Review - Globally Distributed Content Delivery

This article talks about the technology behind Akamai's business, the
pioneer in content delivery networks. The business of Akamai is to deliver
content at the network's edge, giving clients faster response times to get
their content and decreasing the demand on content providers' origin
servers. The rationale behind it is that, Akamai provides capacity on-demand
to its customers on globally distributed fault-tolerant clusters and content
providers do not have to undergo the high costs of excess provisioning for
occasional flash crowd. Instead, they still provision for it, but this cost
that is shared among the various Akamai customers, making it a win-win
situation for everyone.

The strength of this article is that it outlines a few of the ideas that
made Akamai's edge caches successful. An interesting point is that it
illustrates how Akamai uses DNS redirection for load balancing. This DNS
redirection map is computed continuously to provide better response to flash
crowds. This computation is based on heuristics determining "nearest",
"available" and "likely" servers. The other neat idea is that of
fragmentation of dynamic content into different objects of independent
cacheability. The cacheability of static contents is based on a TTL value
and this TTL value depends on the type of object.

Since Akamai deals with secure content too by relaying authorization tokens
from client to origin server and vice-versa, it breaks the end-to-end
property that is desirable to the client. This could be a security threat to
clients; if Akamai was not handling sensitive data as the content providers'
would if they were handling the data. The article also fails to describe how
it handles data that needs to be written back to origin server.

The Akamai idea is a very good one in my opinion. At times, this article
seemed to be a bit of an advertisement campaign, saying how Akamai deploys
quality software, test them online, etc. The article could be improved with
more details like heuristic for TTL values, but again, this is proprietary
knowledge. An area of future work might be to couple content-based caching
within the highly replicated clusters at a single site to improve
performance.

 
Received on Thu Oct 06 2005 - 09:13:07 EDT