[dns-operations] opting in to stupid DNS tricks
jabley at hopcount.ca
Wed Feb 23 13:06:15 UTC 2011
On 2011-02-23, at 00:38, Matthew Pounsett wrote:
> On 21-Feb-2011, at 06:23, Jim Reid wrote:
>> BTW, I still don't understand why CDNs are abusing the DNS to solve something that is actually a routing problem. What's wrong with anycasting the IP address(es) of the web site or whatever? That way, the network figures out the truly optimal path (peering policies aside) between the end client and the content provider's server. Yes, I realise this may break TCP connections sometimes, but how much of a real problem is this? Has anyone got hard data about this?
> As an anycast DNS provider, my answer to this is: because BGP optimizes for the shortest AS path, not the shortest path, or the lowest RTT, or the most bandwidth, or the most capacity. At $DAYJOB we give the same answer from everywhere, but we don't just rely on BGP to give great performance. Anycast gives availability, and helps to sink regional DDoS.. it does not provide performance.
I think this depends very much on what you're optimising for. For any of the clouds that I have had a hand in deploying or running, performance means availability, not RTT.
If you expect your responses to be cached and answered from the cache, then the incremental performance benefit in reducing the initial cache miss by 100ms is close to zero; the performance benefit in serving a (topological) region that otherwise can't get an answer is enormous, however.
More information about the dns-operations