lkml.org 
[lkml]   [2009]   [Jun]   [10]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: Performance regressions in 2.6.30-rc7?
On Tue 09-06-09 14:48:18, Chris Mason wrote:
> On Tue, Jun 09, 2009 at 12:32:08PM +0200, Jan Kara wrote:
> > On Thu 04-06-09 21:13:15, Mike Galbraith wrote:
> > > On Thu, 2009-06-04 at 13:21 +0200, Jan Kara wrote:
> > >
> > > > > Sequential Writes
> > > > > 2.6.30-smp-ordered 6000 65536 32 50.16 508.9% 31.996 45595.78 0.64965 0.02402 10
> > > > > 2.6.29.4-smp-ordered 6000 65536 32 52.70 543.2% 33.658 23794.92 0.71754 0.00836 10
> > > > >
> > > > > 2.6.30-smp-writeback 6000 65536 32 47.82 525.4% 35.003 32588.84 0.56192 0.02298 9
> > > > > 2.6.29.4-smp-writeback 6000 65536 32 52.52 467.6% 32.397 12972.78 0.53580 0.00522 11
> > > > >
> > > > > 2.6.30-smp-ordered 6000 65536 16 56.08 254.9% 15.463 33000.68 0.39687 0.00521 22
> > > > > 2.6.29.4-smp-ordered 6000 65536 16 62.40 308.4% 14.701 13455.02 0.13125 0.00208 20
> > > > >
> > > > > 2.6.30-smp-writeback 6000 65536 16 51.90 281.4% 17.098 12869.85 0.36771 0.00104 18
> > > > > 2.6.29.4-smp-writeback 6000 65536 16 60.53 272.6% 14.977 8637.08 0.21146 0.00000 22
> > > > >
> > > > > 2.6.30-smp-ordered 6000 65536 8 51.09 113.4% 8.700 14856.55 0.06771 0.00417 45
> > > > > 2.6.29.4-smp-ordered 6000 65536 8 56.13 130.6% 8.098 8400.45 0.03958 0.00000 43
> > > > >
> > > > > 2.6.30-smp-writeback 6000 65536 8 50.19 131.7% 8.680 16821.04 0.11979 0.00208 38
> > > > > 2.6.29.4-smp-writeback 6000 65536 8 54.90 130.7% 8.244 4925.48 0.10000 0.00000 42
> > > > It really seems write has some problems... There's consistently lower
> > > > throughput and it also seems some writes take really long. I'll try to
> > > > reproduce it here.
> > >
> > > Looked "pretty solid" to me. I haven't observed enough to ~trust.
> > OK, I did a few runs of tiobench here and I can confirm that I see about
> > 6% performance regression in Sequential Write throughput between 2.6.29
> > and 2.6.30-rc8. I'll try to find what's causing it.
>
> My first guess would be the WRITE_SYNC style changes. Is the regression
> still there with noop?
Thanks for the hint. I was guessing that as well. And experiments show
it's definitely connected. To be more precise with the data:
The test machine is 2 CPU, 2 GB ram, simple lowend SATA disk. Tiobench run
with:
tiobench/tiobench.pl -b 65536 -t 16 -t 8 -d /local/scratch -s 4096
which means 4GB testfile, writes happen in 64k chunks, test done with 16
and 8 threads. /local/scratch is a separate partition always cleaned and
umounted + mounted before each test. The results are (always 3 runs):
2.6.29+CFQ: Avg StdDev
8 38.01 40.26 39.69 -> 39.32 0.955092
16 40.09 38.18 40.05 -> 39.44 0.891104

2.6.30-rc8+CFQ:
8 36.67 36.81 38.20 -> 37.23 0.69062
16 37.45 36.47 37.46 -> 37.13 0.464351

2.6.29+NOOP:
8 38.67 38.66 37.55 -> 38.29 0.525632
16 39.59 39.15 39.19 -> 39.31 0.198662

2.6.30-rc8+NOOP:
8 38.31 38.47 38.16 -> 38.31 0.126579
16 39.08 39.25 39.13 -> 39.15 0.0713364

So with CFQ there is a statistically meaningful difference and with NOOP
there is not.
I've also tried plain simple
dd if=/dev/zero of=/local/scratch bs=65536 count=50k
which gives ~3.6GB file. Also here are noticeable differences alhough
smaller:
2.6.29+CFQ: Avg StdDev
47.5 48.2 48.7 48.133 0.49216

2.6.30-rc8+CFQ:
45.7 45.7 46.5 45.967 0.37712

2.6.29+NOOP:
47.1 48.9 48.5 48.167 0.77172

2.6.30-rc8+NOOP:
46.2 47.1 47.6 46.967 0.57927

So here we see that even with NOOP, 2.6.30-rc8 is still slower while it's
at the margin of statistical meaningfulness (I can gather more data if
people are interested).

Honza
--
Jan Kara <jack@suse.cz>
SUSE Labs, CR


\
 
 \ /
  Last update: 2009-06-10 11:15    [W:0.184 / U:0.536 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site