Probabilistic guarantees on message delays over wide-area networks using in-situ instruments

Abstract

Messages transported over wide-area networks are subject to various delays at the intermediate nodes and hosts. In addition to bandwidth limits, the delays have an apparent "random" component due to the complicated dynamics of the network traffic. We consider that the messages sent over the network are subjected to three types of delays: (a) propagation delays along the links, (b) delays due to bandwidth availability on the links, and (c) "other delays" at the intermediate nodes which are randomly distributed according to unknown distributions. We propose an overlay network of in-situ instruments on such a network to collect delay measurements, and to compute and implement paths for message transport. We propose an algorithm to compute a path whose message delay is close to the optimal expected delay with a high probability, based entirely on measurements. We then present a multiple path method for achieving low end-to-end delays, which is implemented over the Internet using user-level daemons. These daemons realize multiple paths among themselves without explicit support from the underlying network routers, and achieve higher aggregated bandwidths compared to the usual transport methods.