Biz & IT —

Intel wants to kill the traditional server rack with 100Gbps links

New rack design disaggregates and shares CPU, storage, and network components.

Intel wants to kill the traditional server rack with 100Gbps links

Intel is working to replace the traditional server rack with a more efficient architecture that separates CPU, storage, power, and networking resources into individual components that can be swapped out as needed.

Power and cooling would be shared across CPUs, rather than having separate power supplies for each server. Server, memory, network, and storage resources would all be disaggregated and shared across the rack. Incredibly fast interconnects will be needed to prevent slowdowns because disaggregating components pushes them further apart, and Intel is thus building an interconnect that's capable of 100Gbps.

"We are developing a rack-scale architecture," Lisa Graff, VP and general manager of Intel's data center marketing group, said in a briefing with reporters last week. "We're working with end users, OEMs, and ISVs to drive common standards in a reference architecture."

The first version of this reference architecture is expected to be published sometime in 2014. Graff said the idea is to let data center managers "mix and match components instead of forklifting a rack" when pieces need to be replaced. Sharing things like memory and storage across CPUs will allow higher utilization of computing resources, and a design that eliminates unnecessary parts will let data centers cram more computing power into each rack.

Intel

The effort is complementary to Facebook's Open Compute Project. Facebook is already designing its own servers, stripping out extraneous bits of hardware, and it has worked with Intel on possible designs for racks that disaggregate and share resources.

The networking technology used by typical data centers isn't quite fast enough to power disaggregated racks just yet. That's why Intel is developing silicon photonics technology that uses light to move data at up to 100Gbps. Silicon photonics has the added benefit of reducing the amount of cabling needed in a rack.

"Silicon photonics made with inexpensive silicon rather than expensive and exotic optical materials provides a distinct cost advantage over older optical technologies in addition to providing greater speed, reliability, and scalability benefits," Intel said in January, when it announced that it has produced engineering samples of the technology. "Businesses with server farms or massive data centers could eliminate performance bottlenecks and ensure long-term upgradability while saving significant operational costs in space and energy."

Intel isn't the only company trying to speed up networking with silicon photonics. As we reported in December, IBM has "developed a technology that integrates optical communications and electronics in silicon, allowing optical interconnects to be integrated directly with integrated circuits in a chip." IBM said its silicon nanophotonics technology is ready for mass production, although specific implementations haven't been announced.

Corporate data centers won't be receiving an immediate overhaul, but Intel's vision (and IBM's) is one that may be appealing to companies with thousands or tens of thousands of servers.

Intel describes its goals as follows:

  • Physical Aggregation. All non-critical sheet metal removed and key components such as power supplies and fans taken out of individual servers and consolidated at the rack level. Savings are expected due to higher levels of efficiency and lower costs by reducing the number of fans and power supplies.
  • Fabric Integration and Storage Virtualization. Disaggregate and separate out the storage from compute systems with direct attached storage, and achieve higher utilization through storage virtualization. The compute and network fabric is the key technology that is enabling disaggregation of storage without impact to performance. Intel Silicon Photonics interconnects will enable higher speed connections between various computing resources within the rack, thus enabling the eventual disaggregation of server, memory, network and storage within the rack.
  • Future. Ultimately, the industry will move to subsystem disaggregation where processing, memory and I/O will be completely separated into modular subsystems, making it possible to easily upgrade these subsystems rather than doing a complete system upgrade.

Early versions of the rack-scale architecture are planned for deployment by several companies in China, namely Alibaba, Baidu, Tencent, and China Telecom.

New “smartphone-class” chips for the data center

Intel's rack-scale architecture was part of a larger announcement to be made Wednesday at the Intel Developer Forum in Beijing. Intel is introducing new Atom SoCs (system-on-a-chip) for the data center. As we reported last December, Intel is pitching "smartphone-class CPUs" to data centers by combining "extremely low power usage with server-class features, including virtualization technology and Error-Correcting Code (ECC) for higher reliability."

The December 2012 announcement unveiled Intel Atom S1200 processors, the "world's first 64-bit SoC for servers." What's new today is the S12x9 Atom processor for storage systems.

Key features include 40 lanes of integrated PCIe 2.0, the physical paths between I/O and processor; RAID storage acceleration built into the hardware; failover support; native dual-casting that "can allow data to be read from a source and delivered to two memory locations simultaneously"; and Asynchronous DRAM Self-Refresh, which protects DRAM data during power failures.

MacroSAN, Accusys, Qsan, and Qnap are releasing systems based on the new storage chip, Intel said.

Intel also offered brief previews of chips that don't have firm release dates yet. Avoton, the second generation of Intel's 64-bit Atom chips for microservers, is based on Intel's 22nm process.

"Avoton will feature an integrated Ethernet controller and is expected to deliver significant improvements in performance-per-watt," Intel said. "Avoton is now being sampled to customers and the first systems are expected to be available in the second half of 2013."

The Atom S1200 and Avoton are being used to power HP's "Moonshot" servers.

Another Atom-based SoC built on 22nm process technology, codenamed Rangely, is targeted at network and communications infrastructure such as "entry-level to mid-range routers, switches and security appliances." Like Avoton, Rangely is expected to come out in the second half of this year.

Intel will also be refreshing its Xeon E3, E5, and E7 processor lines later this year. The Haswell-based E3 will improve performance for video analytics with integrated graphics, and lower power consumption to as little as 13 watts. Xeon E5 will improve security with several new features including "Intel OS Guard, the next generation of Intel Execute Disable Bit, [which] protects against privilege attacks by preventing malicious code from executing out of application memory space, in addition to data memory."

Xeon E7, targeted at analyzing massive data sets, triples memory capacity to 12TB in an eight-socket node. The new Xeon E5 will debut in the third quarter of 2013, while Xeon E7 is scheduled for the fourth quarter. The refreshed E3 is coming this year, but Intel did not get more specific than that.

Channel Ars Technica