[f-nsp] Load balancing across sties on one real server

[Derek Maurer]

What I've described is a much more simpliflied version of what I'm doing
in reality, but this layout follow the same concept of what I was trying
to do (healthcheck and actual details of the configuration not listed
here).  On the physical server, each website is serving the same content
(there's an underlying application for each website with limited number
of threads, which is the reason load balancing 3 of the same websites
from one server make any kind of sense).

Physical Server	
IP 192.168.1.5, port 10000 (website1), port 10010 (website2), port 10020
(website3).

Foundry setup
Real Server r1: port 10000; port 10010; port 10020
Virtual Server: predictor round-robin, bind http r1 10000 r1 10010 r1
10020

Setup this way, only the real server port 10000 will see any traffic,
unless port 10000 is unavailable due to the healthcheck failing.  I
haven't tried yet, but I'm assuming if I put 2 more IP addresses on the
physical server and assign website2 an website3 to them, configure 2 new
real servers one for each website (2 & 3) and bind those, load balancing
is going to work properly. Ie:

Physical Server (still one server, but multiple IPs)	
IP 192.168.1.5 port 10000 (website1), IP 192.168.1.6 port 10010
(website2), IP 192.168.1.7 port 10020 (website3).

Foundry setup
Real Server r1: port 10000
Real Server r2: port 10010
Real Server r3: port 10020
Virtual Server: predictor round-robin, bind http r1 10000 r2 10010 r3
10020

I guess I'm just trying to confirm that when the documentation said the
predictor directs traffic across 'real servers', they actually ment
configured real servers and that load balancing using round-robin or
least-connections will not work unless there's multiple IP addresses
being bound (not just multiple ports on the same IP).

DRM


-----Original Message-----
From: Oliver Adam [mailto:oadam at madao.de] 
Sent: Thursday, October 30, 2008 6:40 AM
To: Derek Maurer; foundry-nsp at puck.nether.net
Subject: Re: [f-nsp] Load balancing across sties on one real server

I do not get the problem 100%. How many real servers do you have and how
are they linked to the virtual servers? There should not be any need to
configure unique IPs at the real server for each virtual server. What is
the predictor you are using? Have you tried to open multiple connection
simultaneously? Round-robin is normally the best predictor to see a load
balancing during testing in case you do have a single client only.
Ensure that the client is really opening multiple connections - some
browser tend to reuse sessions and this is of course not going to result
in any load balancing.

Do you have stickyness configured? If so: you will have to use different
client IPs to see some kind of load balancing.

- Oliver

At 05:43 30.10.2008, Derek Maurer wrote:
>I've noticed if I have multiple site running on the same real server I 
>can't balance the load across them, , regardless of the predictor.  The

>first real server port bound in the virtual server always seems to gets

>the connections.
>
>Reading the documentation closely, I noticed the predictor balances 
>load across 'real servers'.  Do I really need to configure unique IPs 
>for each site on the same physical server (and matching 'real servers')

>in order to load balance effectly across sites on one server?
>
>DRM
>_______________________________________________
>foundry-nsp mailing list
>foundry-nsp at puck.nether.net
>http://puck.nether.net/mailman/listinfo/foundry-nsp