Root
DNS
DNS
Server
Client
Internet
Figure 1: Conventional Content Routing
name server to receive the address of a nearby con-
tent server, incurring another round-trip time. Fi-
nally, it incurs the round-trip time to access the con-
tent on the designated server. If in this example
the client is located in Turkey, the first round-trip
is likely to go to Norway or London. The second
round-trip may have to travel as far as Redmond,
Washington, and the final might be to a content dis-
tribution site in Germany.
Thus, the conventional content routing design does
not scale well because it requires the world-wide
clients of a site, on a cache miss, to incur the long
round-trip time to a centralized name server as part
of accessing that site, from wherever the client is lo-
cated. This round-trip times are purely overhead,
and are potentially far higher than the round-trip
times to the content server itself; this long latency
becomes the dominant performance issue for clients
as Internet data rates move to multiple gigabits, re-
ducing the transfer time for content to insignificance.
These name requests may also use congested por-
tions of the network that the content delivery system
is otherwise designed to avoid.
DNS-based content routing systems typically use
short time-to-lives on the address records they re-
turn to a client, in order to respond quickly to
changes in network conditions. This places addi-
tional demands on the DNS system, since name re-
quests must be sent more frequently, increasing load
on DNS servers. This can lead to increased latency
due to server loads, as well as increased probability
of a dropped packet and costly DNS timeout. As
shown in [5], DNS lookup can be a significant por-
tion of web transaction latency.
Both of these problems can be ameliorated by intro-
ducing multiple levels of redirection. Higher-level
names (e.g., m.contentdistribution.net) specify
a particular network or group of networks and have a
relatively long time-to-live (30 minutes to an hour).
The records identifying individual servers (such as
s12.m.contentdistribution.net) expire in just
seconds. However, this increases the amount of work
a client (or a client’s name server) must perform on
a “higher-level” cache miss, and requires additional
infrastructure. Also, such a design conflicts with the
desire for high availability, since alternate location
choices are unavailable to a client with cached DNS
records in the event of network failure.
Conventional content routing systems may also suf-
fer from other availability problems. A system which
uses only network-level metrics does not respond
to application-level failure, so a client may be con-
tinually redirected to an unresponsive web server.
Designs which rely upon measurements taken from
server locations may also choose servers which are
not suitable from the client’s perspective, due to
asymmetric routing. A smaller, related, problem is
that DNS requests go through intermediate name
servers, so that the actual location of the client may
be hidden.
Finally, content routing systems may have difficulty
scaling to support multiple content provider net-
works and large numbers of content providers. Some
content providers (such as CNN.com) serve HTML
pages from a central web site but provide graphics
and other high-bandwidth objects from a content de-
livery network; the URLs of these objects are located
under the content delivery network’s domain name.
This has the advantage of increasing the probabil-
ity of DNS cache hits, since the same server location
information (akamai.net, for example) can be used
for other sites. However, it does not help increase
availability of the site’s HTML content or improve
latency to access it. Performing content routing on a
larger set of domain names in order to improve web
latency may result in lower DNS hit ratios, in ad-
dition to the costs of a larger database at a content
delivery network’s name servers.
Obtaining access to the network routing informa-
tion needed to perform content routing may also be
problematic. Content provider networks must ei-
ther obtain routing information from routers near
to their servers (via BGP peering or a proprietary
mechanism) or else make direct network measure-
ments. Both these schemes require aggregating net-
work information for scalability, duplicating the ex-
isting routing functions of the network. It may also
be politically infeasible to obtain the necessary in-
formation from the ISPs hosting content servers.
There is also no clear path for integrating access to
multiple content delivery networks. In order to do