This is regarding the load balancing/sharing across the servers.

Assume I have the following in my setup:
- clients: C-1, C-2, ..., C-m
- 1 load sharing server(LSS)
- Servers: S-1, S-2, ...., S-n

Here are my requirements:
- The load sharing should be done across the servers(S-1,...S-n) only
based on the number of connections(and not based on the number,size of
the requests etc. ) towards them.
- The clients use only one IP address for all the requests(including
for fetching new connections).
- Only the initial connecting 'Bind' request should go via the LSS. The
LSS assigns one of the servers which has less number of connections
towards it(round-robin if all the servers have equal connections). For
this, the LSS maintains a table for keeping the number of connections
towards each of the servers.
- Once the server is assigned, the requests from the clients should be
sent directly to the designated Server.
- Complete routing steps(initial bind requests via LSS; further
application requests to server directly) should be transparent to the
Clients.
- If a connection to the server fails, the client retries to get a new
connection. Now, this request should be transparently sent to the LSS
and the same cycle repeats.

I would like to use such a load sharing mechanism, as I personally feel
it is simply a performance overhead to route the application requests
via the LSS, when the load sharing is done only based on the number of
connections.

Is it possible to implement such a load sharing mechanism using any of
the routing protocols(E.g.VRRP)?

qazmlp1209@rediffmail.com wrote:
> This is regarding the load balancing/sharing across the servers.
>
> Assume I have the following in my setup:
> - clients: C-1, C-2, ..., C-m
> - 1 load sharing server(LSS)
> - Servers: S-1, S-2, ...., S-n
>
> Here are my requirements:
> - The load sharing should be done across the servers(S-1,...S-n) only
> based on the number of connections(and not based on the number,size of
> the requests etc. ) towards them.
> - The clients use only one IP address for all the requests(including
> for fetching new connections).
> - Only the initial connecting 'Bind' request should go via the LSS. The
> LSS assigns one of the servers which has less number of connections
> towards it(round-robin if all the servers have equal connections). For
> this, the LSS maintains a table for keeping the number of connections
> towards each of the servers.
> - Once the server is assigned, the requests from the clients should be
> sent directly to the designated Server.
> - Complete routing steps(initial bind requests via LSS; further
> application requests to server directly) should be transparent to the
> Clients.
> - If a connection to the server fails, the client retries to get a new
> connection. Now, this request should be transparently sent to the LSS
> and the same cycle repeats.
>
> I would like to use such a load sharing mechanism, as I personally feel
> it is simply a performance overhead to route the application requests
> via the LSS, when the load sharing is done only based on the number of
> connections.
>
> Is it possible to implement such a load sharing mechanism using any of
> the routing protocols(E.g.VRRP)?


Your big problem is going to be moving the active end of a TCP
connection from the LSS to the target server. I've seen something like
this possible in some cluster configurations (at least between nodes in
the cluster), but not in a general environment.

In the configuration you're describing, the LSS is usually a proxy or
NAT device, and that's the way it manages to keep the connections
straight.

But why not do DNS based load balancing. Hand out a resolution with a
short TTL to the least loaded server. If the server fails, stop
handing out that address, and when the client retries, it'll get a new
one.