On-line estimation of internet path performance: an application perspective

Estimating end-to-end packet loss on Internet paths is important not only to monitor network performance, but also to assist adaptive applications make the best possible use of available network resources. There has been significant prior work on measuring and modeling packet loss in the Internet, but most of those techniques do not focus on providing real-time information and on assessing path performance from an application standpoint. In this paper, we present an online probing-based approach to estimate the loss performance of a network path, and extend this estimate to infer the performance that an application using the path would see. The approach relies on a hidden Markov model constructed from performance estimates generated from probes, which is then used to predict path performance as an application would experience. The accuracy of the model is evaluated using a number of different metrics, including loss rate and loss burstiness. The sensitivity of the results to measurement and computational overhead is also investigated, and an extension of the base approach using a layered model is explored as a possible solution to capturing time-varying channel behavior while keeping computational complexity reasonably low. The results we present show that the approach is capable of generating accurate, real-time estimates of path performance, and of predicting the performance that applications would experience if routed on the path