During this long cache winter independent software vendors have invested in developing solutions which may share the same name, yet address different needs – capacity savings vs. quality of experience (QoE) improvement, managed content vs. over-the-top (OTT), content specific vs. application specific vs. everything-HTTP vs. protocol agnostic, saving peering vs. saving transport, located at the network core vs. located at network edge.

This spaghetti of solutions triggered operators and ISPs to redefine their cache strategies. Whereas previously an ISP would just require a single file-caching system in the entry to its network, nowadays legacy file-caching is no longer effective for most of the Internet content (which is just not files anymore but rather fragments of streams, example here), and CDNs replace much of file-caching traditional role. The new environment requires new solutions.

Networking gear providers are taking a close look at this market, rehearsing before they play an active role. Speaking to such players I have witnessed significantly different points of view, perhaps emerging from the different positions such players come from.

Yet, they all share the same concerns – cache servers, as the name implies, serve content. They act as a “mini Internet in a box.” Placing such an entity in the network (a) puts constraints on routing and traffic engineering (b) requires endless goose chase after the evolving Internet, its protocols, formats and business logic. The independence of the network and the IT may be jeopardized by introducing application manipulators into the network.

Whereas such approach can be contained for managed content, it raises many concerns with regards to OTT. Content not managed by the operator should arrive from the content providers directly or through their representatives – namely CDNs. If you worry about peering or transport costs, expand your bandwidth using virtual capacity – a content-agnostic solution, equivalent to addition of bandwidth.

From the core southbound – let the network be a network. Make sure you network is efficient as possible by expanding it virtually for all bits without application, protocol or content discrimination. This way you get best of all worlds – optimal network, full design flexibility and zero application-level configuration.

Dear brother,
You are invited to my son’s wedding, two weeks from now.
Looking forward to meeting you.
Your brother.

After learning the price of each word he reasons that Dear brother is redundant since his brother is going to be getting the message by hand, and he already knows how much he loves him. The words you are invited are also redundant, since it is obvious that once there is a wedding his brother is invited, similarly Looking forward to meeting you can be deleted. The words my son’s are also redundant because it wouldn’t make sense to report another man’s wedding, and so the man ended up with the shorter telegram:

wedding in two weeks
which exactly captures the information he wanted to transfer to his brother.

Given a message, in information theory we try to assess how many bits are needed to encode this message, or in other words given an encoded message how many bits are redundant and can be deleted. To do that we first need to quantify the amount of data encoded in a message of bits. In information theory this is referred to as entropy. The higher the entropy the more data is encapsulated in those bits (therefore fewer bits are redundant). Given a message of symbols over an alphabet , of letters the entropy is given by this formula:

,

where is the probability of an arbitrary letter in to be . In the simple case, .

As can be easily seen the higher the entropy is the more random the message seems (i.e. There are less patterns in it), on the other hand, the message aaaaaaaaaa….aa has entropy 0 (remember that . For compression, we want to re-encode a message (with fewer bits) such that its entropy increases. In encryption (the other end of the spectrum), on the other hand, we want to re-encode a message (with the same number of bits) such that its entropy increases.

The joke we opened with goes on, as the man decides to delete the word wedding (because what other reason is there to be sending a telegram in the first place), and the words in two weeks (which is the appropriate time prior a wedding to be sending invitations), and so he returns home without sending any telegram at all.

The memorable scene in Monty Python’s Life of Brian (1979) perfectly reflects the Internet reality. Two millenia have passed, and we are a 7-billion individuals herd. We consume the same content – perhaps from different sources, using different methods and devices, at different times and locations – but we all do the same.

Self-similarity in Internet traffic is amazingly high. It is so high that within 10 minutes about 48% of a sizable-ISP-traffic is not new anymore; it is merely a combination of previously transferred content. An even more ego-oppressing finding suggests that not only that we are very similar to many other people, but from the whole huge Internet we are practically wandering in a domain no larger than 150 Gigabyte at any given time.

Does this mean that you can just cache 150GB and reduce 50% of the Internet load? Actually not. These 150GB are changing all the time, they emerge as fragments of diverse sessions from different sources and in different protocols. File caching would just not work. It takes a real-time bit-level system to leverage our herd-behavior.

Follow our future posts, in which we will share detailed and graphical findings.