Characterizing Internet Delay Asymmetry

Many real-time applications over the wide-area network perform proximity measurements to optimize their performance. These include voice over IP [14], multicast streaming (e.g., [7, 3, 4]), data transfer applications that perform locality-aware redirection and server selection (e.g., [15]), and system infrastructures such as proximityaware DHTs (e.g., [12]) and services such as positioning systems (e.g., [8, 5]). For some applications, the proximity of interests is the round-trip time (RTT) between two end hosts. For example, if the interaction between remote hosts typically involves only one or a few request and reply messages, e.g., a DNS lookup, or a small file download, then RTT is a good indication of the completion time of the interaction. For other applications, the proximity of direct relevance is the one-way delay from the client to the servers/peers (or along the reverse direction). For example, in multicast streaming applications, since the data always flows from a parent node to a child node in the overlay multicast tree, optimizing the oneway delay from the parent to the child is more beneficial. Measuring the one-way delay between two end hosts faces two major obstacles: it requires strict time synchronization between the two hosts and access to both hosts as there is no standard daemon in operating systems that measures and reports one-way delays. Thus, the common practice in most applications that rely on proximity information is to measure RTT and assume one-way delay (OWD) to be half of RTT. In this work, we investigate the validity and implications of this premise (and common practice). We study the extent and severeness of asymmetry in forward and reverse one-way delays between a pair of Internet hosts; we trace the reasons for the observed OWD/RTT relationship; and we investigate the dynamics of the OWD/RTT relationship. In the following, we present preliminary findings and lay out plans for future work.