Comparisons of proximity metrics

Given the above set of metrics for evaluating proximity between client and its local DNS server, we compare their results on a common set of 7,894⁶client-LDNS associations in Table 9. The comparison shows that network clustering is a fine-grained metric, similar to traceroute divergence (TD) count of 1. Hosts within the same network cluster, or which have a TD of 1, are guaranteed to be very close to each other. However, hosts not in the same network cluster, or have a TD bigger than 1, may still be quite close. Thus, these two metrics are quite conservative. AS clustering is the most coarse-grained metric, since an AS can be quite large. This is comparable to the ratio of common to disjoint path length. RTT correlation is also a relatively coarse-grained metric. It is inconclusive and largely dependent on the two probe site locations.

In general, performance-oriented metrics such as round-trip time should provide accurate real-time network latency measurements. CDNs often do real-time network measurements from their servers to clients. Since we can only probe from a limited set of locations, such metrics are inconclusive. Topology-oriented metrics have the advantage of being non-invasive, since they do not incur any network overhead. However, they cannot take network congestion into account.

As we explain in the following section, the applicability of each metric depends on the density of CDN server placement. The denser the placement, the more fine-grained metric is needed.

Table 9: Comparison of four proximity metrics

Proximity metric	Evaluation
AS clustering	78% in the same cluster
Network clustering	23% in the same cluster
Traceroute divergence	16%: TD=1, 32%: TD=2
(TD)	median TD=4, mean TD =5.7
(probe site 2)	65%: $disjointPathLen$
	$\le commonPathLen$
RTT correlation	71%: $t_d^2<t_d^3 \Rightarrow t_c^2<t_c^3$
(probe sites 2, 3)	62%: $\vert t_d^2-t_d^3\vert\le10ms \Rightarrow$
	$\vert t_c^2-t_c^3\vert\le10ms$
	$a=t_d^2-t_d^3$, $b=t_c^2-t_c^3$
	$correl(a,b)=0.13$