Application performance metrics for evaluating delay estimation schemes

Delay estimation schemes have been proposed to estimate the n
²
delays between n hosts using only O(n) measurements. Each scheme has been shown accurate when evaluated using an aggregated metric. However, their accuracy is low using metrics that evaluate their quality of estimations as experienced by the application users. Further, their use in several applications produces delay several times higher than optimal. This paper considers five classes of network applications and proposes three delay performance metrics to evaluate IDMaps, GNP, Vivaldi, and IDES when used in these applications. Our metrics measure the user-perceived performance difference of the applications when using estimation instead of direct delay measurement. Our simulations show that while these schemes have shortcomings, the majority of users receive good results for many applications.