[Devel] Re: [ckrm-tech] [PATCH 4/7] UBC: syscalls (user interface)

[Andrew Morton]

On Fri, 18 Aug 2006 07:45:48 -0700
Dave Hansen <haveblue at us.ibm.com> wrote:

> On Fri, 2006-08-18 at 12:08 +0400, Andrey Savochkin wrote:
> > 
> > A) Have separate memory management for each container,
> >    with separate buddy allocator, lru lists, page replacement mechanism.
> >    That implies a considerable overhead, and the main challenge there
> >    is sharing of pages between these separate memory managers.
> 
> Hold on here for just a sec...
> 
> It is quite possible to do memory management aimed at one container
> while that container's memory still participates in the main VM.  
> 
> There is overhead here, as the LRU scanning mechanisms get less
> efficient, but I'd rather pay a penalty at LRU scanning time than divide
> up the VM, or coarsely start failing allocations.
> 

I have this mad idea that you can divide a 128GB machine up into 256 fake
NUMA nodes, then you use each "node" as a 512MB unit of memory allocation. 
So that 4.5GB job would be placed within an exclusive cpuset which has nine
"mems" (what are these called?) and voila: the job has a hard 4.5GB limit,
no kernel changes needed.

Unfortunately this is not testable because numa=fake=256 doesn't come even
vaguely close to working.  Am trying to get that fixed.