A Measurement-Centered Approach to Latency Reduction

Increasing latency on paths through the Internet has a negative impact on latency-sensitive applications such as audio and video conferencing, as well as on the perception of responsiveness of many other applications. Measurement is key to addressing this issue, in two ways. First, the symptoms of excessive latency can be hard to isolate, and the causes obscure: the deployment of passive and active measurements of latency is crucial to any effort to address the latency problem. Second, while content and service providers have long attempted to address latency issues through the development of content delivery networks (CDNs), an emphasis on bandwidth as the primary measure and unit of comparison of Internet access performance has removed incentive for ISPs providing Internet connectivity (access providers) to pay much attention to latency. On this second point, any effort to reduce latency must include a re-emphasis on latency as a measure of access performance. This, in turn, requires advances in measurement. While they may be inaccurate, bandwidth measurements of access links are easy to understand, and correlate well with perceived access performance. Measuring latency is made more difficult by the need to choose a reference point to which latency can be measured, by nonlinear effects due to queueing, and to complexity introduced by CDNs. The development of such a metric for access provider comparison is as much an education and awareness-raising effort as it is a technical problem to solve; regardless, it is a necessary first step to providing a powerful incentive to reduce latency on the access segment. The latency problem is not by any means a new one. It impacts not just classical latency-sensitive applications such as two-way audio and video: if users of a website perceive too much delay in accessing that site, they are likely to move on. Therefore, 0 0.2 0.4 0.6 0.8 1