That longtime favorite of movie cops, stoners, and Homer Simpson, the donut, is quietly reshaping the future of computing.
Well, sort of. Sugar-glazed, jelly-filled rings of delicious fried dough are not going to have much of an impact on computing, it's true. But something called the torus, which has the distinctive shape of a donut, looks like it already is.
Torus, it turns out, is not the Latin word for donut. It actually means cushion, or that round molding you see at the bottom of architectural columns. Mathematicians have a formal definition of the torus, and so do topologists, those mathematicians who specialize in studying the surfaces of objects -- how points in space are connected to each other, that is, which is where you start to get a hint of the donut's vital role in computing.
It turns out that the surface of a torus is a particularly good topology with which to build high-speed networks, like those needed to make highly-parallel computers tick. When dozens, hundreds, or even thousands of processing nodes need to have reliable, high-speed access to each other, the torus has proven to be a great choice.
In a way, the torus interconnect is a ring network, as used in underground metro-area networks for bringing traffic to buildings, but taken to the next level. Think of the surface, or skin of a donut (ignore, if you can, the powdered sugar). Now, cover it in a regular mesh of nodes, each one connected to its neighbors above and below and to the left and right.
You'll notice that in both directions -- around the larger circumference of the donut (circling its central hole) and around the smaller circle of the "body" -- there is no edge to this mesh. It wraps around to connect with itself in every direction. And that means there are always at least two paths between any two nodes, no matter if other nodes or connections fail. And so, the torus is an extremely hardy interconnect, able to tolerate faults yet move traffic with high bandwidth and low latency.
Several of the world's fastest supercomputers use torus interconnects: IBM's Blue Gene, the Cray XT3, and a couple of Fujitsu models, for instance. But the toruses inside these monsters have n-dimensional surfaces, or topologies, with more connections between neighboring nodes than the four available on the 3D delicacies that Homer Simpson dreams of. Fujitsu's K computer, with 80,000 nodes, uses a 6-dimensional torus interconnect (with the insanely healthy name of Tofu). IBM's BlueGene/Q model, used, among other things, to model nuclear explosions, relies on a 5D torus.
(Were he to eat a sufficient quantity of his favorite toroidial junk food, Homer might reach a state of mind in which he could actually visualize an n-dimensional donut in all its sugar-glazed glory, but that's probably beyond most of us.)
Toroidal interconnects are not limited to the outer reaches of supercomputing, however. This year, they started to hit the business market, namely in the form of a server from AMD called the SM15000. It packs as many as 512 x86 processing cores into a 10-RU chassis, all linked together over a 3D toroidal packet-switching fabric. This fabric brings 10Gbit/s of Ethernet bandwidth to each processor, enabling them to exchange data with each other and to move data in and out of disk drives and an outboard LAN.
Dough! AMD's low-latency fabric, acquired early this year when the chip company bought a firm called SeaMicro, replaces the usual set of Ethernet links and switches that typically connect traditional servers and disk systems in datacenters.
Andrew Feldman, former SeaMicro CEO and now GM and SVP of AMD's Data Center Server Solutions Group (DCSS), tells me that this fabric will eventually stretch beyond a single rack and thus create even larger pools of processors. Already, the fabric can link the processors to as much as 4 petabytes of onboard disk storage.
Introduced last summer, the AMD server is just right, Feldman says, for handling many of the tasks called for in web-scale computing: serving webpages, running Hadoop, handling NoSQL DBMS, and so forth. Running those jobs on traditional servers, which consume much more electricity and physical space, Feldman says, leads to waste and higher TCO.
And so, the lowly donut looks set to attack, if not satisfy, yet another bottomless appetite.