httpd-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Marc Slemko <>
Subject Re: FYI - coredumps on FreeBSD
Date Mon, 07 Apr 1997 05:26:20 GMT
On Sun, 6 Apr 1997, Dean Gaudet wrote:

> On Sun, 6 Apr 1997, Marc Slemko wrote:
> >   I am wondering if reversing the order of searching the
> > queue (to make the queue of processes blocked in accept() be handled LIFO
> > instead of FIFO) could help for using Apache.
> Yes!  do do!  I've hinted this to the linux kernel developers too.  LIFO
> increases L2 cache-coolness.  I'm certain this will work based on the 
> benefits of a proper MaxSpareServer setting.  i.e. MaxSpare too high
> means lots of cpu churning for no apparently good reason, I suspect
> L2 cache boundary conditions.

It is a one or two line change for FreeBSD, so if it actually does work
once I try it I will probably make a patch available; it isn't something
that can be the default behavior, but I'm thinking it could be appropriate
for a setsockopt().  

As David Greenman correctly pointed out and I already knew, it shouldn't
help in caching the code since most caches are of physical memory and the
code is in shared pages, but the unshared parts (incl. data) could be

On top of that, you have the really lame fact that if you have room in
memory for n servers but have n+1 running, each n requests will result (on
most or at least many OSes; don't know that strict ordering is
guarenteed...) in every server being swapped in and out again.  That would
certainly be helped.

View raw message