|
Hi,
> > Suppose a site has more than one POP(supersparrow terminology).
> > Every POP is an LVS.
>
> from the outside an LVS is just a server. You then need a director
> somewhere else to forward packets to each server=LVS. You are going
> to get into loops here. Or am I missing
> something?
Well , I did suggest a mechanism for avoiding loops.
Do you mean that such a mechanism is not possible?
1.> In the first mail I had suggested simply what Lars had mentioned:
i.e.
* more than one POPs
* each POP is a LVS
* there is a master director(which has the VIP).This director
does the load balancing amongst the LVS's(POPs).Hence two tier.
The master director may or may not have realservers behind it.
2.> In the second one I had modified Lars's idea a bit:
* more than one POP.
* each POP is an LVS
* The request is first load balanced by DNS RR or by one of the POPs
(the one that has VIP).
* Each POP(LVS-Director) has a redirection capability.
This may lead to loops but here my mechanism for loop avoidance takes
over. Lets just say that since the redirections would need IP encapsulation,
we have some information to avoid loops in the (outer IP) packets.
* This way , the master director
- wouldn't be needed (DNS-RR)
But DNS RR would simply continue directing even in case of a failure.
Well that means that a master director *is* required and is unavoidable.(?)
I am sorry for troubling you, I guess it was just a rush of blood.
I was simply trying to eliminate the master director and then implement
two tier load balancing with request 're'redirection.
As for the metric part of the idea,
we would need a comprehensive metric covering latency, CPU load, memory
etc. and the metric calculation and *request redirection* mechanism should
be kept independent.
The pseudo request mechanism would take up a lot of time. So I guess
there's no way to find the network wise closest(O.K. even approximately)
LVS other than touching on the routing information.
Is there a case for not re-inventing 'supersparrow' for LVS?
The FAQ says that no one would goto the pain of doing that because of the
success of supersparrow. But wouldn't a Load sharing algorithm based on a
metric derived from the routing information be useful as well?
( I am assuming a simple two tier architecture here as in '1' above(Lars's
idea))
Horms, could you please tell me what you think about this?
All I can say is that such a project would
a.> interest us
b.> be a very good learning experience
c.> may or may not be useful commercially.
As far as the availability of a BGP(or any other) router goes, we could
configure one for ourselves in the Lab.Couldn't we?
Hope I've been explicit enough with the terminology.
Thanks a lot
abhijeet
|
