Subject: Re: UVM/UBC balancing
To: Lars Heidieker <lars@heidieker.de>
From: Chuck Silvers <chuq@chuq.com>
List: tech-kern
Date: 04/28/2001 12:14:26
hi lars,

the whole point of reactivating pages that are "the wrong type" is that
we *want* to do something that isn't LRU, so I don't think there's anything
wrong with this.  ideally we could arrange for the pagedaemon to not even
see pages that would just be reactivated anyway, but we can think about that
more once we are better about selecting which pages to reuse.

some of the other developers have been talking about another idea for this,
but I don't understand it yet.  I'll try to post some more info about this
new idea later today after I find out more about it.

-Chuck


On Fri, Apr 27, 2001 at 09:42:36AM +0200, Lars Heidieker wrote:
> Hi Chuck,
> 
> It just came up in my mind that we might break the LRU approximation of
> the pageoutdaemon havily.
> Reactivating pages because "they are not the right kind" like in your
> aproach with giving each kind of memory (vtext/data/vnode) a low water
> mark or what I have thought of keeping the LRU approximation (with a
> unified inactive queue) until the vnode mem is more than 2/3 of the
> entire memory and then scanning mainly for fscache only pages and
> reactivating others will result in flushing the pages thru the paging
> system by IO and therefor braking LRU approximation.
> If IO wasn't "that fast" we probably wouldn't need to think about it...
> But I think it can't be right that IO breaks the LRU approximation.
> 
> Lars