On Mon, 2005-12-12 at 18:01 +0100, [EMAIL PROTECTED] wrote: 
> To handle pages shared between two or more CKRM classes, I tried a new 
> implementation that adds a new class in the memory controller to manage 
> these
> pages. This class is named the "Shared" class.

First of all, please don't top-post.  It makes it very difficult to
reply to your messages.  Also, you may want to reexamine what mail
client you are using to send messages.  They show up line-wrapped here.

I worry that your "Shared" class is a band-aid solution which attempts
to fix a fundamental problems the memory controller's design.  As you
noticed, even tuning the numbers in lots of different ways can't make it
work right.

I would argue that the allocator is the dead wrong place to do
accounting, and implement the controller.  The allocator needs to be
fast as possible, and have as little overhead as possible.  Same goes
for 'struct page', which needs to be as lean as possible.  These are two
things that are opposed to the current design.

The memory controller also changes fundamental things about how page
reclaim is done.  This is a horribly complex area that is hard to get
right because it has to work for _everybody_.  I've seen patches go by
on this list for very fundamental things like OOMs triggering a the
wrong time.  That worries me that it will take years to stabilize the
memory controller's impact, because its testing is so much smaller than
that of the larger VM.

Instead of patching up the current controller, I'd encourage all of you
to take a hard look at what you're really trying to do, and make sure
that the current code can get you there.  I'd suggest that it can't.

-- Dave



-------------------------------------------------------
This SF.net email is sponsored by: Splunk Inc. Do you grep through log files
for problems?  Stop!  Download the new AJAX search engine that makes
searching your log files as easy as surfing the  web.  DOWNLOAD SPLUNK!
http://ads.osdn.com/?ad_id=7637&alloc_id=16865&op=click
_______________________________________________
ckrm-tech mailing list
https://lists.sourceforge.net/lists/listinfo/ckrm-tech

Reply via email to