September 06, 2013, 3:13 PM —
Image credit: Huangjia Hui
Intel is acknowledging the changing face and internal functions of the data center with a new initiative designed to re-architect the underlying infrastructure, allowing companies and end users to adapt their data centers to a more services- and mobility-oriented environment.
The strategy, as laid out at its Datacenter Day event held in late July and hosted by several Intel executives, is for automation and speed to replace manual, time-consuming and often fixed functions, each with their own independent configuration, said Diane Bryant, senior vice president and general manager of the Datacenter and Connected Systems Group at Intel.
"Today, the network is still manually configured. The process to reconfigure a new network to support the service is a manual process that can take weeks," said Bryant. With the virtualized, software-defined network, the time to provision software and hardware on a new service can be reduced to just minutes.
She also noted that data storage continues to grow at a 40% compound annual growth rate, with 90% of it unstructured. She cited an IBM study that found businesses store on average 18 copies of the same piece of data. "That actually sounds quite logical having run IT for four years," joked Bryant, who was previously CIO of Intel.
Finally, she noted that even with virtualization, server capacity is barely at 50%. "That means 50% of server capacity is unused, which is a true crime," said Bryant.
Christian Perry, senior analyst for data centers, Technology Business Research
To solve this, Intel is looking to rearchitect not just servers but the network for cloud services, and plans to do this via software-defined networks. "Software-defined networks allow us to extract the control function out of the switch, run it globally, run it on standard high volume Intel hardware as just another app running on your Intel architecture. That drives up the utilization and drives down the capital expense through the movement off proprietary servers," said Bryant.
Intel wants to help companies move beyond the standard generation of data centers by offering what it calls the Rack Scale Architecture (RSA), which will virtualize the whole network and every component in it. An application will assemble the CPU, memory, storage and networking it needs from the pool of assembled hardware and build its own virtual server, storage and network.
Why is Intel doing this?
Intel is a chip company yet it's taking charge of a component of downstream technology, the server architecture. For Intel to get into server, network and storage architecture would like Qualcomm, the dominant player in mobile phone chips, deciding the cellular networks are not being well managed by Verizon, AT&T, etc., and it was going to do something about it.
The problem is the big server vendors aren't taking charge. They are distracted to some degree or another. CEO Meg Whitman is slowly righting the ship of HP but that company took a severe body blow in recent years. Dell is in even worse shape with its lingering privatization plans, and IBM isn't interested in the x86 business because it tried to sell the System x business to Lenovo but the two couldn't reach a price agreement.
Nature and business abhors a vacuum and Intel is stepping in. And as Nathan Brookwood, research fellow with Insight64 noted, it's happened before with great success.
"Intel is acting as a leader and there is nothing wrong with Intel picking up the leadership mantle and moving forward. They did it in the past. Who drove WiFi into mass market? It was Intel with Centrino. Before that, WiFi was a curiosity. It took Intel putting Centrino with WiFi in every laptop to make it popular. They also did it with USB and PCI Express," Brookwood said.
Christian Perry, senior analyst for data centers at Technology Business Research, agreed. "They see an opportunity, it's theirs for the taking to define that leadership role. They've probably had that opportunity in the past few years, but why now? We see traction around software-defined data centers and Intel has plenty of software capabilities in terms of programming into the chips. They can help to define software definitions and that underlying layer in the data center," Perry said.
Not only is Intel taking a role that should fall to hardware vendors, it is also taking on the job of defining a software-defined data center, something you'd expect from VMware, Citrix or Microsoft, noted Perry.
"They have a very well-orchestrated definition behind software-defined anything or everything. I never heard a more sensible approach in terms of explaining how everything can work in a software-defined data center. We hear things from EMC and VMware, but Intel really sees the big picture and they should see the big picture because their chips are running the big picture," Perry said.
Quietly building the infrastructure
So it remains to be seen if Intel can do for software-defined networks what it did for WiFi, but it sure will try. Without a lot of hoopla, Intel has made some major moves into networking infrastructure. Intel has introduced Open Network Platform reference designs to help OEMs build and deploy a new generation of networks that it says will maximize bandwidth, reduce costs, and offer flexibility to support new services.
In April 2013, Intel introduced three platforms for software defined networking and network function virtualization: the Open Network Platform Switch Reference Design (ONPS); the Data Plane Development Kit (DPDK) Accelerated Open vSwitch; and the Intel Open Network Platform Server Reference Design.
ONPS allows for automated network management, and coordination between the server switching elements and network switches. The DPDK will improve small packet throughput and workload performance, while the Open Network Platform Server Reference Design, previously codenamed "Sunrise Trail," is based on Intel chips and Wind River software. Intel acquired Wind River in 2009.
These are not new efforts. Intel has been working on them for years and claims customers for this technology include including HP, NEC, NTT Data, Quanta, Super Micro, VMware and Vyatta (a Brocade company).
New edge servers (a recap)
In a departure from traditional Intel operating procedure, the company plans to offer custom chips to big customers. Already Facebook and eBay will get custom low-end Xeon E3 processors in a system-on-a-chip (SoC) design, and there will be more. Expect more of this, said Jason Waxman, general manager of the Cloud Computing Platforms Group at Intel.
At the event, Waxman introduced the Atom C2000, an eight-core processor, known by its codename "Avoton." It will come with an encryption acceleration network device codenamed "Rangeley" as well as with many Xeon features, such as error code correction (ECC), Intel Virtualization Technology and 64GB of memory capacity.
In short, it looks a lot like a Xeon, and that's no accident. "People want consistency. They want 64 bits and software compatibility and error correction code even on the low end," said Waxman.
He also announced a future chip based on the Xeon E3 design but will use a system-on-a-chip (SoC) design instead of the usual discrete design for Xeon servers, which involves several chips on the server board. A Xeon SoC means much lower power consumption and smaller motherboards, since fewer chips are used. The Xeon SoC will be introduced next year with the Broadwell generation of processors. Broadwell will be the current Haswell architecture, built on a 22nm design process, shrunk to 14nm.
That the first customers for this chip are Facebook and eBay is no coincidence. Consumer networks like those use thousands of edge servers to handle their millions of visitors at any given moment and to serve up HTML pages generated on the back end. It has slowly dawned on server vendors that using a powerful processor such as Xeon E5, or even an E3, for a server that simply handles Web connections and spits out HTML pages is overkill. The Atom (and ARM) is more than adequate for that task.
That was the appeal of using Atom and ARM processors in ultra-dense edge servers such as HP's new Project Moonshot and the 10u server from startup SeaMicro, which AMD acquired in 2012. Both make the same point: save your Xeons for the database work and let a chip that uses a fraction of the power to pump out HTML pages.
Perry says that edge servers are a new market for Intel and the industry at large. Not a huge one but a new one nonetheless.
"Microservers won't be huge for revenue but Intel will play in that market. It will be a while before that market grows out. Customers are still in a wait and see mode on microservers," said Perry.
Customers may be waiting but the vendors are not. AMD made its move with the acquisition of SeaMicro, which makes ultra-dense servers using Intel's Atom processors. AMD will eventually put its own chips in those servers. HP has Project Moonshot servers using an ARM-based processor designed by Calxeda.
So Intel doesn't want to be left out. "Intel wants to be a part of that market but it's more of a 'We'll provide anything you need for your workload environments' strategy," said Perry.