|
Hi!
On Thu, 05 Jul 2007, Nick Stephens wrote:
> So if you had to set this up and sell it as a reliable HA system to your
> boss(es), which software package would you use for this? Piranha,
> Ultramonkey, Keepalived?
We're currently using keepalived and vanilla 2.6 kernels (which
already have LVS, so no patching needed). We're also looking into
ldirectord since keepalived has given us some trouble.
As of currently, we have multiple sets of two LB nodes. Failover
is *not* done by any classic HA solution but by speaking BGP to
our central switches. We have a fair amount of BGP know-how in
the company so that seemed natural.
We've been handling several millions of connections per second
with this setup and have not yet hit a ceiling of any kind with
it.
I do recommend modern (i.e. Opteron or Core Duo) CPUs if for the
only reason that their associated chipsets and thus servers
usually are much better at handling large interrupt loads (think
coalescing). I have to admit though that we favour the Opterons
very much, so I can't really speak for the Core Duos.
As for spec, we've had a single loadbalancer see more than 1.5
million conns per second without degrading performance. Note
however, that the buffer zone is very small, as soon as you're
getting near the ceiling of interrupt rate, things are going to
deteriorate rapidly.
All of this is to be taken with a grain of salt of course. This
setup works for us. Your kind of workload, network, workflow
might warrant an entirely different approach.
Regards,
Tobias
--
In the future, everyone will be anonymous for 15 minutes.
|
