lkml.org 
[lkml]   [2000]   [Jan]   [6]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: [RFC] [RFT] [PATCH] memory zone balancing
Date
On Thu, 6 Jan 2000, David S. Miller wrote:
> Date: Thu, 6 Jan 2000 17:05:41 +0100 (CET)
> From: Ingo Molnar <mingo@chiara.csoma.elte.hu>
>
> i think this is pretty much 'type-dependent'. In earlier versions
> of the zone allocator i added a zone->memory_balanced() function
> (but removed it later because it first needed the things your patch
> adds). Then every zone can decide for itself wether it's
> balanced. Eg. the DMA zone is rather critical and we want to keep
> it free aggressively (part of that is already achieved by placing
> it at the end of the zone chain), the highmem zone might not need
> any balancing at all, the normal zone wants some high/low watermark
> stuff.
>
> Let's be careful not to design any balancing heuristics which will
> fall apart on architectures where only one zone ever exists (because
> GFP_DMA is completely meaningless).

We'll have to keep the current constraint (freepages.{min,low,high})
stuff as our overall target anyway. We simply need to add some limits
that the special zones have to target for.

And it won't just be tested from kswapd(); swap_tick() and maybe
even __get_pages() will have to check for the constraints too and
wake up kswapd() or take action themselves...

Possible heuristics (using d for dma, n for normal and high for
high free mem; limit = freepages.*) could use the following targets:

d + n > limit
d > limit/4
n > limit/4
h < limit * 2 (because we don't care about free high pages
and we do care about free normal and dma pages)

Another possibility is to include the highmem (d + n + h > limit)
and keep the single low-mem limits.

IMHO there should be a number of constraints for a heuristic
like this:
- simple
- not force too much memory free
- the algorithm should have freedom in what to free, in order
to avoid excessive scanning (alt: separate queues per zone)
- the limits for dma and normal memory should be high enough
to be able to fulfill kernel allocations
- should have no impact on non-zoned machines

The (extremely simple) constraints above should probably be
enough to actually get the job done. At least, I wouldn't
know why we'd need more...

cheers,

Rik
--
The Internet is not a network of computers. It is a network
of people. That is its real strength.


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2005-03-22 13:55    [W:0.110 / U:0.036 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site