08.06.2010 16:29, Dave Chinner wrote:
On Tue, Jun 08, 2010 at 01:55:51PM +0400, Michael Tokarev wrote:
I've got a.. difficult issue here, and am asking if anyone else
has some expirence or information about it.
Production environment (database). Machine with an Adaptec
RAID SCSI controller, 6 drives in raid10 array, XFS filesystem
and Oracle database on top of it (with - hopefully - proper
Upgrading kernel from 2.6.27 to 2.6.32, and users starts screaming
about very bad performance. Iostat reports increased I/O latencies,
I/O time increases from ~5ms to ~30ms. Switching back to 2.6.27,
and everything is back to normal (or, rather, usual).
I tried testing I/O with a sample program which performs direct random
I/O on a given device, and all speeds are actually better in .32
compared with .27, except of random concurrent r+w test, where .27
gives a bit more chances to reads than .32. Looking at the synthetic
tests I'd expect .32 to be faster, but apparently it is not.
This is only one machine here which is still running 2.6.27, all the
rest are upgraded to 2.6.32, and I see good performance of .32 there.
But this is also the only machine with hardware raid controller, which
is onboard and hence not easy to get rid of, so I'm sorta forced to
use it (I prefer software raid solution because of numerous reasons).
One possible cause of this that comes to mind is block device write
barriers. But I can't find when they're actually implemented.
The most problematic issue here is that this is only one machine that
behaves like this, and it is a production server, so I've very little
chances to experiment with it.
So before the next try, I'd love to have some suggestions about what
to look for. In particular, I think it's worth the effort to look
at write barriers, but again, I don't know how to check if they're
actually being used.
Anyone have suggestions for me to collect and to look at?
Yes, I've seen this. We use xfs for quite long time. The on-board
controller does not have battery unit, so it should be no different
than a software raid array or single drive.
But I traced the issue to a particular workload -- see $subject.
Simple test doing random reads or writes of 4k blocks in a 1Gb
file located on an xfs filesystem, Mb/sec:
read write write
2.6.27 xfs 1.17 3.69 3.80
2.6.32 xfs 1.26 0.52 5.10
2.6.32 ext3 1.19 4.91 5.02
Note the 10 times difference between O_SYNC and O_DIRECT writes
in 2.6.32. This is, well, huge difference, and this is where
the original slowdown comes from, apparently. In 2.6.27 both
sync and direct writes are on-par with each other, in .32
direct write has improved, but sync write is just pathetic now.
And compared with previous o_sync, that's about 6 times the
difference which I reported previously.
We're running a legacy oracle application here, on Oracle8,
which does not support O_DIRECT and uses O_SYNC. So it gets
hit by this issue quite badly - no doubt users start screaming
after switching to .32.
I also tested ext3fs, for comparison. This one does not have
that problem and works just fine in both .32 and .27. I also
tried disabling barriers for xfs, which made no difference
So it's O_SYNC writes on XFS which are problematic. Together
with hw raid apparently, since no one noticed when I switched
other machines (with sw raid) from .27 to .32.
I'll _try_ to find when the problem first appeared, but it is
not that simple since I've only very small time window for