Rebooting the data center
How can data centers be made more efficient, scalable, flexible and secure? Specifically, how can IT service providers achieve lower construction costs, a PUE nearer 1.0, more computing power per watt, lower latency for users, a smaller carbon footprint — and, above all, how can they manage these complex systems better? Here are some of the trends that will underpin the 21st century data center (or Data Center 2.0).
For certain kinds of high-volume, low-compute-power workload — such as web page serving, search engine query execution, or parallelised data processing tasks — a new species of server, the , may occupy less data center space and consume less power than the traditional Xeon- or Opteron-based enterprise server.
HP is among the leading vendors in the emerging microserver space with its Project Moonshot. The first Moonshot server, announced in November 2011, was an ARM-based SoC platform called Redstone, which HP installed as a demonstration system in its Discovery Lab. Based on an existing ProLiant SL6500 chassis with server trays that can hold a mixture of compute or storage cartridges, the Redstone Development Server Platform delivers four times the density of the traditional ProLiant server (accommodating up to 72 compute nodes per tray, or 288 per 4U chassis) while consuming a tenth of the power, according to HP.
In June 2012 HP changed course, announcing that the first production Moonshot microservers, codenamed Gemini, would use compute cartridges based on Intel's 'Centerton' Atom processors — specifically, the Atom S1200 series with TDP (Thermal Design Power) ratings of 6.1-8.5W. Originally scheduled to ship late last year, Gemini servers are now expected in Q2 2013. ARM is by no means out of the Moonshot picture though: the latest recruit to HP's Pathfinder Partner Program, which helps vendors develop for the Moonshot platform, is Texas Instruments with its ARM-based (Cortex-A15 MPCore) Keystone II SoC. Whether Keystone II-based cartridges will appear in a future generation of Gemini servers is not yet clear, however.
Other vendors are involved of course — notably Dell with its ARM-based Copper servers. And last year AMD bought microserver vendor SeaMicro to underpin its Data Center Server Solutions business. SeaMicro's Atom-supporting SM15000 server recently achieved certification for Cloudera's CDH4, a widely deployed Apache Hadoop distribution used for Big Data analytics.
Although microservers are interesting development, they are not applicable to all kinds of workloads, and are unlikely to revolutionise the data center anytime soon. Analyst IHS iSuppli, for example, expects shipments to hit 1.2 million units by the end of 2016 — just 10 percent of the total server market.
Fast, low-power, physically resilient solid-state storage is familiar in client devices — the flash memory in our smartphones, and the SSDs in our tablets and ultrabooks for example. SSD arrays are also increasingly making their way into the data center, where their performance and power efficiency are particularly appealing. However, as with microservers, solid-state storage is not universally applicable in this environment.
The main inhibitor for large-scale SSD uptake is the price, which although dropping remains significantly higher than traditional hard disk-based solutions. Another potential problem with SSDs is limited 'write endurance' — failures in blocks of non-volatile NAND flash storage after the maximum number of P/E (Program/Erase) cycles has been exceeded. This has been addressed by Intel with its High Endurance Technology (HET), which combines hardware and firmware enhancements to improve enterprise-class SSD endurance. HET includes 'background data refresh', which moves data around during inactive periods to avoid areas of heavy read activity.
In Intel's tests (see above), first in a controlled environment and then in a production data center (running a security compliance database), SSD arrays delivered the following benefits over comparable enterprise-class (15K rpm) hard-disk-based storage: up to 5x performance increase on random disk I/O tasks; reduced read/write latency (by up to 10x/7x respectively) and reduced maximum latency (by up to 8x); faster transition from idle to active state and no increase in seek time as drive capacity is approached; and 50 per cent less power consumption plus a third less heat production. Intel also suggested that the higher initial (3x) cost of an SSD solution could be justified thanks to: reduced IT time spent handling unwieldy I/O queue depths; no backlogs in recording monitoring data (eliminating a potential compliance issue); fewer delays in patching systems; performance and capacity to handle workloads for 3-5 years; and a simpler setup than a traditional SAN or NAS solution.
Enterprise-class SSD storage is not only available in SATA-based arrays such as EMC's new XtremeIO product, but also as PCI cards from a number of vendors, including Fusion-io and LSI. An alternative approach is taken by X-IO, which specialises in hybrid storage systems that combine SSDs and hard drives in a single pool of storage, with firmware (Continuous Adaptive Data Placement) that places data on SSD only if measurable performance gains will be realised. According to Mike Wills, CEO of RTW Hosting, replacing a traditional SAN with a hybrid array such as X-IO's can deliver 15-20 times the performance, along with a 60 percent improvement in power efficiency. Where an optimised balance of read/write performance and storage capacity is required, the tiered hybrid approach may be the way to go.
The software-defined data center
Server and storage virtualisation are mature concepts, but for maximum data center efficiency and flexibility, the third major IT component — networking — arguably needs to undergo a similar process of decoupling the control layer from the physical hardware. (SDN) is a new field, and the major networking players are still working out their responses to it — in July 2012 was a key move in this respect.
The standard-setting and promotion body for SDN is the Open Networking Foundation (ONF) and its major vehicle is the OpenFlow protocol. In April 2012, Google — a founding ONF member — disclosed details of a large-scale OpenFlow implementation on its datacenter backbone network, which carries the search giant's internal traffic (Google's other, internet-facing, backbone network carries user traffic).
Google's SDN experience included faster deployment time, the ability to emulate the backbone network in software for testing, easy upgradability, significant improvements in network utilisation (close to 100 percent, compared to an industry average of 30-40%) and a generally high degree of stability. Echoing its strategy with servers, Google built its own network switches for this project — something that, if widely adopted, will make vendors of traditional data center networking equipment (notably Cisco) nervous.
Modular data centers
Traditional custom-built 'bricks and mortar' data centers are costly and time-consuming to build (and expand when capacity becomes limited). Prefabricated modular data centers — often fitted into standard shipping containers containing servers, storage, networking and cooling equipment — allow companies to build the capacity they currently require quickly, with the ability to deploy more modules rapidly and cost-effectively as demand grows.
The Uptime Institute survey referred to earlier noted plenty of interest among its respondents in prefabricated modular data centers. However, this is a trend that's clearly in its early stages: 52 percent of respondents said they were still considering traditional builds for future data center expansion, with 19 percent looking at prefabricated components with little or no traditional construction; meanwhile, 21 percent said that the fully containerised data center was an interesting option.
Only 11 percent of respondents professed themselves 'extremely confident' about the maturity of the modular data center market in 2012, although 71 percent agreed that 'keeping pace with changing demands of successive generations of IT equipment' was more important than 'long useful life of data center infrastructure for a constant set of IT equipment demands' (26%).