On 5/24/06, Scott Dickson <Scott.Dickson at sun.com>
wrote:> I was talking with some customers today and came up with several
> questions that I don''t know the answers to. I''ll post
these as separate
> threads so as not to muddy things up too much.
>
> It seems like ZFS has a lot more knowledge of what is going on all the
> way down to the disk level than other FS. That gives a great
> opportunity to have some sort of QoS or bandwidth management. Is there
> a plan for this underway? The idea we came up with was that particular
> fs within a pool would have essentially shares (like in FSS) of the
> bandwidth to the disks in the pool promised to them. This customer has
> a lot of big batch jobs running on their system. They would like to
> promise that the largest one, which writes very infrequently but needs a
> *lot* of bandwidth when it does, gets a big share of the total when it
> needs it.
>
> Does this make sense?
This would make a swap pool[1] per zone to be much more palatable too.
The thing that has always worried be about memory resource controls
and allowing anyone to do anything substantial[2] with swap is that
its effects will go well beyond the zone being controlled. This could
be an effective method to limit the IO rate associated with swap.
Mike
[1] http://www.opensolaris.org/os/project/rm/pools/msets/
[2] Satisifying reservations is inconsequential, making a lot of use
swap kills performance of everyone on the machine
--
Mike Gerdts
http://mgerdts.blogspot.com/