Q: /sys/block and I/O Schedulers

Q: /sys/block and I/O Schedulers

[Ulrich Windl]

Hi!

I have a question: Reading the docs on I/O Schedulers, I had the impression the docs wanted to tell me that only the low-level devices (i.e. disks) use I/O Schedulers, while higher-level devices (like multipaths, RAIDs, LVs, etc.) don't.

Using the SLES11 SP1 kernel (2.6.32.43-0.4-xen), I found out that LVs seem to use the I/O Scheduler, while RAID1 don't seem to do so. Also a disk in a Xen DomU also uses an I/O Scheduler (in addition to those being used in Dom0 for the same device already)

Is that correct, and if so, why is that? Is there a change to be seen in the future?

My test was as simple as this:
/sys/block # for d in *
> do
> echo $d: $(<$d/queue/scheduler)
> done
cciss!c0d0: noop anticipatory deadline [cfq]
dm-0: noop anticipatory deadline [cfq]
[...]
loop0: none
[...]
md4: none
sda: noop anticipatory deadline [cfq]
[...]

As I'm not subscribed to the list, I'd appreciate any CC: to my address. Thanks!

Regards,
Ulrich

Re: Q: /sys/block and I/O Schedulers

[Milan Broz]

On 08/22/2011 12:10 PM, Ulrich Windl wrote:
> I have a question: Reading the docs on I/O Schedulers, I had the
> impression the docs wanted to tell me that only the low-level devices
> (i.e. disks) use I/O Schedulers, while higher-level devices (like
> multipaths, RAIDs, LVs, etc.) don't.

As you already found, low-level device has always real io scheduler.

Device-mapper use stacked device logic - IOW mapped device
is stacked over some real device(s).

Remapping works on bio level, not on IO request level.
Separate bios are simple remapped to low-level devices where
io scheduler does its job.

There is one exception: dm-multipath which uses "request based" mapping,
IOW it means it uses own scheduler on device-mapper level
(But in old kernels it was on bio level as well,
for more info see https://lkml.org/lkml/2008/2/15/411)

> Using the SLES11 SP1 kernel (2.6.32.43-0.4-xen), I found out that LVs
> seem to use the I/O Scheduler

LVs (e.g. linear mappings) do not use own scheduler. But /queue directory
contains more attributes which need to be visible there.

> My test was as simple as this: /sys/block # for d in *
>> do echo $d: $(<$d/queue/scheduler) done

Btw see "lsblk -t" - here you can see stacked devices with scheduler
and topology info parsed from sysfs (in new util-linux).

Milan