by Steven Levy
April 17, 2012

from Wired Website



In early 1999, an associate computer science professor at UC Santa Barbara climbed the steps to the second floor headquarters of a small startup in Palo Alto, and wound up surprising himself by accepting a job offer.


Even so, Urs Hölzle hedged his bet by not resigning from his university post, but taking a year-long leave.

He would never return. Hölzle became a fixture in the company - called Google.


As its czar of infrastructure, Hölzle oversaw the growth of its network operations from a few cages in a San Jose co-location center to a massive internet power; a 2010 study by Arbor Networks concluded that if Google was an ISP it would be the second largest in the world (the largest is Level 3, which services over 2,700 major corporations in 450 markets over 100,000 fiber miles.)

‘You have all those multiple devices on a network but you’re not really interested in the devices - you’re interested in the fabric, and the functions the network performs for you,’ Hölzle says.

Google treats its infrastructure like a state secret, so Hölzle rarely speaks about it in public.


Today is one of those rare days: at the Open Networking Summit in Santa Clara, California, Hölzle is announcing that Google essentially has remade a major part of its massive internal network, providing the company a bonanza in savings and efficiency. Google has done this by brashly adopting a new and radical open-source technology called OpenFlow.

Hölzle says that the idea behind this advance is the most significant change in networking in the entire lifetime of Google.

In the course of his presentation Hölzle will also confirm for the first time that Google - already famous for making its own servers - has been designing and manufacturing much of its own networking equipment as well.

“It’s not hard to build networking hardware,” says Hölzle, in an advance briefing provided exclusively to Wired. “What’s hard is to build the software itself as well.”

In this case, Google has used its software expertise to overturn the current networking paradigm. If any company has potential to change the networking game, it is Google.


The company has essentially two huge networks:

  • the one that connects users to Google services (Search, Gmail, YouTube, etc.)

  • another that connects Google data centers to each other

It makes sense to bifurcate the information that way because the data flow in each case has different characteristics and demand.


The user network has a smooth flow, generally adopting a diurnal pattern as users in a geographic region work and sleep.


The performance of the user network also has higher standards, as users will get impatient (or leave!) if services are slow. In the user-facing network you also need every packet to arrive intact - customers would be pretty unhappy if a key sentence in a document or e-mail was dropped.

Urs Hölzle

Photo provided by Google.

The internal backbone, in contrast, has wild swings in demand - it is “bursty” rather than steady.


Google is in control of scheduling internal traffic, but it faces difficulties in traffic engineering. Often Google has to move many petabytes of data (indexes of the entire web, millions of backup copies of user Gmail) from one place to another.


When Google updates or creates a new service, it wants it available worldwide in a timely fashion - and it wants to be able to predict accurately how quickly the process will take.

“There’s a lot of data center to data center traffic that has different business priorities,” says Stephen Stuart, a Google distinguished engineer who specializes in infrastructure.


“Figuring out the right thing to move out of the way so that more important traffic could go through was a challenge.”

But Google found an answer in OpenFlow, an open source system jointly devised by scientists at Stanford and the University of California at Berkeley.


Adopting an approach known as Software Defined Networking (SDN), OpenFlow gives network operators a dramatically increased level of control by separating the two functions of networking equipment: packet switching and management.


OpenFlow moves the control functions to servers, allowing for more complexity, efficiency and flexibility.

“We were already going down that path, working on an inferior way of doing software-defined networking,” says Hölzle. “But once we looked at OpenFlow, it was clear that this was the way to go. Why invent your own if you don’t have to?”

Google became one of several organizations to sign on to the Open Networking Foundation, which is devoted to promoting OpenFlow. (Other members include Yahoo, Microsoft, Facebook, Verizon and Deutsche Telekom, and an innovative startup called Nicira.)


But none of the partners so far have announced any implementation as extensive as Google’s.

Why is OpenFlow so advantageous to a company like Google? In the traditional model you can think of routers as akin to taxicabs getting passengers from one place to another. If a street is blocked, the taxi driver takes another route - but the detour may be time-consuming. If the weather is lousy, the taxi driver has to go slower.


In short, the taxi driver will get you there, but you don’t want to bet the house on your exact arrival time.

With the software-defined network Google has implemented, the taxi situation no longer resembles the decentralized model of drivers making their own decisions. Instead you have a system like the one envisioned when all cars are autonomous, and can report their whereabouts and plans to some central repository which also knows of weather conditions and aggregate traffic information.


Such a system doesn’t need independent taxi drivers, because the system knows where the quickest routes are and what streets are blocked, and can set an ideal route from the outset.


The system knows all the conditions and can institute a more sophisticated set of rules that determines how the taxis proceed, and even figure whether some taxis should stay in their garages while fire trucks pass.

Therefore, operators can slate trips with confidence that everyone will get to their destinations in the shortest times, and precisely on schedule.

Making Google’s entire internal network work with SDN thus provides all sorts of advantages. In planning big data moves, Google can simulate everything offline with pinpoint accuracy, without having to access a single networking switch. Products can be rolled out more quickly. And since “the control plane” is the element in routers that most often needs updating, networking equipment is simpler and enduring, requiring less labor to service.

Most important, the move makes network management much easier.

By early this year, all of Google’s internal network was running on OpenFlow.

‘Soon we will able to get very close to 100 percent utilization of our network,’ Hölzle says.

“You have all those multiple devices on a network but you’re not really interested in the devices - you’re interested in the fabric, and the functions the network performs for you,” says Hölzle.


“Now we don’t have to worry about those devices - we manage the network as an overall thing. The network just sort of understands.”

The routers Google built to accommodate OpenFlow on what it is calling “the G-Scale Network” probably did not mark not the company’s first effort in making such devices.


(One former Google employee has told Wired’s Cade Metz that the company was designing its own equipment as early as 2005. Google hasn’t confirmed this, but its job postings in the field over the past few years have provided plenty of evidence of such activities.)


With SDN, though, Google absolutely had to go its own way in that regard.

“In 2010, when we were seriously starting the project, you could not buy any piece of equipment that was even remotely suitable for this task,” says Hotzle.


“It was not an option.”

The process was conducted, naturally, with stealth - even the academics who were Google’s closest collaborators in hammering out the OpenFlow standards weren’t briefed on the extent of the implementation.


In early 2010, Google established its first SDN links, among its triangle of data centers in North Carolina, South Carolina and Georgia. Then it began replacing the old internal network with G-Scale machines and software - a tricky process since everything had to be done without disrupting normal business operations.

As Hölzle explains in his speech, the method was to pre-deploy the equipment at a site, take down half the site’s networking machines, and hook them up to the new system. After testing to see if the upgrade worked, Google’s engineers would then repeat the process for the remaining 50 percent of the networking in the site. The process went briskly in Google’s data centers around the world.


By early this year, all of Google’s internal network was running on OpenFlow.

Though Google says it’s too soon to get a measurement of the benefits, Hölzle does confirm that they are considerable.

“Soon we will able to get very close to 100 percent utilization of our network,” he says.

In other words, all the lanes in Google’s humongous internal data highway can be occupied, with information moving at top speed.


The industry considers thirty or forty percent utilization a reasonable payload - so this implementation is like boosting network capacity two or three times. (This doesn’t apply to the user-facing network, of course.)

Though Google has made a considerable investment in the transformation - hundreds of engineers were involved, and the equipment itself (when design and engineering expenses are considered) may cost more than buying vendor equipment - Hölzle clearly thinks it’s worth it.

Hölzle doesn’t want people to make too big a deal of the confirmation that Google is making its own networking switches - and he emphatically says that it would be wrong to conclude that because of this announcement Google intends to compete with Cisco and Juniper.

“Our general philosophy is that we’ll only build something ourselves if there’s an advantage to do it - which means that we’re getting something we can’t get elsewhere.”

To Hölzle, this news is all about the new paradigm.


He does acknowledge that challenges still remain in the shift to SDN, but thinks they are all surmountable.


If SDN is widely adopted across the industry, that’s great for Google, because virtually anything that happens to make the internet run more efficiently is a boon for the company. As for Cisco and Juniper, he hopes that as more big operations seek to adopt OpenFlow, those networking manufacturers will design equipment that supports it.


If so, Hölzle says, Google will probably be a customer.

“That’s actually part of the reason for giving the talk and being open,” he says.


“To encourage the industry - hardware, software and ISP’s - to look down this path and say, ‘I can benefit from this.’”

For proof, big players in networking can now look to Google.


The search giant claims that it’s already reaping benefits from its bet on the new revolution in networking. Big time.