Re: balance round-robin/sticky on Host: header

Hi Benoit,

On Thu, Sep 04, 2008 at 06:07:18PM +0200, Benoit Plessis wrote:
> Willy Tarreau a écrit :
> >
> >>I'm trying to come up with a configuration for HAProxy to balance by the
> >>'Host' header in a generic fashion (i.e. one that does not require me to
> >>hard code host header values to backends). We're using wildcard DNS and
> >>basically have clients registered to their own host/domain: e.g. -
> >>client1.mydomain.com, client2.mydomain.com, etc.
> >>
> >>All requests for a particular host would stick to the same server. I
> >>think it's pretty similar to the URI hashing scheme. If I understand
> >>that scheme requests are balanced round-robin, but stick to backends
> >>based on a hash of the uri.
> >>
> >>Any ideas? BTW, I'm using HAProxy-1.3.14.1 on Gentoo Linux.
> >>
> >
> >The header hash is in the todo list, with many other things, but is just
> >not implemented right now. There are multiple things to consider for this,
> >among which check/ignore case, specify which occurrence of the header to
> >match, etc... so it will not be implemented tomorrow morning...
> >
> >Right now, I do not see any easy method to ensure that all clients for a
> >given host will get to the same server.
> >
> >May I ask for what specific purpose you need the feature (I'm not trying
> >to escape it's implementation, just trying to find a temporary workaround)
> >?
> >Is it more for performance, for persistence, etc...
> >
>
> Hi, back on this subject.
> I think Jeff and I share the same type of plateforme (load-balancing of
> multiple (thousand for myself) small to medium web sites)
>
> And this feature, of which we already discussed ("Haproxy & un-usual
> session tracking", late 2007) is needed for performance.
> Our Application Server can't cache every site on one back-end, and so
> 'normal' round-robin lead to an high performance impact:
>
> previously we had average number like that:
> dns query: 2ms
> connection time: 26ms
> reception of first octet: 139ms
> total load time: 140ms
>
> now we have number like that:
> dns query: 2ms
> connection time: 26ms
> reception of first octet: 623ms
> total load time: 676ms
>
> and up to that:
> dns query: 2ms
> connection time: 26ms
> reception of first octet: 1771,33ms
> total load time: 1828,33ms

that's a valuable feedback. In this case, couldn't you use :

• URI hashing (the same URL will always be processed by the same server) ?
• source IP hashing (most clients will remain on to the same server) ?

That's just pure guess, of course.
Willy Received on 2008/09/11 22:57