Mike,

Thank you again for your response!

I didn't see any problem on iscsi layer but there is what I saw.

When I run with default queue size settings:
node.session.cmds_max = 128
node.session.queue_depth = 32
And trying to push to sg device more than or near 32 outstanding
commands I noticed the big delays for some IOs.
And when I set queue_depth to 128 I was able to push 90 outstanding
ios without any delays.
But when IO count that I pushed was again closed to 128 I started to
see reasonable delays.

Do you have any idea why this happens. Is it can be somehow connected
to open iscsi driver. I believe the poblem is in midl layer

By the way according to /proc/scsi/sg/debug the IO is in mid layer and
lower ... :

here is example:

  act: id=0 blen=4096 t_o/elap=30000/3750ms sgat=1 op=0x28
  act: id=0 blen=4096 t_o/elap=30000/3750ms sgat=1 op=0x28

I very appriciate your help.

Thanks.Evgeny



On Mar 28, 10:39 pm, Mike Christie <[email protected]> wrote:
> On 03/27/2011 03:08 AM, Evgeny Schmeilin wrote:
>
>
>
>
>
>
>
>
>
> > Hello,
>
> > Recently we have started to test our iscsi targets using Ubuntu hosts
> > with open-iscsi initiator. Our tests use sg to make IOs. We have seen
> > that on high load there are delays for some IOs. Sometime it takes 40
> > sec.
> > My investigation shows that the IOs are delayed somewere in the Ubuntu
> > stack. I see the IOs transmitted on the network long after it is
> > issued to sg and the response from the target is very fast.
>
> > Also I have seen some IOs that was issued to the same LUN  after the
> > "Long IO" but it was send before "Long IO" on the network and took
> > reasonable time.
>
> > So my questions are:
>
> > Is there  something in open-iscsi that can delay IOs ?
> > Is there something in open iscsi that can reorder  read/write
> > commands ?
>
> The iscsi driver sends the IO in the order it got sent to it. We only
> support the data pdus inorder option.
>
> The scsi layer could reorder things if a command times out and the scsi
> eh is run. You might be hitting this. It could also explain why you are
> seeing a delay in IO being sent.
>
> The network layer could reorder things, so it depends on where you are
> peaking at the ordering. If on the other side you are looking at the
> target layer then it should come in the proper order. If on the other
> side you are looking from the network layer then it depends on if you
> are looking at things above or below the tcp layer I guess.
>
> > Is there any way to see per IO trace/log record  in open-iscsi ?
>
> See modinfo iscsi_tcp,  modinfo libiscsi_tcp, and modinfo libiscsi. Do
> echo 1 > /sys/module/libiscsi/parameters/debug_*
> for each of the modules debug params.

-- 
You received this message because you are subscribed to the Google Groups 
"open-iscsi" group.
To post to this group, send email to [email protected].
To unsubscribe from this group, send email to 
[email protected].
For more options, visit this group at 
http://groups.google.com/group/open-iscsi?hl=en.

Reply via email to