End-User Mapping: Next Generation Request Routing for Content Delivery – Public Review

To speed web delivery, websites use content delivery networks (CDNs) to deliver content from servers around the world. How does the CDN direct a client to a nearby server? One approach is to use DNS to return different IP addresses depending on which client queries. This approach is complicated by the DNS architecture, in which the CDN's authoritative name server communicates only with the client's local recursive resolver (LDNS), not the client, and so knows the LDNS IP address, but not the client IP address. If an LDNS only serves clients with similar network locations, then the CDN can direct to a server that should work well for all users. Unfortunately for DNS-based redirection, some LDNS, including public resolvers like Google Public DNS and OpenDNS, serve distributed clients. For such an LDNS, a CDN cannot use standard DNS to select a server that will be optimal for all clients. Partly to overcome this limitation, some content providers have turned to anycast to direct client traffic using BGP routing, rather than DNS redirection. To instead overcome the challenge within DNS, a number of companies proposed the edns-client-subnet DNS extension (currently an IETF draft) in which an LDNS passes a prefix of the client's IP address to the authoritative name server. Previous work showed some CDNs had adopted the extension, but little was publicly known about the performance improvements it enabled. This Experience Track paper describes Akamai's roll-out of end-user (as opposed to LDNS) redirection using edns-client-subnet. While the challenge of mapping clients based on their LDNS was well known, one of this paper's contributions is quantifying the problem for, in essence, all Internet clients in an era that has seen the rise of public resolvers. The paper then measures the performance improvement seen by clients of public re-solvers once Akamai started using end-user mapping. To make the results relevant beyond the particulars of Akamai's deployment, the paper finishes with a study that compares, for different CDN sizes, the impact of adding more sites versus enabling end-user mapping. The SIGCOMM Program Committee was very enthusiastic about this Experience Track paper: the results are new to the research community and based on real client measurements, and even few companies can make the measurements afforded by Akamai's high query volume and large worldwide server deployment. The paper presents actual production performance and load numbers , rather than obfuscating them. Despite the enthusiasm, the Program …