free_list: step=bulk size
>
> Per elem: 294 cycles(tsc) 81.866 ns (step:1)
> Per elem: 214 cycles(tsc) 59.477 ns (step:2)
> Per elem: 199 cycles(tsc) 55.504 ns (step:3)
> Per elem: 192 cycles(tsc) 53.489 ns (step:4)
> Per elem: 188 cycles(tsc) 52.456 ns (step:8)
> Per elem: 184 cycles(tsc) 51.346 ns (step:16)
> Per elem: 183 cycles(tsc) 50.883 ns (step:32)
> Per elem: 184 cycles(tsc) 51.236 ns (step:64)
> Per elem: 189 cycles(tsc) 52.620 ns (step:128)
>
> ARRAY variant: time_bulk_page_alloc_free_array: step=bulk size
>
> Per elem: 195 cycles(tsc) 54.174 ns (step:1)
> Per elem: 123 cycles(tsc) 34.224 ns (step:2)
> Per elem: 113 cycles(tsc) 31.430 ns (step:3)
> Per elem: 108 cycles(tsc) 30.003 ns (step:4)
> Per elem: 102 cycles(tsc) 28.417 ns (step:8)
> Per elem: 98 cycles(tsc) 27.475 ns (step:16)
> Per elem: 96 cycles(tsc) 26.901 ns (step:32)
> Per elem: 95 cycles(tsc) 26.487 ns (step:64)
> Per elem: 94 cycles(tsc) 26.170 ns (step:128)

> The users of the API have been dropped in this version as the callers
> need to check whether they prefer an array or list interface (whether
> preference is based on convenience or performance).

I'll start coding up the page_pool API user and benchmark that.

> Changelog since v4
> o Drop users of the API
> o Remove free_pages_bulk interface, no users

In [1] benchmark I just open-coded free_pages_bulk():
https://github.com/netoptimizer/prototype-kernel/commit/49d224b19850b767c

> o Add array interface
> o Allocate single page if watermark checks on local zones fail

--
Best regards,
Jesper Dangaard Brouer
MSc.CS, Principal Kernel Engineer at Red Hat
LinkedIn: http://www.linkedin.com/in/brouer