Perrin Harkins wrote:
>
> > Odd thing #1: As it gets into evening time, load on the machine drops off
> > and there are fewer httpd children running, but I am not seeing free
> > memory return to that 1.3GB level. At most it comes back up to 400MB or
> > so. I don't think the httpd children are hanging on to memory, because
> > they cycle through pretty quickly - MaxRequestsPerChild is set to 512 and
> > none of the processes are ever more than a couple minutes old when I look
> > in. Is there any reason to think the parent httpd process would hang on
> > to anything?
>
> I have noticed that over time the new processes will spawn with less memory
> shared. I'm not sure exactly why this is, but it does seem to happen. A
> nightly complete restart of the server will reset things, but you may not
> have that option.
>
I have noticed this also. I have a theory about its cause, but I haven't
had the free time to hack with it yet. My theory is that when the VM
subsystem swaps pages of the mod_perl parent process, those pages become
forever unshared (even after they return to memory from swap). So I was
going to test the mlockall() Linux system call to see if it made a
difference. (Search the archives for "mlockall").
However, I am using the Linux platform - the VM behavior may be totally
different on Solaris.
If anyone else has some time to play with mlockall() on Linux, I'd like to
know the results. It'll be a few more weeks till I get a chance.
-Adi
