Decentralized Server Selection Through Joint Proximity and Load Optimization
Abstract:
With the advent of "cloud computing" and the growth of popular Web
services, many networked services are replicated at multiple
geographic locations. Such distributed services face the challenge of
server selection -- that is, directing an incoming client request to
the appropriate server or data center, in the hope of reducing network
latency or carefully tuning server loads. To meet these potentially
conflicting goals, existing approaches use heuristics or rely on
central coordination to perform server selection. In this work, we
apply optimization theory to derive a simple, provably optimal, fully
distributed solution to the server-selection problem. Our approach
defines a global objective for a mapping service, and shows that
decentralized mapping nodes performing small amounts of local
computation and sharing limited information, can achieve the global
objective. We also perform experiments, based on a 24-hour trace of a
real operational CDN, that show that the distributed solution
converges very quickly in practice.