LVS
lvs-users
Google
 
Web LinuxVirtualServer.org

Re: LinuxExpo Paper

To: Rob Nourse <rnourse@xxxxxxxxxx>
Subject: Re: LinuxExpo Paper
Cc: linux-virtualserver@xxxxxxxxxxxx
From: Wensong Zhang <wensong@xxxxxxxxxxxx>
Date: Thu, 08 Apr 1999 19:05:46 +0800
Hello,


Rob Nourse wrote:

> I have noticed an error in the description of the IBM Network Dispatcher
> product in section 2 of the LinuxExpo Paper. The IBM Network dispatcher
> does NOT require any modifications to the actual back end server kernel
> code to function properly. In fact, any mix and match collection of back
> end servers will work.

I didn't describe the IBM Network Dispatcher in the section 2 , but the
TCP-router,
the predecessor of IBM Network Dispatcher.

>
> The trick to dealing with finicky clients is to alias the virtual service
> address to the loopback interface (on the actual webserver node) and then
> bind the webserver directly to the aliased address on lo0. In this way, the
> dispatcher node will leave the destination IP address intact, and add the
> MAC address of the network iterface as normal and when the destination
> webserver picks up the packet (because it contains the webserver's MAC
> address) it will forward it to the lo0 interface where the webserver lives.
>

Yeah, I like this dispatching technique. I did a few modification of the
LinuxDirector
code to implement this feature, I tested it and it worked. However, I don't
have
time to pack it and release it now, and partly because I want to add more
things
in the next release of virtual server patch. The comparision of NetDispatcher
to
virtual server via tunneling is that NetDispatcher doesn't have the overhead of

tunneling, but requires that dispatchers and servers are in the same network.
In the
virtual server via tunneling, the load balancer tunnel requests to the servers
(this
overhead is not high. :-)), and servers can be geographically distributed
through
WAN.

> When the reply is returned to the client, it will appear to emenate from
> the correct IP address. This allows for NAT to occur at a DMZ firewall and
> does not require the dispatcher to be involved in the return portion of the
> traffic. The downside of all this is:
>
> 1) Dispatchers and clients must live on the same physical segment

No, not clients, but real servers. Dispatchers and real servers are required to

have interfaces linked on the same physical network.

>
> 2) A node can either be a client or a server for a particular address, but
> not both. If you want to loadbalance traffic between two (loadbalanced)
> webservers and two (loadbalanced) application servers, you must use two
> different service addresses.
>
> All that said, I have not used the new V2.0 of the dispatcher product
> bundled into Websphere so things may(and probably have) changed.
>
> In either case, I'll let you know how our tril of LinuxDirector works out!
> rn
>

Thank you!

Wensong






<Prev in Thread] Current Thread [Next in Thread>