woody@SPARTA.COM (Robert "Woody" Woodburn) (02/13/91)
I'm involved with two different companies who both are distributed nationally and have several internet connections as well as their own corporate backbones. I have not seen a way of distributing DNS requests appropriately to the name servers. It would be nice for requests from west-coast machines that are topologically near to west-coast name-servers query those servers, rather than have extra traffic across cross-country backbones. Have I missed something, or is this a known problem that is being dealt with? woody@sparta.com woody@cseic.saic.com
almquist@JESSICA.STANFORD.EDU ("Philip Almquist") (02/17/91)
Woody, > I have not seen a way of distributing DNS requests appropriately to the > name servers. It would be nice for requests from west-coast machines that > are topologically near to west-coast name-servers query those servers, > rather than have extra traffic across cross-country backbones. > > Have I missed something, or is this a known problem that is being dealt > with? It depends on your definition of "topologically near". One definition is that servers which are "topologically near" respond more quickly than servers which are not. If you accept that definition then most full-service resolvers (e.g. named) have been doing this for years. If "topologically near" is a function of physical (as the crow flies) distance, it would be possible to include in the DNS information about the location (latitude and longitude) of name servers. This could go in some new record type that could (optionally) be included in any response containing an NS record. However, for this to be useful, it would have to be widely implemented and used, and hosts would have to know their own locations. Ie, this would probably not be a practical solution in the real world. You'd also find many who would argue that "topologically near" has little to do with physical distance. To me, the intuitively obvious definition of "topologically near" would be a function of how far (in miles and hops) the packet has to go. This (and most other definitions of "topologically near") could not be implemented without devoting far more bandwidth to the problem than we lose by using the current "suboptimal" approximation. That could change some day as the routing technology improves, but I wouldn't count on it. Philip