Comparison of overlay construction algorithms

In Fig. , we compare the rejection probability, network usage, and server stress for BF, BF-delay, and BF-delay-approx, with the server placed at the transit domain. One observation is that both the BF-delay and BF-delay-approx outperform BF algorithm in terms of rejection probability and network usage. We also see that the performances of BF-delay and BF-delay-approx are close. Furthermore, we note that the server stress of BF is much less than that of BF-delay and BF-delay-approx. BF encourages the requesting client to connect to a client with the most abundant bandwidth, even if that client is farther away from the requesting client than other candidate clients. Since bandwidth is consumed over more links, this potentially increases the rejection probability for future arrivals and the overall network usage. Nevertheless, by pushing requesting clients to other clients, the server is less stressed.

Figure 13: Comparison of different tree construction algorithms: rejection probability

Figure 14: Comparison of different tree construction algorithms: network usage

Figure 15: Comparison of different tree construction algorithms: server stress

We also conducted experiments with the server in the stub domain, and similar results are observed [11].