Re: Our recent Apache woes (fwd)

[Randy Terbush <ra...@zyzzyva.com>]

> Dean Gaudet wrote:
> > 
> > I am 100% convinced that serialized accepts are the way to go.  I'm still
> > convincing myself that weakly serialized (i.e. a maximum of N tasks inside
> > the loop) is a good thing as well (consider SMP boxes). 
> > 
> > The problem described below regarding the load is probably still explained
> > by the lack of serialization.  The difference is probably that hpux uses a
> > different formula for calculating its load avg than sunos 4.1.x.  I'm not
> > sure how to relate the rest of the problem though. 
> > 
> > We could start defining one of the serialization directives in 1.2.1
> > according to the reports we've received for:  osf, hpux, sunos.  Use
> > whatever the users have reported works for them.  (Or whoever uses those
> > architectures could tell me which of fcntl or flock works for you.) 
> 
> This is weird... I'm finding out that with 1.2.1 :) and 1.3 that
> FreeBSD and A/UX _also_ work better using serialized accepts...
> I'm sure no one else runs A/UX, but I know we have a few FreeBSDers
> here: Have you seen the same?

I've been running with flock'ed serialized accepts on FreeBSD for a 
couple of months now. The server "runs better". Without 
serialization, processes seemed to get in a state where they would 
never be selected to handle a request. Just sitting in the queue 
chewing up memory. I still seem to have problems of this type 
though.


> -- 
> ====================================================================
>       Jim Jagielski            |       jaguNET Access Services
>      jim@jaguNET.com           |       http://www.jaguNET.com/
>             "Look at me! I'm wearing a cardboard belt!"

Re: Our recent Apache woes (fwd)

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

In 1.2.1, do you think we should default freebsd and aux to
USE_FLOCK_SERIALIZED_ACCEPT?  sunos should work with that as well.  I
don't know what to use for hpux, ultrix, osf1, paragon, sequent, ...

Dean

On Fri, 27 Jun 1997, Randy Terbush wrote:

> > Dean Gaudet wrote:
> > > 
> > > I am 100% convinced that serialized accepts are the way to go.  I'm still
> > > convincing myself that weakly serialized (i.e. a maximum of N tasks inside
> > > the loop) is a good thing as well (consider SMP boxes). 
> > > 
> > > The problem described below regarding the load is probably still explained
> > > by the lack of serialization.  The difference is probably that hpux uses a
> > > different formula for calculating its load avg than sunos 4.1.x.  I'm not
> > > sure how to relate the rest of the problem though. 
> > > 
> > > We could start defining one of the serialization directives in 1.2.1
> > > according to the reports we've received for:  osf, hpux, sunos.  Use
> > > whatever the users have reported works for them.  (Or whoever uses those
> > > architectures could tell me which of fcntl or flock works for you.) 
> > 
> > This is weird... I'm finding out that with 1.2.1 :) and 1.3 that
> > FreeBSD and A/UX _also_ work better using serialized accepts...
> > I'm sure no one else runs A/UX, but I know we have a few FreeBSDers
> > here: Have you seen the same?
> 
> I've been running with flock'ed serialized accepts on FreeBSD for a 
> couple of months now. The server "runs better". Without 
> serialization, processes seemed to get in a state where they would 
> never be selected to handle a request. Just sitting in the queue 
> chewing up memory. I still seem to have problems of this type 
> though.
> 
> 
> > -- 
> > ====================================================================
> >       Jim Jagielski            |       jaguNET Access Services
> >      jim@jaguNET.com           |       http://www.jaguNET.com/
> >             "Look at me! I'm wearing a cardboard belt!"
> 
> 
> 
>