lkml.org 
[lkml]   [2022]   [Nov]   [23]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [f2fs-dev] [PATCH v3 14/23] f2fs: Convert f2fs_write_cache_pages() to use filemap_get_folios_tag()
On Tue, Nov 22, 2022 at 6:26 PM Vishal Moola <vishal.moola@gmail.com> wrote:
>
> On Mon, Nov 14, 2022 at 1:38 PM Vishal Moola <vishal.moola@gmail.com> wrote:
> >
> > On Sun, Nov 13, 2022 at 11:02 PM Chao Yu <chao@kernel.org> wrote:
> > >
> > > On 2022/10/18 4:24, Vishal Moola (Oracle) wrote:
> > > > Converted the function to use a folio_batch instead of pagevec. This is in
> > > > preparation for the removal of find_get_pages_range_tag().
> > > >
> > > > Also modified f2fs_all_cluster_page_ready to take in a folio_batch instead
> > > > of pagevec. This does NOT support large folios. The function currently
> > >
> > > Vishal,
> > >
> > > It looks this patch tries to revert Fengnan's change:
> > >
> > > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=01fc4b9a6ed8eacb64e5609bab7ac963e1c7e486
> > >
> > > How about doing some tests to evaluate its performance effect?
> >
> > Yeah I'll play around with it to see how much of a difference it makes.
>
> I did some testing. Looks like reverting Fengnan's change allows for
> occasional, but significant, spikes in write latency. I'll work on a variation
> of the patch that maintains the use of F2FS_ONSTACK_PAGES and send
> that in the next version of the patch series. Thanks for pointing that out!

Here are some numbers for reference to performance. I'm thinking we may
want to go with the new version, but I'll let you be the judge of that.
I ran some fio random write tests with block size 64k on a system with 8 cpus.

1 job with 1 io-depth:
Baseline:
slat (usec): min=8, max=849, avg=16.47, stdev=12.33
clat (nsec): min=253, max=751838, avg=346.51, stdev=2452.10
lat (usec): min=9, max=854, avg=17.00, stdev=12.74
lat (nsec) : 500=97.09%, 750=1.73%, 1000=0.57%
lat (usec) : 2=0.41%, 4=0.09%, 10=0.06%, 20=0.04%, 50=0.01%
lat (usec) : 100=0.01%, 1000=0.01%

This patch:
slat (usec): min=9, max=3690, avg=16.61, stdev=17.36
clat (nsec): min=28, max=380434, avg=336.59, stdev=1571.23
lat (usec): min=10, max=3699, avg=17.13, stdev=17.51
lat (nsec) : 50=0.01%, 500=97.95%, 750=1.42%, 1000=0.33%
lat (usec) : 2=0.19%, 4=0.05%, 10=0.03%, 20=0.03%, 50=0.01%
lat (usec) : 100=0.01%, 250=0.01%, 500=0.01%

Folios w/ F2FS_ONSTACK_PAGES (next version):
slat (usec): min=12, max=13623, avg=19.48, stdev=48.94
clat (nsec): min=265, max=386917, avg=380.97, stdev=1679.85
lat (usec): min=12, max=13635, avg=20.06, stdev=49.27
lat (nsec) : 500=93.55%, 750=4.62%, 1000=0.92%
lat (usec) : 2=0.65%, 4=0.09%, 10=0.10%, 20=0.06%, 50=0.01%
lat (usec) : 100=0.01%, 250=0.01%, 500=0.01%

1 job with 16 io-depth:
Baseline:
slat (usec): min=8, max=3907, avg=16.89, stdev=23.39
clat (usec): min=12, max=15160k, avg=11115.61, stdev=265051.86
lat (usec): min=137, max=15160k, avg=11132.68, stdev=265051.75
lat (usec) : 20=0.01%, 250=57.66%, 500=39.56%, 750=1.96%, 1000=0.22%
lat (msec) : 2=0.16%, 4=0.06%, 10=0.01%, 2000=0.29%, >=2000=0.08%

This patch:
slat (usec): min=9, max=1230, avg=17.15, stdev=12.95
clat (usec): min=4, max=39471k, avg=14825.22, stdev=588237.30
lat (usec): min=80, max=39471k, avg=14842.55, stdev=588237.27
lat (usec) : 10=0.01%, 250=38.78%, 500=59.53%, 750=1.12%, 1000=0.16%
lat (msec) : 2=0.04%, 2000=0.34%, >=2000=0.02%

Folios w/ F2FS_ONSTACK_PAGES (next version):
slat (usec): min=9, max=1188, avg=18.74, stdev=14.12
clat (usec): min=5, max=15278k, avg=8936.75, stdev=214230.09
lat (usec): min=90, max=15278k, avg=8955.67, stdev=214230.10
lat (usec) : 10=0.01%, 250=9.68%, 500=86.49%, 750=2.74%, 1000=0.54%
lat (msec) : 2=0.18%, 2000=0.32%, >=2000=0.04%


> How do the remaining f2fs patches in the series look to you?
> Patch 16/23 f2fs_sync_meta_pages() in particular seems like it may
> be prone to problems. If there are any changes that need to be made to
> it I can include those in the next version as well.

\
 
 \ /
  Last update: 2022-11-23 08:56    [W:0.796 / U:0.032 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site