Wondering what is costs more? HTCP\ICP or frontend instance?

I had a question that I am almost sure I know the answer but I want to
hear somebody else about the issue.(please)

While LoadBalancing Couple instances of squid(machines or couple over
the same machine) using route policy there is a big benefit.
The benefit is that the "front" level of the load balancing is being
done on the route level in the kernel which is almost 100% SMP friendly
utilizing available CPU for the task.
(not talking about IPTABLES LB which is another thing)

So for 6-10 plus instances that each and one of them is listening on
different IP:port there is either a LB using route policy which takes in
account statically that each core can take up to let say 1.5k(just a
number) users tops and route the traffic towards this IP.
This is a full "square" balancing not taking anything more then IP level
but it is the simplest way to LB the traffic no overheads in a way.

Now these instances will probably will talk HTCP or ICP, HTCP preferred.
Now I am wondering what will "cost" more?
using one Frontend instance that will LoadBalance the traffic or plain HTCP?

I am not sure that I have taken the right things in account for the what
so called "calculation" but I asked about it in the past in other forms
and with less details.

What do you think also about this idea:
LB using BGP(or other protocol) to "infect" the LB router(s?) with the
right\LoadBalanced routes towards the proxy servers?

Eliezer
Received on Thu Nov 07 2013 - 20:00:17 MST