[prev in list] [next in list] [prev in thread] [next in thread] 

List:       lse-tech
Subject:    Re: [Lse-tech] free_shortage & inactive_shortage
From:       Jesse Barnes <jbarnes () sgi ! com>
Date:       2001-07-14 2:10:06
[Download RAW message or body]

On Fri, 13 Jul 2001, Kanoj Sarcar wrote:

> At one point, my idea was to queue the zones that
> needed to be replenished to kswapd, instead of kswapd
> having to scan all zones to see which ones need
> replenishing. This was happening somewhere around
> 2.3.43 or so ... the idea was a poor man's version 
> of a per node kswapd. Some platforms could choose
> to have a per node kswapd, others could do with a
> global kswapd, others could have a perzone kswapd
> in the extreme case.
>
> Around 2.3.43 or so, there was also logic that 
> if zones needed to be replenished, the page 
> stealer would only reclaim/steal pages for the
> specific zones under the watermark (contrary
> to what it does currently, but either logic has
> its bad poitns). Unfortunately,
> people removed most of this logic as part of
> performance tuning.

That would work fine if the zones were treated differently, i.e. if a zone
could span multiple nodes, like the single zone on sn1 does in a logical
sense.
 
> A good direction to go in might be to see what it
> takes to have a pernode kswapd ... if there are 
> cpus on the node, the node kswapd can run on that
> node, else it might need to run on a cpu on another
> node.
> 
> Lastly, I believe the SGI machines have only a 
> single zone per node, so if allocations are 
> spread out properly, this should not be an issue.

Shouldn't allocations occur first on the node where the process is
running, as much as possible (but not up the point of waking up kswapd)?

> Kanoj
> 


_______________________________________________
Lse-tech mailing list
Lse-tech@lists.sourceforge.net
http://lists.sourceforge.net/lists/listinfo/lse-tech

[prev in list] [next in list] [prev in thread] [next in thread]