Re: Verification failures with mixed read/write case

[Jens Axboe <jens.axboe@oracle.com>]

On Thu, Feb 18 2010, Ricardo M. Correia wrote:
> (Please CC me in the reply, I'm not subscribed)
> 
> Hi,
> 
> I've been observing verification failures with the following
> configuration file:
> 
> -----------
> [global]
> ioengine=libaio
> iodepth=30
> rw=randrw
> bs=512-128k
> direct=1
> rwmixcycle=40
> verify=sha256
> verify_async=8
> 
> [/dev/sda2]
> --------------------
> 
> This happens on both RHEL5.4 and Ubuntu 9.10.
> 
> I was observing that with fio-1.36, but I just tried cloning fio from
> the git repo and I'm experiencing the same problem.
> 
> Any ideas of what might be wrong?
> 
> I found this email of a few months ago:
> http://www.spinics.net/lists/fio/msg00205.html
> 
> And I just verified that this patch is in my just-cloned git tree.

Mixed read-write workloads generally don't work well with verify. The
reason being that if you do verify && read, then fio will assume that
you want to verify previously written data. That should not happen for a
generated read as part of a mixed workload, I'll check what happens
here.

> Also, an unrelated issue is that fio complains about 'rwmixcycle' being
> deprecated but the man page doesn't specify what parameter I should be
> using instead.

The time based approach didn't work well for buffered writes, so it was
switched to an issue based count instead only. So you use
rwmixread/rwmixwrite to set the percentages.

-- 
Jens Axboe