Here are my conjectures on how Akamai works. These are based on some experiments done on April 4th, 2000 using mainly "dig". These were conducted from some machines at UW and one at MIT.

1. What an Akamaized page looks like?

Suppose you enter in your browser. This fetches the index.html file from the cnn server. In that file there will be images which will be pointing to the Akamai servers. Those URLs look like

Getting a web site in the above form is not too tough either. Akamai has a simple tool, Free Flow Launcher, for its customers that they use to Akamaize their pages. The users will specify what content they want to be served through Akamai and the tool will go ahead and Akamaize the URLs. This way the customers still have complete control of what gets served through Akamai and what they still are in charge of. Typicall all the dynamic content and stuff like transactions and cookies are taken by of by the customer's server only. Now the customer is responsible only for the content he choses to server himself and first few hits of other content till the Akamai caches warm up.

2. DNS Black Magic

I do not know the DNS system inside out, so the information here could be incomplete or simply wrong. Believe it at your own risk. Below is the chronology of steps that happen when an object from an Akamai server is to be fetched.

Step 1

From the top level domain you first get the name server of domain. Interesting things happen here itself. There are 8 name servers reported z[A-H]  This information obtained is good on the scale of days.

Below is part of "dig" output.
The columns are <Machine-Name> <Lifetime of cached information> <IN (for query-type)> <A (for query-class)> <IP Address.
(There is no need to worry about the query type and query class here.)

ZA.akamaitech.NET.      3h30m6s      IN A
ZB.akamaitech.NET.      3h30m6s      IN A
ZC.akamaitech.NET.      3h30m6s      IN A
ZD.akamaitech.NET.      3h30m2s     IN A
ZE.akamaitech.NET.      3h30m2s      IN A
ZF.akamaitech.NET.      1d44m5s      IN A
ZG.akamaitech.NET.      23h24m25s IN A
ZH.akamaitech.NET.      23h24m25s IN A

zA.akamaitech.NET.      1d16m48s IN A
ZB.akamaitech.NET.      1d16m48s IN A
ZC.akamaitech.NET.      1d16m48s IN A
ZD.akamaitech.NET.      1d16m48s IN A
ZE.akamaitech.NET.      1d16m48s IN A
ZF.akamaitech.NET.      1d16m48s IN A
ZG.akamaitech.NET.      1d16m48s IN A
ZH.akamaitech.NET.      1d16m48s IN A

A couple of points are of interest here.

Step 2

From one of the name servers above you go and get the name servers for domain Now this step is where most of the stuff happens. A few observations about the information returned at this step. Below is a part of dig output from MIT and one of the machines UW (the format of the output explained above)

n1g.akamaitech.NET.     27m48s IN A
n2g.akamaitech.NET.     16m45s IN A
n7g.akamaitech.NET.     31m44s IN A
n3g.akamaitech.NET.     18m14s IN A
n4g.akamaitech.NET.     18m14s IN A
n8g.akamaitech.NET.     18m14s IN A
n5g.akamaitech.NET.     18m14s IN A
n0g.akamaitech.NET.     18m14s IN A
n6g.akamaitech.NET.     18m14s IN A

n2g.akamaitech.NET.     46m13s IN A
n3g.akamaitech.NET.     16m13s IN A
n8g.akamaitech.NET.     16m13s IN A
n4g.akamaitech.NET.     16m13s IN A
n5g.akamaitech.NET.     16m13s IN A
n6g.akamaitech.NET.     16m13s IN A
n0g.akamaitech.NET.     16m13s IN A
n1g.akamaitech.NET.     31m13s IN A
n7g.akamaitech.NET.     31m13s IN A

Note that all the IP addresses are not unique in a set. My take is that this makes future expansions easier with changed restricted to lesser places.

My belief is that this is THE step where all the Akamai DNS magic is. It will hand out different set of name server IPs to client contacting from different IP addresses. How it determines the nearest set of servers from IP addresses is anybody's guess (proprietary, but allegedly, they use BGP peering with ISPs that host the Akamai cluster, thus giving them a rough estimate of the distance of requesting user from that site - courtesy Tommy Larsen). Apart from wire latency other factors they claim to consider are load on their servers and Internet congestion. They also claim to be able to monitor their servers in real time (once per second). This means that gives out different sets of name server IPs at different times, which explains the short lifetime (30mins - 1hour) of this information. For instance following is part of dig output from the same UW machine at different times (or even different machines at same time - see below).

n2g.akamaitech.NET.     46m11s IN A
n3g.akamaitech.NET.     16m11s IN A
n8g.akamaitech.NET.     16m11s IN A
n4g.akamaitech.NET.     16m11s IN A
n5g.akamaitech.NET.     16m11s IN A
n6g.akamaitech.NET.     16m11s IN A
n0g.akamaitech.NET.     16m11s IN A
n1g.akamaitech.NET.     31m11s IN A
n7g.akamaitech.NET.     31m11s IN A

@UW2 (same machine at a different time)
n3g.akamaitech.NET.     21m28s IN A
n4g.akamaitech.NET.     21m28s IN A
n6g.akamaitech.NET.     21m28s IN A
n5g.akamaitech.NET.     21m28s IN A
n0g.akamaitech.NET.     21m28s IN A
n7g.akamaitech.NET.     36m28s IN A
n1g.akamaitech.NET.     36m28s IN A
n2g.akamaitech.NET.     21m28s IN A
n8g.akamaitech.NET.     21m28s IN A

This has another nice (for Akamai load balancing), or not so nice (for things like organization wide caches) side effect. Different machines could be downloading the same object from different Akamai servers at the same time, if  those machines connect to different primary name servers within in the department (like our setup at UW-CSE).

Step 3

In the final step you go to one of the n[09] name server and get the IP address of the machine you are looking for (e.g.: The server will return two IP addresses for each machine name. For instance, see the partial dig output below  20S IN A  20S IN A

Observe that the lifetime of this information is a mere 20 seconds (which corresponds to one or two web pages viewed) . So after this time period you will go back to the Akamai name server to get the IP addresses.  What this means is that even if both the machines go down, it is highly unlikely that this will be seen by the client.  (assuming that the Akamai name servers find this out and return a different IP in a failure scenario).

It is likely that all the machines don't host the content of all the customers. Suppose that there are three servers (three different machines) - A,B and C at a particular site (a site will typically have multiple machines). And the customers are X, Y and Z. So A will host X,Y, B will host Y,Z and C will host Z,X. This kind of an arrangement has a two-fold advantage
1) No server has to host all the customers' content. Easing the load on it and also making the content serving faster.
2) If any one server goes down, no customer is fully disconnected as there is another server (potentially more) with its tree.

A customer's content could be present at more than two servers at a site, but it makes sense to return the same two machines (till they are up) because of file caching, the object does not have to retrieved from the disk most of the times.

Another observed feature is that all the name servers in the set (n[0-9] returned in above step, give the same two IP addresses for a queried server (like This could mean that the configuration of all the servers in a set is identical and multiple servers are there just for sharing the load.

Akamai caches (courtesy Neal Cardwell)
The akamai machines at the edge are PCs running Linux and a slightly modified version of the squid cache. They are doing on-demand caching rather than push-based replication.

Name server differences
An artifact of the above exploration is the observation that different versions of named  might be running in the department (@UW-CSE).
While (bs4) always returns the two IP addresses in different order (to get some sort of load balancing), (bs1) does no such thing.
And the one at MIT ( seems to be returning the IP addresses in a random order.

Akamai Documents (courtesy Jay BivensBob Devine)
  1. Free Flow (pdf)
  2. Technical FAQ (ps)
  3. General FAQ (ps)

As stated above, these are just conjectures. Corrections/Comments welcome.
Last Modified : 9/03/01