On Sat, Oct 3, 2009 at 3:30 PM, Richard L. Hamilton <rlha...@smart.net> wrote:
> Do sparse root zones share pages of inherited files in the page cache
> across zones?

AFAIK, yes. And measurements support that.

> If so, doesn't that mean that, even though whole root on cloned zfs
> would result in the same _disk_ savings as a sparse zone, it wouldn't
> get similar savings in the page cache?  Has anyone actually measured
> the impact of this, i.e. of having multiple rather than single copies
> of libc and various other OS libs and programs resident at once?

I spent some time a while back measuring the memory footprint of
a sparse root zone - it's down at about 10Meg. Most pages are shared,
I think that svc.configd and svc.startd are two of the bigger hitters.

Any non-sparse zone is going to be way more than that. Although if
you have zfs clones then the zfs arc cache is shared.

Needless to say, even if that puts the per-zone footprint up to 100Meg
it's still an order of magnitude less than the full virtualization solutions
that it's competing with. And still small compared to running  a "modern"
application (java, for example) in the zone.

> If that works the way I think it might, I imagine that systems with
> lots of sparse zones might be likely to take a noticable hit changing
> over to the whole root plus zfs clone approach.
> I ask since with SXCE going away, and sparse root zones AFAIK not
> (yet?) supported in OpenSolaris, that might catch some folks by surprise.

For me, it's not just the memory savings. It's the fact that I can modify
the shared filesystems in the global zone and have the changes
immediately visible in the zones that can be convenient.

-Peter Tribble
http://www.petertribble.co.uk/ - http://ptribble.blogspot.com/
zones-discuss mailing list

Reply via email to