Re: [PATCH 2.5.43-mm2] New shared page table patch

["Martin J. Bligh" <mbligh@aracnet.com>]

>> > I think it will for most of the situations we run aground with now
>> > (normally 5000 oracle tasks sharing a 2Gb shared segment, or some
>> > such monster).
>> 
>> 10 GB pagetable overhead, for 2 GB of data.  No customer I
>> know would accept that much OS overhead.
>> 
>> To reduce the overhead we could either reclaim the page
>> tables and reconstruct them when needed (lots of work) or
>> we could share the page tables (less runtime overhead).
> 
> Or you use 4MB pages.  That tends to work much better and has less 
> complexity.  Shared page tables don't work well on x86 when you have 
> a database trying to access an SGA larger than the virtual address 
> space, as each process tends to map its own window into the buffer 
> pool.  Highmem with 32 bit va just plain sucks.  The right answer is 
> to change the architecture of the application to not run with 5000 
> unique processes.

Bear in mind that large pages are neither swap backed or file backed
(vetoed by Linus), for starters. There are other large app problem scenarios 
apart from Oracle ;-)

M.