...Zhuoqing Morley Mao¹

Zhuoqing Morley Mao (email: zmao@cs.berkeley.edu) is a Computer Science graduate student at University of California, Berkeley. This work was done during her internship at AT&T Research Labs.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... Fred Douglis²

Current affiliation: IBM Research

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... identifier³

Our authoritative DNS server [6] allows host names to be wildcarded, so we can set an address for *.rd.example.com.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... image.⁴

We tested with the default setting without any special options. Some older versions of both browsers were also tested giving the same behavior.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... 112.74.197.163⁵

For privacy concerns, the IP addresses have been anonymized.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... 7,894⁶

Only 7,894 of all associations can be reached from both probe sites 2 and 3.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... them⁷

The main tradeoff here is fewer peering links traversed in multi-ISP CDNs versus less traffic between access and backbone routers as well as lower costs in single-ISP CDNs.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... query.⁸

Note, for fault-tolerance, most CDN DNS servers usually return multiple IP addresses. In this case, we pick the first one, since clients also typically choose the first IP address.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... locations⁹

In many cases, the router's DNS name has an indication of the geographic location [14].

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... metric.¹⁰

We were unable to include CDN Y in the traceroute experiment, since most of its CDN servers are unreachable using traceroute.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.