The DNS Round-Trip Times method relies on each local name server tracking the round-trip time of packets as it performs recursive name service resolution. Over time, the local recursive name server will favor authoritative servers with the lowest latency. This "optimization strategy" results in resolution being performed against the authoritative DNS server "closest" to the client's name server with respect to the networkor so the theory goes.
Rephrased and simplified a bit, the client attempts to resolve images.example.com from each authoritative name server, eventually settling on the name server with the quickest responses.
So, all we need to do is put a publicized, authoritative name server for images.example.com alongside each image cluster at our four locations. Each name server will serve the IP addresses of the onsite image servers. As shown in Figure 6.9, the name server in Germany serves the IP addresses for the cache servers in Germany, the name server in Japan serves the IP addresses for the cache servers in Japan, and so on.
Figure 6.9. Geographically separate image serving clusters with DNS RTT-based convergence.
This method is easy to implement and requires no special infrastructure, but it suffers from some rather serious shortcomings. The most obvious is that the algorithm calls for the "eventual" convergence on the closest name server, and until each client's local name server converges, you have misdirection. Additionally, sudden changes in network topology, such as a collapsed route or the peering relationship between two autonomous systems (ASs) changing, will cause the answer we spent so much time reaching to be suboptimal. The Internet is a rapidly changing placea sea of endpoints. Point A will most likely always be able to communicate with point B, but the routers and networks those packets traverse can change at any moment.