swapping of mod_perl parent process / mlockall()

[Adi Fairbank <ad...@certsite.com>]

Is it correct that when the Apache/mod_perl parent process swaps to disk, a
large part of it (swapped pages) becomes unshared?  Even after the kernel
restores the pages from swap, do they remain unshared?  So once the parent
process becomes unshared, new apache children that are spawned only share
the parent's pages that have never been swapped?

This is what I've grok'ed from my experiments with top/GTop.

If this is the case, it would be helpful to prevent the parent process from
*ever* swapping to disk.

The Linux kernel has a system call mlockall() which disables all memory
paging for the current process.  This sounds like exactly the recipe for
preventing parent process swapping.  It won't affect spawned child
processes, so shouldn't pose a threat of consuming all physical memory and
crashing the system.

I want to play around with this, but I'm not sure where in the mod_perl
source to call mlockall().  It would need to be before the parent spawns any
children..  Has anyone tried using Linux's mlockall()/mlock() with mod_perl?

-Adi

Re: swapping of mod_perl parent process / mlockall()

[Doug MacEachern <do...@covalent.net>]

On Mon, 16 Jul 2001, Adi Fairbank wrote:
 
> Actually, I don't want child processes to inherit the page locks across a
> fork.  I just wanted to experiment with performance issues when only the
> parent process is locked in memory.  (I have a theory that when the parent
> process swaps to disk, the swapped pages become unshared for the rest of the
> server's life)
> 
> I was hoping you could give me a hint as to where in the source code I could
> call mlockall(), e.g. file mod_perl.c, line NNN..

you should just be able to create a module with a .xs something like:
#include "EXTERN.h"
#include "perl.h"
#include "XSUB.h"
#include "sys/mman.h"

MODULE = Mlock		PACKAGE = Mlock		

int
mlockall()

  CODE:
  RETVAL = (mlockall(MCL_CURRENT) == 0);

  OUTPUT:
  RETVAL

and call Mlock::mlockall(); in a startup script.

Re: swapping of mod_perl parent process / mlockall()

[Adi Fairbank <ad...@certsite.com>]

Doug MacEachern wrote:
> 
> On Tue, 26 Jun 2001, Adi Fairbank wrote:
> 
> >
> > If this is the case, it would be helpful to prevent the parent process from
> > *ever* swapping to disk.
> >
> > The Linux kernel has a system call mlockall() which disables all memory
> > paging for the current process.  This sounds like exactly the recipe for
> > preventing parent process swapping.  It won't affect spawned child
> > processes, so shouldn't pose a threat of consuming all physical memory and
> > crashing the system.
> >
> > I want to play around with this, but I'm not sure where in the mod_perl
> > source to call mlockall().  It would need to be before the parent spawns any
> > children..  Has anyone tried using Linux's mlockall()/mlock() with mod_perl?
> 
> according the manpage:
> "Child processes do not inherit page locks across a fork."
> 
> so you would need to call it in a child init handler, but:
> "Only root processes are allowed to lock pages."

Actually, I don't want child processes to inherit the page locks across a
fork.  I just wanted to experiment with performance issues when only the
parent process is locked in memory.  (I have a theory that when the parent
process swaps to disk, the swapped pages become unshared for the rest of the
server's life)

I was hoping you could give me a hint as to where in the source code I could
call mlockall(), e.g. file mod_perl.c, line NNN..

Thanks,
-Adi

Re: swapping of mod_perl parent process / mlockall()

[Doug MacEachern <do...@covalent.net>]

On Tue, 26 Jun 2001, Adi Fairbank wrote:

> 
> Is it correct that when the Apache/mod_perl parent process swaps to disk, a
> large part of it (swapped pages) becomes unshared?  Even after the kernel
> restores the pages from swap, do they remain unshared?  So once the parent
> process becomes unshared, new apache children that are spawned only share
> the parent's pages that have never been swapped?
> 
> This is what I've grok'ed from my experiments with top/GTop.
> 
> If this is the case, it would be helpful to prevent the parent process from
> *ever* swapping to disk.
> 
> The Linux kernel has a system call mlockall() which disables all memory
> paging for the current process.  This sounds like exactly the recipe for
> preventing parent process swapping.  It won't affect spawned child
> processes, so shouldn't pose a threat of consuming all physical memory and
> crashing the system.
> 
> I want to play around with this, but I'm not sure where in the mod_perl
> source to call mlockall().  It would need to be before the parent spawns any
> children..  Has anyone tried using Linux's mlockall()/mlock() with mod_perl?

according the manpage:
"Child processes do not inherit page locks across a fork."

so you would need to call it in a child init handler, but:
"Only root processes are allowed to lock pages."