Results: Stretch vs. Infrastructure Size

Figure: The 90% stretch of MIP with triangular routing (``tri''), bidirectional tunneling (``bi''), and ROAM. In all plots, $x$-axis represents the number of servers (home agents or $i3$servers) on a log scale, and $y$-axis represents the latency stretch. Each plot corresponds to a mobility model (``m'') and a communication model (``c''). ``uniform'', ``Pareto home'', and ``Pareto foreign'' indicate how the location is chosen, i.e., according to a uniform distribution, Pareto distribution from home network, and Pareto distribution from the foreign network.

Figure 11 shows a series of graphs which compare the 90th percentile stretch of MIP with triangular routing and bidirectional tunneling (``bi''). Each graph shows a different combination of mobility model and communication model.

In the transit-stub network, ROAM matches or exceeds MIP's stretch when more than 1-2% of the transit-stub domains have a server. ROAM matches MIP when one or both of the communication end points (the MH and CH) is close to the HN. We expect that these are the optimal cases for MIP. Indeed, Figures 11 (b), (d), (e), and (f) show that MIP's stretch drops sharply as the number of deployed HAs increases. More HAs increase the likelihood that a HA will be in the HN, thus decreasing the stretch incurred by triangular routing or bidirectional tunneling when the CH and/or MH are close to the HN. However, the figures also show that ROAM's stretch converges with MIP's when more than 50-100 servers are deployed in the network (corresponding to 1-2% of the transit-stub domains having a server). This is because ROAM is able (through its trigger server caching algorithm) to dynamically find $i3$servers which are as close to the MH and CH as a statically configured HA.

ROAM significantly improves on MIP's stretch when neither the MH or CH are close to the HN. We expect this to be the worst case for MIP. Figures 11(a) and (c) validate this. Increasing the number of HAs in these cases does not decrease MIP's stretch because having a HA close to the HN does not put it any closer to the CH or MH. In contrast, ROAM's stretch decreases as more servers are deployed because it can still dynamically find closer trigger servers. Figure 11(a) shows that even when the CH, MH, and HN form a triangle with equal distribution of distance on each leg, ROAM's stretch is 40% that of MIP. When the CH and MH are Pareto close (as shown in Figure 11(c)), then ROAM has a stretch $1/400$th that of MIP with triangular routing. The difference is so large because the maximum latency in our transit-stub topology is over 1000 ms, while the minimum latency is only 1ms, so the impact of poor routing is very large.