|
On 1999-09-26T22:18:18,
Wensong Zhang <wensong@xxxxxxxxxxxx> said:
> I don't think there is a need to use the netmask to select the proxy server.
> For a dialing up session, the dynamically allocated IP address won't be
> changed in a single session. The current persistent port feature of LVS keep
> the connections from the same client assigned to the same server that is
> allocated for the first connection. It will make all proxy requests go to
> the same proxy server.
>
> The static partition method (iph->saddr & netmask) is really not necessary,
> which will make the load imbalance worse among the proxy servers.
Sorry, but I apparently made my point incorrectly.
Yes, the _user_ keeps the same IP during the same session. However, he is
going to a proxy cluster which does NOT implement session persistence, so he
gets directed to all 4 proxy servers during a single dialup session. The
proxy servers of course come to the LVS with 4 different src ips, and the
client gets redirected (quite correctly, from the point of the LVS) to
multiple different servers during the same dialup session.
So what I need the netmask feature for is to work around non-persistent proxy
clusters, as implemented by T-Online. Since T-Online is the single largest ISP
in Germany, our customer wants a solution which works around the T-Online
problem.
I have since been told that AOL and maybe Compuserve proxy servers may exhibit
the same behaviour to the outside.
Of course, I _know_ this is a tradeoff, since it will mean that even proxy
clusters which DO implement session persistence correctly will get bundled on
a single real server. Thats why I want to make the netmask for the template
configurable, if you do not wish to use it, it can just be set to
255.255.255.255, resulting in the old behaviour.
I am not concerned about sessions being persistent for multiple dialup
sessions. That is something which is already taken care of in the application.
However, having the same real client hit multiple real servers during a single
transaction is a really really bad hit on performance which I need to work
around.
I hope I managed to explain the problem, the solution and that I am well aware
that it is not perfect.
If I were to add such a patch, would it get into the mainstream LVS ?
Sincerely,
Lars Marowsky-Brée
--
Lars Marowsky-Brée
Network Management
teuto.net Netzdienste GmbH
|
