Over a wide area network, ALMI has to cope with the dynamics of network paths, such as distortion of delay measurements and transient link failures. ALMI needs to prevent the multicast tree from diverging from an efficient construction. To demonstrate that ALMI is able to achieve a cost-efficient tree, we have conducted experiments over 9 sites scattered in both US and Europe.
The experiment was run as follows. We started ALMI at all 9 sites and configured the ALMI controller to re-calculate the multicast tree every 5 minutes. Simultaneously, we run traceroute from each site to every other site periodically, every 5 minutes. The output from traceroute provides us with a benchmark of the network delay experienced between nodes during our experiment. We then compare the total delay of an MST computed from the traceroute measurements to that of the ALMI multicast tree computed by the ALMI controller. For this experiment, we used the traceroute measured delay as the ALMI tree link cost in order to achieve a fair comparison. In other words, the comparison reflects only the difference of tree composition, excluding the distortion caused by delay measurements at the application level.
Figure 5 shows the result of a six hour test run of a single multicast session. Initially, the cost of ALMI multicast tree is very high, since the ALMI controller does not have a priori topological knowledge about group members and randomly connects members to each other at the beginning of the session. However, the ALMI tree cost was quickly brought down at the next re-calculation of the tree and stays close to the real MST cost, as the controller periodically gathers measurement reports from group members and updates the ALMI MST. There are two spikes in the ALMI MST, at time units 22 and 36 respectively. Analyzing the traces, we found that both points are caused by transient network failures. In the first case, one of a pair of two nodes, who are very close to each other, detects the other end as unreachable and connects to a much higher cost neighbor. In the second case, one node experiences temporary network failure and is timed out at the controller. The network recovers after approximately 15 minutes and the node re-joins the group but is randomly assigned a new parent. The presence of a new member, either at the session beginning or during the session, always introduces sub-optimality of the tree since they are randomly connected to the rest of the ALMI multicast tree. A more intelligent controller may be able to use one of the Internet services such as in [9,16,13] to estimate the topological information of a new member and initialize its connection more efficiently. We conclude from this experiment that ALMI is able to use application perceived delay to construct an efficient multicast distribution tree in a highly dynamic network environment.
