Multicast vs. unicast for loss tomography on tree topologies

Abstract

Loss tomography using multicast measurements and using unicast measurements have been investigated separately. In this paper we compare the performance of the two methods on tree structures. We prove identifiability of unicast measurements on tree structures with no degree-2 nodes. To theoretically compare multicast and unicast, we develop an observation model for multicast on trees and derive expressions for calculating the Fisher Information Matrix. We apply optimal experiment design for unicast on trees and develop a simple and insightful solution. Using a packet level simulator, we evaluated and compared the per-link MSE of multicast and unicast under varying parameter settings including link weights, link success rates and tree size. The results show that in contrast to the general belief that multicast always outperforms unicast, unicast can outperform multicast under tight constraint on the probing budget, especially in terms of a weighted average of per-link MSEs. On the other hand, multicast achieves more consistent performance with respect to varying link success rates or tree size.