[nylug-talk] LOWMEM vs. HIGHMEM performance advantage?

[Adrian Sai-Wah TAM]

On Thu, Apr 17, 2008 at 6:26 PM, Ajai Khattri <ajai at bitblit.net> wrote:
> On Thu, 17 Apr 2008, jh wrote:
>
>
> > The reason why I ask: using this app with smaller jobs uses less memory,
>  > and hence, I was wondering if the himem/lowmem split could explain
>  > anything. Otherwise, it is likely a peculiarity of the app itself, and
>  > the question is better directed to the developer.
>
>  Ah, I see what you mean. You might be right. Anything close to 1Gb or more
>  involves an extra layer of page mapping from HIGHMEM into LOWMEM zones
>  before the kernel can access it so yeah that might account for slower
>  processing when running close to 1Gb.
>
>  (Actually, the normal mapping is more like 0 to 896Mb since the kernel
>  grabs 128Mb for itself so you dont get the full 1Gb to play with in the
>  norml zone).
>
>  Note: Im talking about the traditional 1Gb/3Gb split here - kernel patches
>  exist to do 2Gb/2Gb split if you want.

The only way to confirm: make the kernel do 2GB/2GB split and see if
the case improves. (but you probably don't want to do it if rerender02
is a production machine).