Towards accurate Internet distance prediction
Abstract
Internet distance (latency) prediction provides proximity information without extensive network measurements. The pairwise ( O( N2)) distances between N nodes can be estimated using only O(N) measurements. Predicted distances can be applied to improve the performance of many Internet applications: for instance, to select the closest server from multiple replicas, or to select nearby neighbors in constructing a topology-aware overlay network. In recent years, many algorithms have been proposed for predicting Internet distances. However, the impact of the prediction accuracy on applications has not been systematically studied. This thesis first considers Internet distance prediction from an application's perspective. Our study reveals that when prediction is used, the application performance can be dramatically worse than the application performance when measured distances are used. The accuracy of existing prediction algorithms is short of being satisfactory in order to achieve desirable application performance. We further investigate how to improve the distance prediction accuracy and the performance of prediction-based applications. First, we explore a selective measurement scheme that can significantly improve the accuracy of selecting the shortest link with a small number of measurements. In addition, we study the impact of landmark selection on the prediction accuracy. Our experience with various landmark selection schemes suggests that it may be fundamentally difficult to accurately predict short and long links using single coordinate per node. Based on this observation, we propose a hierarchical prediction approach that utilizes multiple coordinate sets at multiple distance scales, with the most suitable scale being chosen for predicting the target distance. Experiments show that the hierarchical approach is very promising for improving the accuracy of network distance prediction.
Degree
Ph.D.
Advisors
Hu, Purdue University.
Subject Area
Electrical engineering
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.