|
well with rr dns there's no real error handling. if one of those machines
goes down, nothing's going to stop traffic from trying to go to that ip
regardless, unless you use a method of dynamic dns and use really low time
to live fields i guess, but i won't go there. or use monitoring and set
something up so when one of those machines dies, another sends gratuitous
arps (fake, whatever) to take over its tasks.
it just seems to me like utilizing dns for something like load balancing
is pretty "hack"-ish, and not really something nameservice is supposed to
handle.
what about using lvs to direct traffic to set of lvs machines behind it?
i guess the flaw there is that the incoming traffic is still hitting 1
machine at first instead of 2. not sure if you'd even see any real
performance benefit there. i guess maybe if you kept the table on that
front machine small and using a simple rr algorithm or what not, traffic
could go through that machine pretty fast...? shrug.
i've never heard that google _is_ using lvs, so i assume they're not. we
could ask them!
-tcl.
On Fri, 9 Jun 2000, Jeremy Hansen wrote:
>
> Has anyone ever tried a combination of round robin dns and lvs machines?
> The VIP's on the lvs machines would be in a rr dns pool making like a
> second level of load balancing across lvs machines which could have either
> the same real servers or a different group of real server behind each lvs
> machine? Just throwing around the idea. Any reasons this would be cool?
> And reasons it would not?
>
> Anyone know if Google uses lvs for load balancing their huge cluster of
> machines?
>
> -jeremy
>
>
>
>
|
