Question on Lazy VM Restore

Question on Lazy VM Restore

[Prateek Sharma]

Hello all,
 	I was wondering if there is an existing mechanism in Xen for 
restoring a saved/checkpointed VM lazily. By that, I mean that the pages 
of the VM should be mapped only on access.
 	Currently, restoring a VM from a snapshot seems to take a long 
time [~20 seconds for 2 GB].
 	Assuming that there is no lazy-restore functionality, what would 
be a good way to implement this? Using parts of Xen-paging? Snowflock? 
Something else?

Thanks!
--Prateek

Re: Question on Lazy VM Restore

[Andrew Cooper]

On 13/12/2013 16:06, Prateek Sharma wrote:
> Hello all,
>     I was wondering if there is an existing mechanism in Xen for
> restoring a saved/checkpointed VM lazily. By that, I mean that the
> pages of the VM should be mapped only on access.
>     Currently, restoring a VM from a snapshot seems to take a long
> time [~20 seconds for 2 GB].
>     Assuming that there is no lazy-restore functionality, what would
> be a good way to implement this? Using parts of Xen-paging? Snowflock?
> Something else?
>
> Thanks!
> --Prateek

There is no current mechanism which I am aware of.  I recall a few
academic papers looking in to the possibility, but no code as a result.

The restoring from a shapshot has a number of large inefficiencies in
the datapath in the privcmd driver.  Fixing these would make quite an
improvement on those 20 seconds, which I suspect might be a rather
easier improvement than implementing lazy restore.

Having said that, if you are looking for a project to do, I suspect the
xen-paging infrastructure would be able to be used for lazy restore,
although Andres (author of xen-paging, CC'd) will likely be able to
speak rather more about the feasibility of it.

~Andrew

Re: Question on Lazy VM Restore

[Andres Lagar Cavilla]

[-- Attachment #1.1: Type: text/plain, Size: 1608 bytes --]

On Sat, Dec 14, 2013 at 11:28 AM, Andrew Cooper
<andrew.cooper3@citrix.com>wrote:

> On 13/12/2013 16:06, Prateek Sharma wrote:
> > Hello all,
> >     I was wondering if there is an existing mechanism in Xen for
> > restoring a saved/checkpointed VM lazily. By that, I mean that the
> > pages of the VM should be mapped only on access.
> >     Currently, restoring a VM from a snapshot seems to take a long
> > time [~20 seconds for 2 GB].
> >     Assuming that there is no lazy-restore functionality, what would
> > be a good way to implement this? Using parts of Xen-paging? Snowflock?
> > Something else?
> >
> > Thanks!
> > --Prateek
>
> There is no current mechanism which I am aware of.  I recall a few
> academic papers looking in to the possibility, but no code as a result.
>
> The restoring from a shapshot has a number of large inefficiencies in
> the datapath in the privcmd driver.  Fixing these would make quite an
> improvement on those 20 seconds, which I suspect might be a rather
> easier improvement than implementing lazy restore.
>
> Having said that, if you are looking for a project to do, I suspect the
> xen-paging infrastructure would be able to be used for lazy restore,
> although Andres (author of xen-paging, CC'd) will likely be able to
> speak rather more about the feasibility of it.
>

Thanks Andrew. Credit where is due, Olaf Hering from Suse had most to do
with xen-paging.

Prateek: yes you could use xen-paging to build a lazy restore. Snowflock
has mutated/evolved into Gridcentric's commercial offering, so happy to
discuss that off list.

Best,
Andres

>
> ~Andrew
>

[-- Attachment #1.2: Type: text/html, Size: 2292 bytes --]

[-- Attachment #2: Type: text/plain, Size: 126 bytes --]

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