Kai
On 11/29/05, Chandra Seetharaman <[EMAIL PROTECTED]> wrote:
On Tue, 2005-11-29 at 15:29 -0500, kai unc wrote:
> Hi Chandra,
Hi Kai,
Sorry that I missed your earlier email. Just realized that you sent an
email after seeing this one.
>
> Looks like __alocate_pages is for all physical page allocations ... so
> in any case, a process which occupies more RSS
> than its failover will have __alocate_pages return NULL. In my test,
> VM will just kill the process.
>
This is the way it works as of now.
> Is that possible, instead of return NULL, that we force paging RSS of
> this process to disk space and use the reduced
> space for new pages ? I understand this will have IO impact, but isn't
> that smoother than killing the process ? The process
> will just run slower but not dead..
Valerie is looking into solving this problem.
> Just for my education ...
>
> In case of malloc, it seems Virtual memory is allocated
> immediately, ( malloc return), but RSS part does
> not increase. The RSS will increase after same memory is first
> time written to. When are the new physical
> memory pages created in this case? malloc or the first time
> write to?
I guess it will be in the first write.
>
> Another case, if a page was swapped from physical memory to
> swapspace, later, it is about to be
> swapped back for access, however, the task is over its
> failover. Is this a new physical memory
> pages allocation?
Yes, it will be a new physical page allocation.
>
> And third case, if something like some_lib.so is paged out
> from memory, and because there is a copy of
> some_lib.so in filesystem, there isn't a copy in physical
> memory and swap space. Later, when the
> some_lib.so is pagein to physical memory, is this a new
> physical memory pages allocation ?
Yes, this will be similar to swap.
>
> Actually, I try to understand when might the page allocation
> be fail, with mainline kernel, even the physical
> memory is not enough, as long as there is SWAP space, there
> will not be access issues to memory, right ?
> will ckrm behave same ?
Nope. As of now ckrm does not behave that way. We are working on getting
it closer.
alloc_page does not know which (of the above) type of alloc it is. One
thing we can do is to try to free up some pages (from the class) before
failing the alloc.
>
>
> Hope this clarifies.
>
> BTW, we are now focusing only on f-series. Testing
> functionality in f-
> series may be a good idea. (Realized that memory
> controller is not
> uploaded. Just did it now).
>
>
> Thanks I will get new version and test it.
>
>
> >
> > Thanks for your clarification.
> >
> > Best Regards,
> > Kai
> >
> >
> >
> >
> --
>
> ----------------------------------------------------------------------
> Chandra Seetharaman | Be careful
> what you choose....
> - [EMAIL PROTECTED]
> | .......you may get it.
> ----------------------------------------------------------------------
>
>
>
>
--
----------------------------------------------------------------------
Chandra Seetharaman | Be careful what you choose....
- [EMAIL PROTECTED] | .......you may get it.
----------------------------------------------------------------------
