Re: [RFC] Process requests instead of bios to use a scheduler

[Sebastian Parschauer <sebastian.riemer@profitbricks.com>]

On 02.06.2014 12:20, NeilBrown wrote:
> On Mon, 02 Jun 2014 11:51:52 +0200 Sebastian Parschauer 
> <sebastian.riemer@profitbricks.com> wrote:
>> 
>> Nope, we have our RAID-1+0. So it is more or less a RAID-10 and
>> putting the scheduler to this RAID-0 layer makes sense for us.
> 
> I still cannot imagine how this would work.  RAID-0 has no
> decisions to make, so no where for a scheduler to fit.
> 
> Just to clarify: is this md/raid0 over md/raid1 or md/raid0 over 
> hardware/raid1?

We have both variants but tested the scheduler with servers without HW
RAID and only 4 HDDs. On the production servers there are 24 HDDs,
every 2 HDDs are in md/raid1 or HW RAID-1 and these 12 RAID-1 devices
form an md/raid0 device.

This is the only PV for LVM and LVs are the customer volumes exported
via SCST/SRP. These are used by virtual machines on other servers. So
the scheduler has to bring some fairness to the customer volumes so
that a streaming customer can't block a database customer completely
with his big sequential IOs as these would have priority. But our goal
with a scheduler is to reduce latency for everyone.

I hope this is clear, now. Just tell me if not. ;-)