On Wed, Feb 27, 2008 at 2:59 PM, Martin Schwidefsky
<[EMAIL PROTECTED]> wrote:

>  The EX_TABLE entry will take care of the fault. As far as I can tell
>  Heikos patch is fine.

Yes, and I see now that this is only in the __init part so it should
not bother me.

>  > PS I did read that Brian Wade's experiments were done with cmma=on and
>  > use the CP settings to enable / disable the feature. When my
>  > assumptions about the implementation are correct, then that was indeed
>  > the right way to do it. But it does not reveal what will happen to
>  > customers who end up with a mix of penguins.
>
>  This would be an interesting experiment. The difference between cmma=on
>  and cmma=off for a page cache page that has been paged-out by z/VM is
>  not big. If a guest accesses such a page with cmma=off the z/VM system
>  will initiate an i/o to get the page and deliver a pfault interrupt to
>  the guest so that the guest can schedule another process. If a guest
>  uses cmma=on it will get a discard fault on access of a paged-out page.
>  In this case it is Linux that initiates the i/o. The process will have
>  to wait for page and another process is scheduled.

Our first assumption was that if some of the guests play the game nice
with cmma=on telling CP what pages they can afford to miss, they would
be very easy target to take pages from. Much easier than guests that
are not co-operative and CP must assume their entire working set is
stable. It did not turn out to work that way.

The situation where CP assumes the guest is co-operative (because it
issued an ESSA) but Linux decides not to flag pages would make the
entire virtual machine as stable (I assume). I understand the Linux
view on this. But apart from that, I believe it would influence which
pages CP takes away when it can pick. Now I often see CP take pages
that are paged out already, which is not good. If the cmma stuff does
not change the z/VM demand scan then I think it's a missed
opportunity.

Rob
--
Rob van der Heij
Velocity Software, Inc
http://velocitysoftware.com/

----------------------------------------------------------------------
For LINUX-390 subscribe / signoff / archive access instructions,
send email to [EMAIL PROTECTED] with the message: INFO LINUX-390 or visit
http://www.marist.edu/htbin/wlvindex?LINUX-390

Reply via email to