To improve performance, large-scale Internet systems require clients to access nearby servers. While centralized systems can leverage static topology maps for rough network distances, fully-decentralized systems have turned to active probing and network coordinate algorithms to scalably predict inter-host latencies. Internet applications seeking immediate adoption, however, must inter-operate with unmodified clients running existing protocols such as HTTP and DNS. This paper explores a variety of active probing algorithms for locality prediction. Upon receiving an external client request, peers within a decentralized system are able to quickly estimate nearby servers, using a minimum of probes from multiple vantages. We find that, while network coordinates may play an important role in scalably choosing effective vantage points, they are not directly useful for predicting a client’s nearest servers.
Kevin P. Shanahan, Michael J. Freedman