DNS Based Load Balancers

Wed Jul 5 00:27:53 UTC 2006

On Tue, 4 Jul 2006, Sam Stickland wrote:

> We work with a couple of different technologies here - our own GSS's, cache
> farms and also external CDNs (for overflow). This is currently and area that
> is currently under evaluation for a quite significant expansion.
>
> Are you able to give some kind of description as to the problems you
> experienced whilst using your own appliances? It would be very useful to be
> able to avoid making the same mistakes.

>From my experience with F5's GSLB product 3dns, my issues with 
geographic load balancing via an appliance can be reduced to the 
criteria they have available to decide what answer to give a query.

- Pings and traceroutes are both subject to rapid state change. 
Paths and latencies change for a number of reasons not related to 
network proximities. Traceroute hops in particular are a terrible 
metric to use in judging proximity, as it could be very easy for a 
14-hop path inside the US to trump a 4-hop transatlantic path. 
Pings/traceroutes also take a long time, and are only valuable for 
repeat queries from the same client, dumping the first on some 
default pool. Not so load balanced.

- BGP aspath length. This is actually probably the 'best' data that 
a geopraphical load balancing system can use. The data is detailed 
and metrics for any inbound connection are already in the 'db'. 
However, expecting a corporate IT or ops department to configure bgp 
peering on their load balancer is probably expecting a bit much. To 
the best of my knowledge, no appliance uses aspath length.

- Maps of RIR allocations and their geographic locations. OK. I can 
see how these might be useful for balancing traffic roughly across 
global regions, but the lack of granularity makes this a somewhat 
elaborate way to skin this particular cat. Also, we all know how 
well RIR allocation corresponds to actual location in the real 
world. At work, I am pleasantly surprised by 'geolocation' tools 
that claim my office in Redwood City is actually in Washington DC or 
London.

If you're looking to distribute traffic across several data centers, 
across many geographic regions, why not anycast a set of auth 
nameservers, with each pointing at their own data center in answers? 
This solution probably gives you a better 'correct' hit rate than 
any commercial appliance, and can be implemented yourself, or with 
the help of a commercial provider who specializes in this sort of 
thing. No vendor lock-in, no arm and a leg for a substandard PC in a 
rack mount case. (but you dont get the big fancy logo either)

matto

--matt at snark.net------------------------------------------<darwin><
   Moral indignation is a technique to endow the idiot with dignity.
                                                 - Marshall McLuhan