The New York Times article on Google is quite an eye-opener. I’ve been reading similar stuff recently about Google’s ginormous computing complex being built next to the Columbia river where the aluminium plants stand.
There comes a point when you scale some system enough that it becomes a fundamentally different beast. When access to cheap electricity starts to become a key consideration in your “computer” design, we’re on to something big. Whether or not Google continues to succeed and dominate, the trend will continue.
Going back to an old favourite theme, the end-to-end principle decried efforts to add value in the network because all too often you ended up with value-subtract instead. Only the edge nodes have the context to know what reliability, security or processing is required.
Google starts to challenge this idea, if somewhat indirectly. We’re seeing a new network architecture emerge, where storage and processing power is dotted around the network. Google isn’t the only player; Akamai have been doing this for ages, and you’re going to see home gateways start to get a lot smarter and also blur the “edge” definition.
Also, the topology of the network itself morphs in response to these resources; “all paths lead to Google” is hyperbole, but makes the point. When some “edge” nodes are millions of times bigger than others, we’re into a new game. Google itself starts to change what is connected to whom, in which case it’s hard to pretend that the connectivity is really “application agnostic” when it all points to Googleplexes dedicated to search and correlation functions. The raw bits start to acquire a taint of meaning merely from the conduits they pass through.
The chink in the end-to-end principle seems to be that certain problems require bringing very large data sets together, backed by very large storage and processing requirements to correlate and add value, and then more processing to push that data out. The idea of “intelligence at the edge” is almost a PC-centric architecture with an assumption of processing being dominated by local CPU, storage and memory with network communications (over a bus a thousand times slower) being the subsidiary case. We’re not invalidating it: just testing whether the idea scales from “point to point” to “multi-point” to “every damn point”.
The “GoogleNet” is not “intelligence in the network” in the traditional sense, I’d certainly grant; the bits are still wired up with dumb pipe. What perhaps we’re seeing is a hierarchy of design principles emerging:
Note that these are all just engineering and economic principles or guidelines: by testing the edges of their applicability we’re not saying they’re wrong in any way. I suspect I’m just spinning my intellectual wheels in re-discovering the Seven Fallacies of Distributed Computing, with Google being the new king of that universe.
The tricky bit is going to be working out the business model impact on telcos wishing to exploit any gaps in the end-to-end principle/IP abstraction. But I suspect you’re going to have to pay me to tell you the answer to that one :)
Posted by Martin Geddes at 11:43 AMTrackBack URL for this entry:
http://www.telepocalypse.net/cgi-sys/cgiwrap/mgeddes/MT/mt-tb.cgi/756.
Listed below are links to weblogs that reference Is the middle is the new edge?:
»
Stupid Networks and the Need for Application Performance Monitoring from IP Convergence: Beyond VoIP, Beyond Cost Savings
Folks, convergence is gaining real ground, enterprises globally are reaping the benefits of a single pipe providing voice, data and video.
These benefits are realized in both soft and hard cost savings.
Hard costs in that enterprises no longer need a sepa
[Read more]
I assume you are aware of the work of Barabasi, in his book "Linked". It seems that Google is fanatically pushing the power-law relationship between number of nodes and the amount of connections to that node up to the maximum limit.
Some real-world test of the paradigm....
Martin,
I need to challenge your stupid network concept as I'm seeing a need for smarter networks.
Although the concept of a packet, packet switching and routing is far simpler the legacy concept of understanding service to the network element - there's a growing need for more intelligence as to how a network is performing and relation to the application's the rely on it.
You can always bolt a protocol analyzer onto a network but they are costly and never around when something burped like last nite.
You could use a synthesized analyzer like Cisco SAA - but simulating traffic and a pattern can be fooled by real ebs and flows of impairments.
Brings me to utilizing a product like Visual Networks which offers a perspective on real traffic , and ability to look back in time as to why an application burped last nite.
My thoughts are that by adding intelligence (application layer) into a network - that the realization of convergence and IP is greatly enhanced.
Thoughts?