Disk IO scheduling in XEN

Disk IO scheduling in XEN

[Paresh Nakhe]

[-- Attachment #1.1: Type: text/plain, Size: 742 bytes --]

Hi,

We are working on a project to modify the disk scheduling mechanism in XEN a
bit in an attempt to improve it. In the paper "Xen and the Art of
Virtualization"  following is mentioned.

*"Xen services batches of requests from competing domains in a
simple round-robin fashion; these are then passed to a standard elevator
scheduler before reaching the disk hardware*"

We were going through linux-jeremy source code in an attempt to map the
above in code. We could not however do so. On the contrary we came to the
conclusion that there is no such mechanism. Domain 0 services requests as
soon as it receives a hypercall from a guest domain. Are we right?

Which one would be better first come first serve or round robin and why?

Thanks

[-- Attachment #1.2: Type: text/html, Size: 816 bytes --]

[-- Attachment #2: Type: text/plain, Size: 138 bytes --]

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

Re: Disk IO scheduling in XEN

[John Weekes]

[-- Attachment #1.1: Type: text/plain, Size: 1148 bytes --]

On 3/24/2011 1:19 PM, Paresh Nakhe wrote:
> Hi,
>
> We are working on a project to modify the disk scheduling mechanism in 
> XEN a bit in an attempt to improve it. In the paper "Xen and the Art 
> of Virtualization"  following is mentioned.
>
> /"Xen services batches of requests from competing domains in a
> simple round-robin fashion; these are then passed to a standard elevator
> scheduler before reaching the disk hardware/"
>
> We were going through linux-jeremy source code in an attempt to map 
> the above in code. We could not however do so. On the contrary we came 
> to the conclusion that there is no such mechanism. Domain 0 services 
> requests as soon as it receives a hypercall from a guest domain. Are 
> we right?

Ian Campbell forward-ported the 2.6.18-xen.hg patches for this but I 
don't think that they've been merged into the main git 2.6.32 stable 
tree yet. Check the list archives for the subject line "unfair servicing 
of DomU vbd requests" for more information on this.

> Which one would be better first come first serve or round robin and why?

Of the two, RR, because domains can become starved otherwise.

-John

[-- Attachment #1.2: Type: text/html, Size: 1834 bytes --]

[-- Attachment #2: Type: text/plain, Size: 138 bytes --]

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

RE: Disk IO scheduling in XEN

[James Harper]

> Hi,
> 
> We are working on a project to modify the disk scheduling mechanism in
XEN a
> bit in an attempt to improve it. In the paper "Xen and the Art of
> Virtualization"  following is mentioned.
> 
> "Xen services batches of requests from competing domains in a
> simple round-robin fashion; these are then passed to a standard
elevator
> scheduler before reaching the disk hardware"
> 
> We were going through linux-jeremy source code in an attempt to map
the above
> in code. We could not however do so. On the contrary we came to the
conclusion
> that there is no such mechanism. Domain 0 services requests as soon as
it
> receives a hypercall from a guest domain. Are we right?
> 
> Which one would be better first come first serve or round robin and
why?
> 

There was a discussion about this recently. It turns out that one DomU
can starve the others and there are real life examples of this
happening. Patches have been proposed and posted but I don't know into
what trees.

James

RE: Disk IO scheduling in XEN

[Ian Campbell]

On Thu, 2011-03-24 at 22:29 +0000, James Harper wrote:
> There was a discussion about this recently. It turns out that one DomU
> can starve the others and there are real life examples of this
> happening. Patches have been proposed and posted but I don't know into
> what trees.

They are in Linus' tree so they will be in 2.6.39-rc1.

It isn't yet in Jeremy's tree, I have a branch at:

git://xenbits.xen.org/people/ianc/linux-2.6.git irq-fairness-2.6.32

John Weekes has tested it and reported success
(<4D7935C0.9010509@nuclearfallout.net>) so, Jeremy please pull:

The following changes since commit 6d6ba2f4ea5f5a11f31ed707445ec4a57d225eb6:
  Jeremy Fitzhardinge (1):
        Merge branch 'stable-2.6.32.x-dom0' of git://xenbits.xen.org/people/sstabellini/linux-pvhvm into xen/next-2.6.32

are available in the git repository at:

  git://xenbits.xen.org/people/ianc/linux-2.6.git irq-fairness-2.6.32

Ian Campbell (1):
      xen: events: Make last processed event channel a per-cpu variable.

Keir Fraser (3):
      xen: events: Clean up round-robin evtchn scan.
      xen: events: Make round-robin scan fairer by snapshotting each l2 word
      xen: events: Remove redundant clear of l2i at end of round-robin loop

Scott Rixner (1):
      xen: events: Process event channels notifications in round-robin order.

 drivers/xen/events.c |   80 +++++++++++++++++++++++++++++++++++++++++++++-----
 1 files changed, 72 insertions(+), 8 deletions(-)

Re: Disk IO scheduling in XEN

[Jeremy Fitzhardinge]

On 03/24/2011 08:19 PM, Paresh Nakhe wrote:
> Hi,
>
> We are working on a project to modify the disk scheduling mechanism in
> XEN a bit in an attempt to improve it. In the paper "Xen and the Art
> of Virtualization"  following is mentioned.
>
> /"Xen services batches of requests from competing domains in a
> simple round-robin fashion; these are then passed to a standard elevator
> scheduler before reaching the disk hardware/"
>
> We were going through linux-jeremy source code in an attempt to map
> the above in code. We could not however do so. On the contrary we came
> to the conclusion that there is no such mechanism. Domain 0 services
> requests as soon as it receives a hypercall from a guest domain. Are
> we right?

There's no explicit Xen-specific disk scheduling.  Linux has a great
deal of intrinsic mechanisms for setting up disk scheduling, including
using cgroups to schedule groups of processes, so you can use those to
get any particular scheduling policy you desire.

> Which one would be better first come first serve or round robin and why?

I think there are many more options for disk scheduling than those.

For example: "is it better for guests to schedule their own requests, or
should they just pass requests straight through?".

    J