Re: on large numbers of virtual hosts and memory use

[Dmitry Khrustalev <di...@xyzzy.machaon.ru>]

Walking pconf in standalone_main before starting daemons and doing
mprotect(.., PROT_READ) on each memory block should help.

>The workaround, since we know the child should never be writing to
>them so COW is simply an excuse for not doing shared memory, is
>to actually implement that via shared memory.  Then you get rid
>of the pages mapped in each child, and are much happier.  This
>doesn't look to be _that_ major an undertaking to me.  Comments?

This is not true for BUFFERED_LOGS case.

	-Dima

Re: on large numbers of virtual hosts and memory use

[Dmitry Khrustalev <di...@bog.msu.su>]

On Sat, 18 Oct 1997, Marc Slemko wrote:
> On Sat, 18 Oct 1997, Dmitry Khrustalev wrote:
> 
> > 
> > Walking pconf in standalone_main before starting daemons and doing
> > mprotect(.., PROT_READ) on each memory block should help.
> 
> I'm not convinced that will change anything because a process may change
> the protection back to read/write at any time, so the kernel (on systems
> that really like reserving swap space) will reserve it anyway.

Indeed. what's strange, solaris does not really like reserving swap space
- if you do not touch allocated memory, it would not reserve anything. 

mmap(MAP_PRIVATE|MAP_NORESERVE) work.

	-Dima

> 

Re: on large numbers of virtual hosts and memory use

[Marc Slemko <ma...@worldgate.com>]

On Sat, 18 Oct 1997, Dmitry Khrustalev wrote:

> 
> Walking pconf in standalone_main before starting daemons and doing
> mprotect(.., PROT_READ) on each memory block should help.

I'm not convinced that will change anything because a process may change
the protection back to read/write at any time, so the kernel (on systems
that really like reserving swap space) will reserve it anyway.

Re: on large numbers of virtual hosts and memory use

[Dean Gaudet <dg...@arctic.org>]

Yeah this is something that I meant to look at ... never had the time
though. Given that I mostly use IRIX and Linux which both have optimistic
COW it's not a huge issue for me.  But it's a pain in the butt on the few
solaris servers I deal with.

To do this I think we'd need three pools:

    pconf	like current pconf, but should only be stuff which is
		shared RO
    ptemp	like current ptemp
    punshare	probably created before init_modules() and contains stuff
		which is unshared -- but allocated by the parent

Dean

On Sat, 18 Oct 1997, Dmitry Khrustalev wrote:

> 
> Walking pconf in standalone_main before starting daemons and doing
> mprotect(.., PROT_READ) on each memory block should help.
> 
> >The workaround, since we know the child should never be writing to
> >them so COW is simply an excuse for not doing shared memory, is
> >to actually implement that via shared memory.  Then you get rid
> >of the pages mapped in each child, and are much happier.  This
> >doesn't look to be _that_ major an undertaking to me.  Comments?
> 
> This is not true for BUFFERED_LOGS case.
> 
> 	-Dima
>