Re: acls and httpclose

On Sat, Apr 24, 2010 at 08:47:24AM +0200, Angelo Höngens wrote:
> On 23-4-2010 21:11, Willy Tarreau wrote:
> > On Fri, Apr 23, 2010 at 01:14:29PM +0200, Angelo Höngens wrote:
> >>> My observations are that http-server-close is about twice as fast as
> >>> httpclose, so you could save about half of the CPU usage here.
> >>>
> >>> Willy
> >>>
> >>
> >> Trying to wrap my head around this.. I read the configuration guide, and
> >> let me see if I understand correctly: when I use the http-server-close
> >> option, connections between haproxy and the backends or not persistent,
> >> but for each request, a new connection will be established.
> >>
> >> But this adds the overhead of connection establishment, and should
> >> increase cpu usage, not descrease it, right? What am I missing?
> >
> > You're not missing anything, that's precisely why http-server-close
> > (which closes one side) is faster than httpclose (which closes both
> > sides).
> >
> > Willy
>
> Aha, now it makes sense. I thought the httpclose option was closing only
> client connections.
>
> Perhaps for you it is clear, but for me the documentation, although
> elaborate, is a bit hazy on the subject.

OK, probably I need to rework the doc on the subject.

> What would be more clear for me, would be two basic options:
> http-server-close and http-client-close. and perhaps
> http-server-forceclose and http-client-forceclose as well, and the
> general http-close and http-forceclose option that closes both
> connections. But it's your party ;)

I see your point. The fact is that we try our best to remain compatible with older versions, and some keywords such as httpclose were inherited from 7-8 years back. When I see some old setups running 1.1.32 with hundreds of instances, I'm very glad that 1.4.4 still supports them without changing one line of config.

Regards,
Willy Received on 2010/04/24 09:12