Second, we plan to improve existing DNS-based server selection algorithms by considering the properties of known client-LDNS associations for an LDNS that requests a server name resolution. The following characteristics of the associations can be explored based on data collected using our methodology: known client proximity to the LDNS, known client distribution, and hidden load factor.
Given a name resolution request from an LDNS, if the known client proximity to the LDNS is good, then a CDN server close to the LDNS would also be close to its clients. If the proximity correlation is low, known client distribution and client cluster request patterns would be considered. If the majority of HTTP requests belong to a single network cluster, finding a CDN server close to or within that network cluster would also be close to clients issuing a majority of requests. Along with these factors, the hidden load factor of the LDNS is also considered to select lightly loaded CDN servers for an LDNS with a large hidden load factor. If the proximity correlation is low between LDNS and its clients, then server selection is optimized using other metrics such as server load.
Finally, we would like to apply the results of this work to improving content distribution internetworking (CDI), which refers to the interoperation among multiple CDNs for additional flexibility. A prototype of CDI, called CDN Brokering [6], uses a DNS-based brokering mechanism to forward requests among DNS servers of the interoperating CDNs. As a third area of future work, we plan to improve CDN brokering algorithms by using hidden load factors and client-LDNS proximity information. The client-LDNS proximity findings in our work justify DNS-based brokering, because the majority of the clients and their LDNS belong to the same AS.