DatacenterDynamics coined the term ‘zettastructure’ to express the magnitude of the change that is set to sweep the world’s technology infrastructure.
A zettabyte (Zbyte) is a trillion gigabytes, or 10²¹ bytes. In 2006, the total capacity of all hard drives was roughly a sixth of a zettabyte. Ten years on, the Internet is carrying more than three zettabytes of data center traffic a year, and network giant Cisco predicts that by 2019, global data center traffic will be more than 10Zbyte per month.
Things are hooking up
There’s more. Thanks to the so-called Internet of Things (IoT), it’s not just people creating the deluge. More data is being generated by machines and objects – from cash registers to traffic lights – and Cisco’s Cloud Index this year predicted that by 2019, the total amount of data produced by people and things will top 500Zbytes per year.
Where is it all going? Not all of it is kept, but in 2014, the world had 1.4Zbytes of stored data, and the total was growing so fast that hard drive king Seagate predicted a capacity gap in 2016, because the world’s storage factories would not be able to produce enough drives. Thankfully, that doesn’t seem to be the case – but all other predictions of exponential growth are coming true in spades, and Cisco reckons we’ll have 3.5Zbytes stored by 2019.
Within this growth there are shifts. Data and workloads are rapidly moving off mobile devices and local hard drives, and out of private enterprise data centers into the public cloud. They can be consolidated more efficiently there, and the gap is widening. By 2019, in-house data centers will improve, from running an average of two workloads per server to more than three. Big deal. In the same period, public cloud servers will increase their capacity from five to eight or 10 workloads. At the same time, a new surge or “edge” capacity will be needed to manage IoT data and streaming media for the people and things in a locality. So it’s all changing. Let’s look at some specific strands of how that will work out.
Webscale data centers are built differently, and they are finally getting the racks they deserve. Giants such as Google have specified their own racks for some time, but now those ideas are emerging on the open market through initiatives such as the Facebook-led Open Compute Project (OCP), and the more proprietary offering of Intel’s Rack Scale Design.
There are common features to all of this. Racks, or trays within racks, are becoming the new unit of computing, and they are sharing power and cooling more effectively with data center distribution through the rack. Multiple Gbps of network are available, and more of that is delivered via fiber over open-switch architectures provided by SDN and NFV within the facility. Everything is accessible from the front of the rack, so the back can be arranged as a hot-aisle at temperatures uncomfortable to humans.
Data and workloads are rapidly moving off mobile devices and local hard drives, and out of private enterprise data centers into the public cloud. They can be consolidated more efficiently there, and the gap is widening.
OCP’s Open Rack takes it to a radical extent, increasing the rack equipment’s width to 21 inches. That’s more than some large enterprises can stomach, so LinkedIn has produced and shared an alternative – Open19 – which can accommodate existing equipment.
The processors behind the zettastructure are changing. Intel has ruled the roost and server architectures have remained stable for some time, but there are signs of change.
Energy is a huge data center cost, so ARM processors, widely in use in mobile devices, have long been proposed as an alternative, although ARM fans have been frustrated at the sluggish progress of their chips in data centers, as vendors such as Calxeda and AMD tried and failed to make an impact there. The latest effort, by Fujitsu, takes ARM processors and tunes them with scalable vector extension (SVE) for the high performance computing (HPC) market.
Meanwhile, IBM has encouraged other vendors to adopt its OpenPower processor, again for specialist markets including HPC. OpenPower is being built into servers compliant with the OCP rack standards, and implemented by cloud providers.
Much of the real action is in GPUs, the graphics processing units originally created for accleration cards used in gaming. As it turns out, the fast parallel processing required for speedy graphics is useful in more general activity. The world’s fastest supercomputers are now all built using GPUs. Nvidia is ahead of Intel, and has been striking deals with other vendors, including one that teams its GPUs with IBM’s OpenPower chips.
Networking has been shaken to its foundations by the arrival of software-defined networking (SDN), which allows topology, and functions can be centralized and expressed as software. This allows network functions virtualization (NFV), breaking the stranglehold of large network vendors and turning networks into the flexible resources that the zettastructure demands.
This is having a radical effect on the network within data centers, with functions aggregating into “white label” switches at the top of the rack, which can be defined on an open basis by groups such as OCP. This in turn is speeding the arrival of faster fiber links into the data centers and easing their integration with the world at large, where bigger fiber networks are being laid between countries.
Networks have become a key differentiator for data center providers that entice new customers by promising fast access to their existing customers through direct links within and between their premises. It’s no exaggeration to say that without SDN and NFV, the zettastructure could simply not exist.
For a couple of years we’ve been hearing about the all-flash data center, which will switch its most-used storage to solid state, for swift access, and to save power. But hard drives keep evolving to remain one step ahead in price-performance, while behind the scenes future generations of storage are evolving, which will change the rules again.
Solid-state drives (SSDs) are now available in sizes up to 60TB, and a fifth of businesses expect to soon be using flash as their primary storage medium. Future developments include phase-change memory, coming from IBM, Intel and Micron, and ReRAM from vendors including SanDisk.
But beware of believing the hype: 3D Xpoint, promised by Intel, looks to have been shipped before it was ready, leading to an embarrassing climb-down.
If all your data is in the cloud, that’s where the threat comes from. Data centers can now be attacked from anywhere, at any time. The toolkit of attackers includes distributed denial-of-service (DDoS) threats, where a data center can be swamped with apparently real traffic. There’s also a continual effort to find weaknesses in the software that would allow access to the infrastructure.
There’s an arms race between those looking for weaknesses to exploit and those patching them. At the same time, while specialist companies find ways to filter traffic and protect data centers from DDoS attacks, others are finding new ways to hit them. Some of these attacks enter the political arena. China, Russia and the US government are all known to have backed attacks on bodies within other nations. It’s also known that the US hoarded unpublished weaknesses in software to use as cyber-weapons.
Meanwhile, there’s also a new generation of physical attacks to data centers, based on their non-IT components. All too often, physical infrastructure has unconsidered interfaces to the Internet, which can be found and exploited, potentially shutting down a data center. Another devious trick might be to deliberately book new cloud instances in such a way as to overwhelm the power and cooling within a site.
These attacks have been overlooked in the past, but as the zettastructure begins to handle more real-world tasks in the Internet of Things, these attacks could actually threaten lives by sabotaging critical functions at, for example, hospitals or in transport systems.
This article appeared in the October issue of DatacenterDynamics magazine