Chasing the Big-Data Server
As an EE Times reporter back in the 2000’s, I was a hardware guy, so I mainly followed the cloud computing hardware not the algorithms. I was trying to figure out what this sea change in our digital lives would mean for my engineering readers. What sort of systems would people build? What kind of chips would they need?
At first, the outlook seemed gloomy. Google was rumored
to have designed its own PC server motherboards to run its services. Its data
centers were so large -- so the story went – that cutting five cents on the cost
and a few watts on the power of each board would amount to millions maybe tens
of millions of dollars in savings.
But no one knew for sure because the Googles and
Amazons kept quiet. They didn’t want anyone else to learn the hard lessons of
cloud computing they were figuring out.
So, I started going to events where execs spoke about
these big data centers. Often the events didn’t produce an EE Times
story, but I knew I was on the trail of one.
At a panel discussion hosted by eBay someone mentioned a Berkeley computer science professor who did a sabbatical with Google. I made sure to find out his name afterwards.
Randy Katz turned out to be an excellent source. Like many other academics in tech, he was objective and opinionated. I asked him every question I could think of; he told me what he had learned.
The headline of my article was simply, “Goooogled,” mimicking the
company’s web site that put the extra o’s in its name to indicate the page
after page of search results it proudly delivered in those days.
The timing was good, given the rise of Web savvy
cellphones and TVs.
“There are something like three billion handsets now
becoming first-class citizens of the internet,” said Katz. “The phones are
limited in what they can do locally, so you will have an ever-increasing demand
for internet data centers,” he said.
In our 2008 interview, he estimated some 400,000
developers were already using Amazon's Elastic Computing Cloud service launched
in 2006, making it a de facto standard for business computing.
Even historically technophobic Hollywood was jumping
on the bandwagon. Top movie studios along with many tech giants were drafting
their own cloud computing standard for what we would later call streaming
media. The Digital Entertainment Content Ecosystem aimed to let movie makers serve
their customers directly for the first time, disrupting DVD and set-top box
makers as well as the more traditional movie houses that distributed their
shows.
It was a sea change for computer makers too, as I
noted in a sidebar that sketched out elements of one of Google's most secretive
products -- its unique PC server design. I didn’t have many details, but I had
the gist of it: the PC server business, a rare profitable chunk of the PC
business, was getting disrupted. The big data centers who were quickly emerging
as the some of the largest customers for PC servers had determined the current products
were overengineered, too costly, power hungry and heavy. A six-foot high rack
crammed with as many server boards as users could keep cool was becoming the
new unit of computing.
In April 2011, Facebook decided it was time someone in
the cloud business spilled the beans. It revealed its own low-cost designs and
launched a project to rally the industry around making them.
The
Open Compute Project included servers using highly efficient
power supplies on simplified systems. Dell was already delivering them to the
rapidly growing company that, along with Twitter, had become the face of social
media.
Facebook claimed its latest data center in Prineville,
Oregon, had set new records in its power efficiency and was among the first to
use fans and ambient air, rather than expensive air conditioners.
Dileep Bhandarkar, a tech executive who managed
Microsoft's data centers at the time, attended the Facebook event. He praised
the effort, but stopped short of joining it. “We have been sharing our best
practices [quietly with computer makers] for several years now,” said Dileep.
Eventually, Microsoft did join the rapidly growing
project. Dileep later left Microsoft to help Qualcomm, a designer of cellphone
chips, create its first server processors so it could get a slice of the cloud
computing business, too.
A little more than a year after its event, Facebook PR
asked if I’d like to meet its server engineers and see up close their designs,
already in a third generation. Of course, I was delighted to. EE Times ran
a story
with a slideshow of the Facebook systems and people behind
them.
Despite its big ambitions, Facebook’s small server engineering team was lean.
“Our philosophy is to keep small, fast teams -- huge teams move slowly,” said Matt Corddry (above), senior manager of Facebook’s 30ish-person hardware engineering group who formerly held a similar job at Amazon. “We typically have less than ten people per team. We love to go fast and have small teams with everything they need to break rules and build things quickly.”
A week later, Google followed suit, giving Wired
magazine an
exclusive on its systems and data centers. I followed up with a
story summarizing the news for my audience. Essentially it
proved Google was huge and was hugely disrupting the fat profit margins of
computer makers.
Google told Wired it deployed its one-millionth
server on July 9, 2008, and just one of its data centers included nearly 50,000
servers. It had already gone through a dozen generations of its own proprietary
server designs since it crafted the first Google server in 1999.
The first Google server cost $1,500, and was made from
parts sourced at local Silicon Valley electronics shops, saving a whopping
$3,500 on the cost of an off-the-shelf computer. Google claimed the stripped-down
design not only slashed costs and power consumption, it improved reliability.
Wired noted that Google was
not yet designing any of its own chips, but it was open to the possibility. In
short, Google and its peers were writing a new chapter in the history of
computing, making warehouse-sized data centers the next big platform and
designing the boards that went into them, too.
For its part, Facebook continued for years a drumbeat for
open-source data-center hardware, staging huge events with keynotes full of tech
news and exhibit floors crowded with demos. By March 2015, I
reported Apple, Bank of America, Cisco Systems,
Hewlett-Packard and Juniper Networks were among the members of its Open Compute Project,
“an effort to let big cloud computing operators like Facebook help set the
agenda for systems design.”
The social networking giant claimed since 2009 its
work had helped it save $2 billion. That was probably at least $2 billion off
the top line of the computer companies who read EE Times.
Comments
Post a Comment