PAE kernel and 4GB of memory

[Patrick O'Callaghan]

On Thu, 2008-07-24 at 14:40 -0700, Nifty Fedora Mitch wrote:
> On Thu, Jul 24, 2008 at 01:49:34PM +0100, Howard Wilkinson wrote:
> > 
> > I am looking for a definitive answer to the question of where the PAE  
> > kernels become useful. I have seen various articles that mention needing  
> > PAE kernels if you have more then 4GB of physical memory in a 32-bit  
> > processor environment. I have also seen statements that say you need  
> > them if you have 4GB or more of memory. Now which is right? Also, even  
> > if you need a PAE kernel because the last few bytes are not addressable  
> > when you have exactly 4GB is this useful or is the trade off of larger  
> > page tables and pages going to eat any benefit of being able to address  
> > these few bytes and if so when does the PAE kernel become useful?
> >
> 
> http://en.wikipedia.org/wiki/Physical_Address_Extension
> 
> Can you be more specific and define 'useful'.
> 
> In general on a 32 bit system you will have 32 bit pointers by default...
> signed arithmetic gives you an effective 2GB process size.   Compare and
> contrast lseek() and lseek64()... sizeof(off_t something).

32 bits gives you a 4GB address space. Pointers are not signed (and the
off_t type reflects this, which is why it's different from int).

> But if you have six 2GB processes running on a 6.x GB system is that useful?
> Are you playing with one Big process.
> 
> Do you have a test case  or pointer to a test case (best) so folks with
> large memory systems can sanity test this for you?

I suspect the question is related to accessing *physical* (not virtual)
memory.

poc