Re: [PATCH][RFC] simpler __alloc_pages{_limit}

[Roger Larsson <roger.larsson@skelleftea.mail.telia.com>]

On Saturday den 25 August 2001 13:55, Stephan von Krawczynski wrote:
> On Sat, 25 Aug 2001 02:48:28 +0200
>
> Roger Larsson <roger.larsson@norran.net> wrote:
> > Hi again,
> >
> > [two typos corrected from the version at linux-mm]
> > [...]
> > Doing this - the code started to collaps...
> > __alloc_pages_limit could suddenly handle all special cases!
> > (with small functional differences)
> >
> > Comments?
>
> Hi Roger,
>
> I tested your page against straight 2.4.9 (where it applied mostly, the
> rest I did manually) and experience the following: 
> 1) system gets slow, even in times where plenty of free memory is
> available. There must be some overhead inside.

It is not unlikely because it care too much about the higher order 
allocations. It needs a higher order page and really tries...

> 2) It does not really work around the basic problem of too
> many cached pages in case of heavy filesystem action, I do get the already
> known "kernel: __alloc_pages: 2-order allocation failed." by simply copying
> files a lot. 

Is this with raiserfs and/or nfs? And without highmem support?
Why is 2-order allocations needed???
Can anyone answer?
Higher order allocs during normal operation is not that nice...

> 3) Even in high load situations the CPU load seems to get
> worse, I made it up to 7 with normal file copying on a SMP 1GHz 1GB RAM
> machine.

Might also be related to the higher order. Freeing too much inactive pages
to satisfy the request...
SMP might be a factor since the patch will go harder on the locks...

>
> Hm, I guess that doesn't really work as you expected.

Well, I make a version that gives up on higher order allocations more 
quickly...

But the real problem might be - why are the higher order allocations
needed anyway?

/RogerL

-- 
Roger Larsson
Skellefteå
Sweden