named

[Dan Ritter]

On Thu, Dec 02, 2010 at 10:16:58AM -0500, Richard Pieri wrote:
> On Dec 2, 2010, at 1:28 AM, Derek Martin wrote:
> A bunch of stuff that my experience does not bear out.  In over 20 years of being a DNS admin with a variety of ISPs involved, I have rarely seen a local DNS cache prove to be inferior to going out to the ISP for every lookup.  To the contrary, running a local cache provides superior lookup times for all but the first lookup in all but the most dodgy edge cases.

This is true.

> The CDN argument falls flat.  If the CDN looks at the ISP's source addresses on DNS lookukps then it still isn't going to get you something geographically local.  A node on Speakeasy in Maine gets a list of name servers geographically located in New York City even though Boston is closer.

If you have Speakeasy DSL in Maine, your next network hop is probably
ALWAYS in NYC even though Boston is geographically closer -- because
that's how Speakeasy's physical network is set up. The network topology
determines performance.

But the CDN generally looks at the source address of the incoming DNS
request, and it does so according to maps which are pretty good at
relating the network topology rather than the geographic distance.

> 
> The super-short TTL argument falls even flatter.  Suffice it to say that DNS roulette is not load balancing.  There have been many papers written about why it is bad practice.  That Akamai uses DNS roulette doesn't make it better.  In fact, it's a lot of why Akamai is derided for making the web go slow.

Akamai doesn't use short TTLs for load balancing. Akamai uses short
TTLs for failover, so that your preferred node going down doesn't
simultaneously kill Apple, CNN, the NYTimes, and all their other customers
and end up causing you to complain to your ISP that half the Net is down.

-dsr- (last worked at Akamai almost 10 years ago)

-- 
http://tao.merseine.nu/~dsr/eula.html is hereby incorporated by reference.
You can't defend freedom by getting rid of it.