Check out the new USENIX Web site. next up previous
Next: Forming Groups Up: Role Classification of Hosts Previous: Defining Similarity


Role Classification

The role classification problem is not difficult to solve in ideal situations, such as the network shown in Figure 1, in which two nodes that share the same logical role communicate with the identical set of machines. Clearly, such a situation does not reflect the connection patterns in typical enterprise networks. Three major challenges of the role classification problem are:
  1. Two hosts that share the same logical role may communicate with drastically different sets of machines.
  2. A host may potentially be classified into more than one role.
  3. The grouping results that network administrators desire may vary from network to network and therefore the role classification algorithm must provide flexibility for them to control its mechanics so that meaningful grouping results can be achieved.
In a typical network setting for a technology company, each lab or test machine may be dedicated to a single engineer. Thus, each of these lab machines, despite sharing the same role, can have a connection pattern that is very different from the rest of the lab machines. To be able to correctly group such machines together, the grouping algorithm must take into account the potential roles of neighboring hosts rather than comparing the neighbor sets. Furthermore, some hosts may potentially be classified into more than one role. For instance, there could exist a machine in the network in Figure 1 that communicates with both sets of machines with which many engineering machines and sales machines communicate respectively. In such cases, the connection patterns of hosts must be evaluated carefully to ensure that each host is grouped with other hosts with which it has the strongest similarity in connection habits. The role classification problem is not trivial for the aforementioned reasons. Not only does the computation of the similarity measure matter, but the process of how nodes are grouped based on the similarity values among node pairs is also important. The grouping algorithm consists of two phases: i) the group formation phase and ii) the group merging phase. The group formation phase identifies each group of hosts that have similar sets of neighbors using a simple similarity measure such as the one described in Section 3. The purpose of the group formation phase is two-fold: i) to efficiently identify various groups of hosts, each of which has drastically different overall connection patterns, and ii) to prepare for the second phase of the algorithm. The formation phase of the algorithm can efficiently find the desired partitioning for the example network in Figure 1 but may fail for many networks since it does not take into account the potential roles of neighboring hosts as explained earlier. In general, the group formation phase may generate a partitioning that contains more groups than desired. The group merging phase decides whether groups, produced by the formation phase, can further be merged using a much more sophisticated similarity measure. This phase provides network administrators with fine-grained control over the merging process so that the grouping results reflect their intuition of the network structure.

Subsections
next up previous
Next: Forming Groups Up: Role Classification of Hosts Previous: Defining Similarity
Godfrey Tan 2003-04-01