|
On Tue, 11 Apr 2000, Horms wrote:
> On Wed, Apr 12, 2000 at 10:45:26AM +1000, Benjamin Lee wrote:
> >
> > I was thinking that a user space program could sit on each of the real
> > servers and then tell the LVS how much it *wants* another connection. This
> > *weight* value could be inserted into the list that holds the available
> > real servers, and could be incorporated into one of the existing
> > scheduling algorithms maybe. Obviously the overhead is a TCP connection
> > from the LVS box to each real server. But I think this is acceptable.
>
> I don't know if any implementations have been made available
> but certainly such a scheme is workable. ldirectord which ships
> as both part of heartbeat and part of IPVS monitors back
> end servers and takes them in and out of the pool when failures
> occur. In some ways what you are suggesting would be an extension of
> this.
>
Excellent. I shall take a look at ldirectord.
This is good news as I was hoping to investigate some sort of intelligent,
adaptive weighting scheme into LVS's, say, weighted RR scheduling
algorithm. And because the *weigh* calculation would be done by each real
server there wouldn't be any extra load on the LVS-box.
Sincerely,
Ben
|
