Coverage

Informally, we say an aggregate LP provides a coverage of its participant individual LPs if the following hold:

Figure 1: Example of Factors that Impact Coverage of Clusters

• The Internet distances between each client/content server pair, in the cluster, is comparable. An Internet distance can be determined based on parameters such as available capacity of links and network topology.
• The path(s) between clients and content servers in a cluster must be comparable with respect to points of congestion.
• Source feature vectors extracted from individual latency profiles must be comparable.
• The paths between pairs of clients and servers must exhibit similarity of client and server ASes, and BGP based AS-PATH routes.

Figure 1 illustrates some clients and servers. Suppose the Internet distance between the client and the three servers in the figure are presumed similar. A router located close to server S₂ acts as a potential point of congestion, and therefore we may conclude that server S₂ should not be included in the cluster. Similarly, we observe that server S₃ is in a different time zone from the other servers, excluding server S₃ from the cluster, due to the significance of time of day in latency estimation.

Determining the Internet distance or identifying points of congestion may be intrusive and expensive. In contrast, BGP routing information is often publicly accessible and is often stable over relatively long periods of time. Therefore, our initial efforts are focused on two factors, namely the feature vectors from latency distributions that were used by the WebPT, and BGP based AS-PATHs between clients and servers.