How a data center is like a disk drive
Masters also said that the shift towards hyperscale computing is forcing the change in the data center. Big apps require massively scalable, cut-down systems where the redundancy is in the software and in the quantity of hardware, not in any particular server that is equipped with all kinds of redundancies, because this cuts overall acquisition and operation costs.
The companies that Red Hat is talking to in the hyperscale data center racket are looking at "fail in place" scenarios in which they treat a data center like a disk drive, and while they admire the compact and tuned nature of SoC server nodes with integrated switch fabrics, they are not even thinking about things at the rack level anymore, but at the data center level.
With a fail-in-place data center, you load it up with a few tens of thousands of nodes that have networking on each server node, pipe in external networking and power, and you never do maintenance on it. If a server fails, you mark it as bad, like a bad block on a disk drive, and you just leave it in there and let the network heal around it.
At some point, all the server nodes are dead, but by then you have built more data centers with shiny new servers with even better performance and thermals. Presumably, you eventually recycle all the internal parts.
So just how big is the potential market for ARM servers? Well, if you just hitch it to the hyperscale data center market, says Wesselman, then that bit of the server racket is accounting for anywhere between 20 and 25 per cent of server volumes a year and, to boot, it is also the fastest growing part of the server business.
"There is a lot of interest in that segment," says Wesselman, and by the way, "this is the segment that is addressing mobility. Hyperscale requires extreme quantities of servers at massive density, power efficiency, and easy of manageability across all those different systems. And for these reasons we need to look at new technology at the processor level and even at the total system and get more integration to reduce costs and to simplify the solution."
The one thing everyone wants to know is what is the potential for ARM-based server sales? Wesselman says that the research that HP relied on as it was putting together its Project Moonshot hyperscale server effort in 2011 was based on publicly available ARM roadmaps of the time and put together by Stacy Rasgon of Wall Street analyst firm Sanford Bernstein.
Rasgon came to the conclusion that by 2015, ARM-based servers could address anywhere from 15 to 20 per cent of shipments and therefore get maybe 10 to 12 per cent of revenues. That is not actual an revenue prediction, but merely a statement about the part of the business where such ARM-based machines could play.
"This is a whole new segment, and frankly, nobody knows how big and how fast this is all going to take off," said Wesselman. "Sort of like how the market had no ability to forecast what the iPhone or the iPad would do – and we have all seen how that history has transpired."
What no one said on the chat was that for ARM servers to get that business, it would have to take a very, very big chunk of the hyperscale server business away from Intel. Even as fast as that part of the server racket is growing, there is no other way for the math to work.
The irony, of course, is that you will essentially have a cluster of ARM-based nodes with funky networking and access to storage, and with not much more oomph than a smartphone or tablet serving up data to smartphones and tablets. Or, for another way of looking at it, isn't that a perfect kind of symmetry? No more big iron, just big fabrics.
John Chambers just fell off his chair, and so did Ginni Rometty.
The other important thing is that this original market forecast may not reflect some of the beefier ARM chip designs that are coming down the pike from Calxeda, Applied Micro Circuits, Nvidia, AMD, Cavium, and maybe even Samsung.
"The truth is, we don't know," conceded Ferguson, who in his job at ARM is wrestling with that question night and day. "We are at a very exciting stage. We really are empowering people to do innovation on hardware and around software."
If you add up all of the server, storage, and networking market, the total addressable market is north of $100bn annually, and a chunk of each one of these is up for grabs. In some cases, with ARM processors and integrated switching (not network interfaces, but switching), you get rid of switches entirely.
Chambers just fell off his chair again, and Andy Bechtolsheim looked up from his desk and started daydreaming about his own ARM superchip to fit inside of Open Compute enclosures that did everything anyone could ever want.
The important thing is to break free of the Wintel duopoly on the server, said Shuttleworth, as ARM has done on mobile devices, and to not get trapped again on servers.
"At the risk of being controversial," Shuttleworth said, "one of the reasons why the existing server ecosystem has found itself constrained is because there are parts of the stack that are immovable. If you have to support Windows, then you are forced to look, and act, and present to the OS in exactly the same way. And in every ecosystem where that's true, people feel it is unhealthy.
"We have to be very careful not to throw the baby out with the bathwater. In mobile ecosystems today, we have a tremendous amount of diversity at a certain level. People are concerned that there is not enough diversity further up the stack, and that is making it very hard for different manufacturers to differentiate."
Canonical, which is the company behind the Ubuntu Linux distribution, has certainly thrown itself into ARM servers. It's the first company to offer several releases of Linux supported on ARM platforms, and is eager to catch the next wave in hyperscale data centers.
"This industry does tear up the rulebook every generation," said Shuttleworth. "And we are starting to see that this scale-out computing shift is a generational shift – on the order of the generational shift from mainframes to microcomputers to the modern server and the shift to fabric-oriented hyperscale is certainly on the path.
"And there are ways to prepare for that shift, today, by moving to cloud computing and moving to scale-out versions of existing applications and starting to think about managing down legacy infrastructure that doesn't get that new way of working." ®