|
Replies:
6
-
Last Post:
Mar 18, 2006 2:36 PM
by: comay
|
|
|
Brad Diggs
Bradley.Diggs@Sun.COM
|
|
|
|
Parallel Creation/Deletion at Serial Speed...
Posted:
Mar 17, 2006 4:00 PM
|
|
Hi all,
I recently had the pleasure of doing some zones creation/deletion
performance analysis on an x4200 server. One of the things that I
was trying to determine was if I attempt to create multiple zones
(say 10-20) in parallel, will the time to create the zones be
reduced any. The net result was that although all the zones were
flawlessly created, the net time to create them was about the same
as if I had created the zones serially.
I suppose this makes sense in order for package management to ensure
consistency for each zone. However, it sure would be nice if I
parallel zone creation performance could be improved.
You may be wondering why in the world anyone would want to quickly
create lots of zones in parallel. In my case it was so that I could
quickly setup a shared lab environment for about 20 students where
each student had their own zone to perform the lab. In my experience
it takes about 45-60 minutes to create a whole root (S10u1/OEM) zone.
So lab setup for 20 students takes roughly 20 hours.
One amazing side note worthy of mentioning though is that creating
the zones on the x4200 SAS disks consistently took 17-20 minutes.
That was a significant reduction from IDE, SCSI, and even a RAID5 T3
work group array.
With all that said, is there any work going into how to safely and
efficiently parallelize zone creation?
Thanks in advance,
Brad
_______________________________________________
zones-discuss mailing list
zones-discuss at opensolaris dot org
|
|
|
Posts:
370
From:
San Francisco, CA
Registered:
3/9/05
|
|
|
|
Re: Parallel Creation/Deletion at Serial Speed...
Posted:
Mar 17, 2006 4:06 PM
in response to: Brad Diggs
|
|
On Fri, Mar 17, 2006 at 06:00:47PM -0600, Brad Diggs wrote:
> Hi all,
>
> I recently had the pleasure of doing some zones creation/deletion
> performance analysis on an x4200 server. One of the things that I
> was trying to determine was if I attempt to create multiple zones
> (say 10-20) in parallel, will the time to create the zones be
> reduced any. The net result was that although all the zones were
> flawlessly created, the net time to create them was about the same
> as if I had created the zones serially.
>
> I suppose this makes sense in order for package management to ensure
> consistency for each zone. However, it sure would be nice if I
> parallel zone creation performance could be improved.
>
> You may be wondering why in the world anyone would want to quickly
> create lots of zones in parallel. In my case it was so that I could
> quickly setup a shared lab environment for about 20 students where
> each student had their own zone to perform the lab. In my experience
> it takes about 45-60 minutes to create a whole root (S10u1/OEM) zone.
> So lab setup for 20 students takes roughly 20 hours.
>
> One amazing side note worthy of mentioning though is that creating
> the zones on the x4200 SAS disks consistently took 17-20 minutes.
> That was a significant reduction from IDE, SCSI, and even a RAID5 T3
> work group array.
>
> With all that said, is there any work going into how to safely and
> efficiently parallelize zone creation?
The new "zoneadm clone" subcommand is probably what you're looking for.
Check out http://learningsolaris.com/archives/2005/11/30/zones-move-clone/.
I'm not sure what the S10 timeline for it is, though...
Cheers,
- jonathan
--
Jonathan Adams, Solaris Kernel Development
_______________________________________________
zones-discuss mailing list
zones-discuss at opensolaris dot org
|
|
|
|
|
Posts:
807
From:
US
Registered:
3/9/05
|
|
|
|
Re: Parallel Creation/Deletion at Serial Speed...
Posted:
Mar 17, 2006 5:51 PM
in response to: jwadams
|
|
On Fri 17 Mar 2006 at 04:06PM, Jonathan Adams wrote:
> On Fri, Mar 17, 2006 at 06:00:47PM -0600, Brad Diggs wrote:
> > Hi all,
> >
> > I recently had the pleasure of doing some zones creation/deletion
> > performance analysis on an x4200 server. One of the things that I
> > was trying to determine was if I attempt to create multiple zones
> > (say 10-20) in parallel, will the time to create the zones be
> > reduced any. The net result was that although all the zones were
> > flawlessly created, the net time to create them was about the same
> > as if I had created the zones serially.
> >
> > I suppose this makes sense in order for package management to ensure
> > consistency for each zone. However, it sure would be nice if I
> > parallel zone creation performance could be improved.
> >
> > You may be wondering why in the world anyone would want to quickly
> > create lots of zones in parallel. In my case it was so that I could
> > quickly setup a shared lab environment for about 20 students where
> > each student had their own zone to perform the lab. In my experience
> > it takes about 45-60 minutes to create a whole root (S10u1/OEM) zone.
> > So lab setup for 20 students takes roughly 20 hours.
> >
> > One amazing side note worthy of mentioning though is that creating
> > the zones on the x4200 SAS disks consistently took 17-20 minutes.
> > That was a significant reduction from IDE, SCSI, and even a RAID5 T3
> > work group array.
> >
> > With all that said, is there any work going into how to safely and
> > efficiently parallelize zone creation?
>
> The new "zoneadm clone" subcommand is probably what you're looking for.
> Check out http://learningsolaris.com/archives/2005/11/30/zones-move-clone/.
> I'm not sure what the S10 timeline for it is, though...
Yep; additionally, we have a good handle on why (sparse) zone creation
is slow: the contents file.
During the creation of a zone, we wind up writing and rewriting the
/var/sadm/install/contents file hundreds of times, which really hurts,
since it is huge.
Turns out this is also hurting regular install time, and so this is
an area of active work. I anticipate a dramatic zone installation
speedup (for sparse zones, anyway) when we get this resolved.
-dp
--
Daniel Price - Solaris Kernel Engineering - dp at eng dot sun dot com - blogs.sun.com/dp
_______________________________________________
zones-discuss mailing list
zones-discuss at opensolaris dot org
|
|
|
|
|
Posts:
1,264
From:
US
Registered:
8/5/05
|
|
|
|
|
|
Posts:
962
From:
US
Registered:
3/9/05
|
|
|
|
Re: Parallel Creation/Deletion at Serial Speed...
Posted:
Mar 18, 2006 2:36 PM
in response to: mgerdts
|
|
> If the zone attach code
> (http://www.opensolaris.org/jive/thread.jspa?threadID=6843&tstart=0)
> could make it into Update 2 along with ZFS, you could likely script
> the zfs snapshot, zfs clone, zoneadm attach sequence to do this almost
> as fast as my 0.922 seconds per zone. Of course, you are probably
> using someting faster than an Ultra 2. :)
Unfortunately, the cloning work is scheduled for one of the Solaris
Update releases coming after Update 2. And there are still some issues
that need to be worked out so that patching and upgrading will be
successful for systems where zone's root are on ZFS file systems.
As we know more with respect to schedules, we'll definitely roll that
out here.
dsc
_______________________________________________
zones-discuss mailing list
zones-discuss at opensolaris dot org
|
|
|
|
|
Posts:
244
From:
Registered:
5/25/05
|
|
|
|
Re: Parallel Creation/Deletion at Serial Speed...
Posted:
Mar 18, 2006 10:20 AM
in response to: jwadams
|
|
To add some numbers to this: I just did some experiments with cloning zones. Cloning a whole root zone is almost twice as fast as installing. Cloning a sparse root zone is more than thirty times faster than installing! See my blog for the details.
Menno
|
|
|
|
|
Posts:
962
From:
US
Registered:
3/9/05
|
|
|
|
Re: Re: Parallel Creation/Deletion at Serial Speed...
Posted:
Mar 18, 2006 2:32 PM
in response to: menno
|
|
> To add some numbers to this: I just did some experiments with cloning
> zones. Cloning a whole root zone is almost twice as fast as installing.
> Cloning a sparse root zone is more than thirty times faster than
> installing! See my blog for the details.
One of the nice attributes of using clone is the subsequent clones (if
they occur in roughly the same time period) are even faster since the
file data is in buffer cache.
For those looking for Menno's blog, btw, it can be found at
http://blogs.sun.com/menno
dsc
_______________________________________________
zones-discuss mailing list
zones-discuss at opensolaris dot org
|
|
|
|
|
