LVS
lvs-users
Google
 
Web LinuxVirtualServer.org

Re: load balancing and lvs

To: Brian Edmonds <bedmonds@xxxxxxxxxxx>
Subject: Re: load balancing and lvs
Cc: lvs-users@xxxxxxxxxxxxxxxxxxxxxx
From: Horms <horms@xxxxxxxxxxxx>
Date: Sun, 15 Oct 2000 03:47:58 -0400
On Wed, Oct 11, 2000 at 10:53:18AM -0700, Brian Edmonds wrote:
> "Ted Pavlic" <tpavlic@xxxxxxxxxxx> writes:
> > What can be done is to put half of your load balanced VIPs on one
> > director and half on the other.
> 
> I've been thinking, why not have two IPs per virtual service.  Thus in
> the haresources you might have:
>  lvs1 10.1.1.1
>  lvs2 10.1.1.2
> 
> And in the DNS you could have
> 
> service IN A 10.1.1.1
>         IN A 10.1.1.2
> 
> Then the DNS will (roughly) round robin clients (or sites, depending on
> the client's DNS setup and caching) between the directors, and the
> directors will load balance to the real servers.  I've not given it
> extensive thought, but I don't see any reason why this should not work
> in a very straightforward fashion.

In general using DNS for load balancing traffic between hosts
that are on the same network doesn't work particularly well.
Most of the problems relate to the TTL (time to live) of
DNS replies. Making the TTL low increases DNS traffic, increasing
response time and increasing the chance of a DNS failure. Making
the TTL too high makes the granularlity on balancing traffic
to coarse.

Additional problems are that BIND for one won't let you assign weights
to round-robin entries - though this may not be a problem depending
on the setup. A bigger problem is that DNS cannot determine
if the request was made on behalf of a single host or will
be cached and used by many users. This comes back to the TTL.

Using DNS is definately a useful tool, but IMHO load balancing
such as that provided by LVS is always going to give you
better control on how traffic flows through your network.

-- 
Horms


<Prev in Thread] Current Thread [Next in Thread>