Re: dramatic I/O slowdown after upgrading 2.6.38->3.0+

[Mike Christie <michaelc@cs.wisc.edu>]

On 04/11/2012 04:40 AM, Michael Tokarev wrote:
> On 10.04.2012 19:13, Jan Kara wrote:
>> > On Tue 10-04-12 10:00:38, Michael Tokarev wrote:
> []
>>> >>   2.6.38:
>>> >>   # dd if=/dev/sdb of=/dev/null bs=1M iflag=direct count=100
>>> >>   100+0 records in
>>> >>   100+0 records out
>>> >>   104857600 bytes (105 MB) copied, 1.73126 s, 60.6 MB/s
>>> >>
>>> >>   3.0:
>>> >>   # dd if=/dev/sdb of=/dev/null bs=1M iflag=direct count=100
>>> >>   100+0 records in
>>> >>   100+0 records out
>>> >>   104857600 bytes (105 MB) copied, 29.4508 s, 3.6 MB/s
>>> >>
>>> >> That's about 20 times difference on direct read from the
>>> >> same - idle - device!!
>> >   Huh, that's a huge difference for such a trivial load. So we can rule out
>> > filesystems, writeback, mm. I also wouldn't think it's IO scheduler but
>> > you can always check by comparing dd numbers after
>> >   echo none >/sys/block/sdb/queue/scheduler

Did you try newer 3.X kernels or just 3.0?

We were hitting a similar problem with iscsi. Same workload and it
started with 2.6.38. I think it turned out to be this issue:

// thread with issue like what we hit:
http://thread.gmane.org/gmane.linux.kernel/1244680

// Patch that I think fixed issue:
commit 3deaa7190a8da38453c4fabd9dec7f66d17fff67
Author: Shaohua Li <shaohua.li@intel.com>
Date:   Fri Feb 3 15:37:17 2012 -0800

    readahead: fix pipeline break caused by block plug