Re: ext4 write performance regression in 3.6-rc1 on RAID0/5

[Yuanhan Liu <yuanhan.liu@linux.intel.com>]

On Wed, Aug 22, 2012 at 12:07:26PM +0800, Shaohua Li wrote:
> On 8/22/12 11:57 AM, Yuanhan Liu wrote:
> > On Fri, Aug 17, 2012 at 10:25:26PM +0800, Fengguang Wu wrote:
> >> [CC md list]
> >>
> >> On Fri, Aug 17, 2012 at 09:40:39AM -0400, Theodore Ts'o wrote:
> >>> On Fri, Aug 17, 2012 at 02:09:15PM +0800, Fengguang Wu wrote:
> >>>> Ted,
> >>>>
> >>>> I find ext4 write performance dropped by 3.3% on average in the
> >>>> 3.6-rc1 merge window. xfs and btrfs are fine.
> >>>>
> >>>> Two machines are tested. The performance regression happens in the
> >>>> lkp-nex04 machine, which is equipped with 12 SSD drives. lkp-st02 does
> >>>> not see regression, which is equipped with HDD drives. I'll continue
> >>>> to repeat the tests and report variations.
> >>>
> >>> Hmm... I've checked out the commits in "git log v3.5..v3.6-rc1 --
> >>> fs/ext4 fs/jbd2" and I don't see anything that I would expect would
> >>> cause that. The are the lock elimination changes for Direct I/O
> >>> overwrites, but that shouldn't matter for your tests which are
> >>> measuring buffered writes, correct?
> >>>
> >>> Is there any chance you could do me a favor and do a git bisect
> >>> restricted to commits involving fs/ext4 and fs/jbd2?
> >>
> >> I noticed that the regressions all happen in the RAID0/RAID5 cases.
> >> So it may be some interactions between the RAID/ext4 code?
> >>
> >> I'll try to get some ext2/3 numbers, which should have less
> >changes
> on the fs side.
> >>
> >> wfg@bee /export/writeback% ./compare -g ext4
> lkp-nex04/*/*-{3.5.0,3.6.0-rc1+}
> >> 3.5.0 3.6.0-rc1+
> >> ------------------------ ------------------------
> >> 720.62 -1.5% 710.16 lkp-nex04/JBOD-12HDD-thresh=1000M/ext4-100dd-1-3.5.0
> >> 706.04 -0.0% 705.86 lkp-nex04/JBOD-12HDD-thresh=1000M/ext4-10dd-1-3.5.0
> >> 702.86 -0.2% 701.74 lkp-nex04/JBOD-12HDD-thresh=1000M/ext4-1dd-1-3.5.0
> >> 702.41 -0.0% 702.06 lkp-nex04/JBOD-12HDD-thresh=1000M/ext4-1dd-2-3.5.0
> >> 779.52 +6.5% 830.11 lkp-nex04/JBOD-12HDD-thresh=100M/ext4-100dd-1-3.5.0
> >> 646.70 +4.9% 678.59 lkp-nex04/JBOD-12HDD-thresh=100M/ext4-10dd-1-3.5.0
> >> 704.49 +2.6% 723.00 lkp-nex04/JBOD-12HDD-thresh=100M/ext4-1dd-1-3.5.0
> >> 704.21 +1.2% 712.47 lkp-nex04/JBOD-12HDD-thresh=100M/ext4-1dd-2-3.5.0
> >> 705.26 -1.2% 696.61 lkp-nex04/JBOD-12HDD-thresh=8G/ext4-100dd-1-3.5.0
> >> 703.37 +0.1% 703.76 lkp-nex04/JBOD-12HDD-thresh=8G/ext4-10dd-1-3.5.0
> >> 701.66 -0.1% 700.83 lkp-nex04/JBOD-12HDD-thresh=8G/ext4-1dd-1-3.5.0
> >> 701.17 +0.0% 701.36 lkp-nex04/JBOD-12HDD-thresh=8G/ext4-1dd-2-3.5.0
> >> 675.08 -10.5% 604.29
> lkp-nex04/RAID0-12HDD-thresh=1000M/ext4-100dd-1-3.5.0
> >> 676.52 -2.7% 658.38 lkp-nex04/RAID0-12HDD-thresh=1000M/ext4-10dd-1-3.5.0
> >> 512.70 +4.0% 533.22 lkp-nex04/RAID0-12HDD-thresh=1000M/ext4-1dd-1-3.5.0
> >> 524.61 -0.3% 522.90 lkp-nex04/RAID0-12HDD-thresh=1000M/ext4-1dd-2-3.5.0
> >> 709.76 -15.7% 598.44 lkp-nex04/RAID0-12HDD-thresh=100M/ext4-100dd-1-3.5.0
> >> 681.39 -2.1% 667.25 lkp-nex04/RAID0-12HDD-thresh=100M/ext4-10dd-1-3.5.0
> >> 524.16 +0.8% 528.25 lkp-nex04/RAID0-12HDD-thresh=100M/ext4-1dd-2-3.5.0
> >> 699.77 -19.2% 565.54 lkp-nex04/RAID0-12HDD-thresh=8G/ext4-100dd-1-3.5.0
> >> 675.79 -1.9% 663.17 lkp-nex04/RAID0-12HDD-thresh=8G/ext4-10dd-1-3.5.0
> >> 484.84 -7.4% 448.83 lkp-nex04/RAID0-12HDD-thresh=8G/ext4-1dd-1-3.5.0
> >> 470.40 -3.2% 455.31 lkp-nex04/RAID0-12HDD-thresh=8G/ext4-1dd-2-3.5.0
> >> 167.97 -38.7% 103.03
> lkp-nex04/RAID5-12HDD-thresh=1000M/ext4-100dd-1-3.5.0
> >> 243.67 -9.1% 221.41 lkp-nex04/RAID5-12HDD-thresh=1000M/ext4-10dd-1-3.5.0
> >> 248.98 +12.2% 279.33 lkp-nex04/RAID5-12HDD-thresh=1000M/ext4-1dd-1-3.5.0
> >> 208.45 +14.1% 237.86 lkp-nex04/RAID5-12HDD-thresh=1000M/ext4-1dd-2-3.5.0
> >> 71.18 -34.2% 46.82 lkp-nex04/RAID5-12HDD-thresh=100M/ext4-100dd-1-3.5.0
> >> 145.84 -7.3% 135.25 lkp-nex04/RAID5-12HDD-thresh=100M/ext4-10dd-1-3.5.0
> >> 255.22 +6.7% 272.35 lkp-nex04/RAID5-12HDD-thresh=100M/ext4-1dd-1-3.5.0
> >> 243.09 +20.7% 293.30 lkp-nex04/RAID5-12HDD-thresh=100M/ext4-1dd-2-3.5.0
> >> 209.24 -23.6% 159.96 lkp-nex04/RAID5-12HDD-thresh=8G/ext4-100dd-1-3.5.0
> >> 243.73 -10.9% 217.28 lkp-nex04/RAID5-12HDD-thresh=8G/ext4-10dd-1-3.5.0
> >
> > Hi,
> >
> > About this issue, I did some investigation. And found we are blocked at
> > get_active_stripes() in most times. It's reasonable, since max_nr_stripes
> > is set to 256 now. It's a kind of small value, thus I tried with
> > different value. Please see the following patch for detailed numbers.
> >
> > The test machine is same as above.
> >
> > From 85c27fca12b770da5bc8ec9f26a22cb414e84c68 Mon Sep 17 00:00:00 2001
> > From: Yuanhan Liu <yuanhan.liu@linux.intel.com>
> > Date: Wed, 22 Aug 2012 10:51:48 +0800
> > Subject: [RFC PATCH] md/raid5: increase NR_STRIPES to 1024
> >
> > Stripe head is a must held resource before doing any IO. And it's
> > limited to 256 by default. With 10dd case, we found that it is
> > blocked at get_active_stripes() in most times(please see the ps
> > output attached).
> >
> > Thus I did some tries with different value set to NR_STRIPS, and
> > here are some numbers(EXT4 only) I got with different NR_STRIPS set:
> >
> > write bandwidth:
> > ================
> > 3.5.0-rc1-256+: (Here 256 means with max strip head set to 256)
> > write bandwidth: 280
> > 3.5.0-rc1-1024+:
> > write bandwidth: 421 (+50.4%)
> > 3.5.0-rc1-4096+:
> > write bandwidth: 506 (+80.7%)
> > 3.5.0-rc1-32768+:
> > write bandwidth: 615 (+119.6%)
> >
> > (Here 'sh' means with Shaohua's "multiple threads to handle
> >strips"
> patch [0])
> > 3.5.0-rc3-strip-sh+-256:
> > write bandwidth: 465
> >
> > 3.5.0-rc3-strip-sh+-1024:
> > write bandwidth: 599
> >
> > 3.5.0-rc3-strip-sh+-32768:
> > write bandwidth: 615
> >
> > The kernel maybe a bit older but I found that the data are still kind of
> > valid. Though, I haven't tried Shaohua's latest patch.
> >
> > As you can see from those data above: the write bandwidth is increased
> > (a lot) as we increase NR_STRIPES. Thus the bigger NR_STRIPES set, the
> > better write bandwidth we get. But we can't set NR_STRIPES with a too
> > large number, especially by default, or it need lots of memory. Due to
> > the number I got with Shaohua's patch applied, I guess 1024 would be
> > nice value; it's not too big but we gain above 110% performance.
> >
> > Comments? BTW, I have a more flexible(more stupid, in the meantime) way:
> > change the max_nr_stripes dynamically based on need?
> >
> > Here I also attached more data: the script I used to get those number,
> > ps output, and iostat -kx 3 output.
> >
> > The script does it's job in a straight way: start NR dd in background,
> > trace the writeback/global_dirty_state event in background to count the
> > write bandwidth, sample the ps out regularly.
> >
> > ---
> > [0]: patch: http://lwn.net/Articles/500200/
> >
> > Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
> > ---
> > drivers/md/raid5.c | 2 +-
> > 1 files changed, 1 insertions(+), 1 deletions(-)
> >
> > diff --git a/drivers/md/raid5.c b/drivers/md/raid5.c
> > index adda94d..82dca53 100644
> > --- a/drivers/md/raid5.c
> > +++ b/drivers/md/raid5.c
> > @@ -62,7 +62,7 @@
> > * Stripe cache
> > */
> >
> > -#define NR_STRIPES 256
> > +#define NR_STRIPES 1024
> > #define STRIPE_SIZE PAGE_SIZE
> > #define STRIPE_SHIFT (PAGE_SHIFT - 9)
> > #define STRIPE_SECTORS (STRIPE_SIZE>>9)
> 
> does revert commit 8811b5968f6216e fix the problem?

Hi Shaohua,

Quote those numbers again:
 write bandwidth:
 ================
 3.5.0-rc1-256+:
 	write bandwidth: 280
 3.5.0-rc1-1024+:
 	write bandwidth: 421 (+50.4%)
 3.5.0-rc1-4096+:
 	write bandwidth: 506 (+80.7%)
 3.5.0-rc1-32768+:
 	write bandwidth: 615 (+119.6%)

Where the above kernel does not include commit 8811b5968f6216e; it's bit
old kernel.

The following kernel does, which I applied your patch
series(http://thread.gmane.org/gmane.linux.raid/38711)

 3.5.0-rc3-strip-sh+-256:
 	write bandwidth: 465

 3.5.0-rc3-strip-sh+-1024:
 	write bandwidth: 599

 3.5.0-rc3-strip-sh+-32768:
 	write bandwidth: 615


And yes, the kernel is old. But from Fengguang's data, I don't see that
new kernel matters too much.

Thanks,
Yuanhan Liu