This week, the New York Times ran a series of articles about data center power use (and abuse) “Power, Pollution and the Internet” (http://nyti.ms/Ojd9BV) and “Data Barns in a Farm Town, Gobbling Power and Flexing Muscle” (http://nyti.ms/RQDb0a). Among the claims made in the articles were that data centers were “only using 6 to 12 % of the energy powering their servers to deliver useful computation. Like a lot of media broadsides, the reality is more complex than the dramatic claims made in these articles. Technically they are correct in claiming that of the electricity going to a server, only a very small fraction is used to perform useful work, but this dramatic claim is not a fair representation of the overall efficiency picture. The Times analysis fails to take into consideration that not all of the power in the data center goes to servers, so the claim of 6% efficiency of the servers is not representative of the real operational efficiency of the complete data center.
On the other hand, while I think the Times chooses drama over even-keeled reporting, the actual picture for even a well-run data center is not as good as its proponents would claim. Consider:
A new data center with a PUE of 1.2 (very efficient), with 83% of the power going to IT workloads.
Then assume that 60% of the remaining power goes to servers (storage and network get the rest), for a net of almost 50% of the power going into servers. If the servers are running at an average utilization of 10%, then only 10% of 50%, or 5% of the power is actually going to real IT processing. Of course, the real "IT number" is the server + plus storage + network, so depending on how you account for them, the IT usage could be as high as 38% (.83*.4 + .05).
In the last couple of weeks, I finally put a couple of pieces together . . . the tech industry is pushing hard, down two parallel tracks, toward much more resource-efficient computing architectures.
Track 1: Integrated systems. Computer suppliers are putting hardware components (including compute, network, and storage) together with middleware and application software in pre-integrated packages. The manufacturers will do assembly and testing of these systems in their factories, rather than on the customer's site. And they will tailor the system — to a greater or lesser degree, depending on the system — to the characteristics of the workload(s) it will be running.
The idea is to use general-purpose components (microprocessors, memory, network buses, and the like) to create special-purpose systems on a mass-customization basis. This trend has been evident for a while in the Oracle Exadata and Cisco UCS systems; IBM's Pure systems introductions push it even further into pre-configured applications and systems management.
Track 2. Modular data centers. Now, zoom out from individual computing systems to aggregations of those systems into data centers. And again, assemble as much of the componentry as possible in the factory rather than on-site. Vendors like Schneider, Emerson, and the systems shops like IBM and HP are creating a design approach and infrastructure systems that will allow data centers to be designed in modular fashion, with much of the equipment like air handling and power trucked to the customer's site, set up in the parking lot, and quickly turned on.
Next up in the 2012 lineup for the Intel E5 refresh cycle of its infrastructure offerings is Cisco, with its announcement last week of what it refers to as its third generation of fabric computing. Cisco announced a combination of tangible improvements to both the servers and the accompanying fabric components, as well as some commitments for additional hardware and a major enhancement of its UCS Manager software immediately and later in 2012. Highlights include:
New servers – No surprise here, Cisco is upgrading its servers to the new Intel CPU offerings, leading with its high-volume B200 blade server and two C-Series rack-mount servers, one a general-purpose platform and the other targeted at storage-intensive requirements. On paper, the basic components of these servers sound similar to competitors – new E5 COUs, faster I/O, and more memory. In addition to the servers announced for March availability, Cisco stated that it would be delivering additional models for ultra-dense computing and mission-critical enterprise workloads later in the year.
Fabric improvements – Because Cisco has a relatively unique architecture, it also focused on upgrades to the UCS fabric in three areas: server, enclosure, and top-level interconnect. The servers now have an optional improved virtual NIC card with support for up to 128 VLANs per adapter and two 20 GB ports per adapter. One in on the motherboard and another can be plugged in as a mezzanine card, giving up to 80 GB bandwidth to each server. The Fabric Interconnect, the component that connects each enclosure to the top-level Fabric Interconnect, has seen its bandwidth doubled to a maximum of 160 GB. The Fabric Interconnect, the top of the UCS management hierarchy and interface to the rest of the enterprise network, has been up graded to a maximum of 96 universal 10Gb ports (divided between downlinks to the blade enclosures and uplinks to the enterprise fabric.
Last week it was Dell’s turn to tout its new wares, as it pulled back the curtain on its 12th-eneration servers and associated infrastructure. I’m still digging through all the details, but at first glance it looks like Dell has been listening to a lot of the same customer input as HP, and as a result their messages (and very likely the value delivered) are in many ways similar. Among the highlights of Dell’s messaging are:
Faster provisioning with next-gen agentless intelligent controllers — Dell’s version is iDRAC7, and in conjunction with its LifeCyle Controller firmware, Dell makes many of the same claims as HP, including faster time to provision and maintain new servers, automatic firmware updates, and many fewer administrative steps, resulting in opex savings.
Intelligent storage tiering and aggressive use of flash memory, under the aegis of Dell’s “Fluid Storage” architecture, introduced last year.
A high-profile positioning for its Virtual Network architecture, building on its acquisition of Force10 Networks last year. With HP and now Dell aiming for more of the network budget in the data center, it’s not hard to understand why Cisco was so aggressive in pursuing its piece of the server opportunity — any pretense of civil coexistence in the world of enterprise networks is gone, and the only mutual interest holding the vendors together is their customers’ demand that they continue to play well together.
At its recent financial analyst day, AMD indicated that it intended to differentiate itself by creating products that were advantaged in niche markets, with specific mention, among other segments, of servers, and to generally shake up the trench warfare that has had it on the losing side of its lifelong battle with Intel (my interpretation, not AMD management’s words). Today, at least for the server side of the business AMD made a move that can potentially offer it visibility and differentiation by acquiring innovative server startup SeaMicro.
SeaMicro has attracted our attention since its appearance (blog post 1, blog post 2), with its innovative architecture that dramatically reduces power and improves density by sharing components like I/O adapters, disks, and even BIOS over a proprietary fabric. The irony here is that SeaMicro came to market with a tight alignment with Intel, who at one point even introduced a special dual-core packaging of its Atom CPU to allow SeaMicro to improve its density and power efficiency. Most recently SeaMicro and Intel announced a new model that featured Xeon CPUs to address the more mainstream segments that were not for SeaMicro’s original Atom-based offering.
On Monday, February 13, HP announced its next turn of the great wheel for servers with the announcement of its Gen8 family of servers. Interestingly, since the announcement was ahead of Intel’s official announcement of the supporting E5 server CPUs, HP had absolutely nothing to say about the CPUs or performance of these systems. But even if the CPU information had been available, it would have been a sideshow to the main thrust of the Gen8 launch — improving the overall TCO (particularly Opex) of servers by making them more automated, more manageable, and easier to remediate when there is a problem, along with enhancements to storage, data center infrastructure management (DCIM) capabilities, and a fundamental change in the way that services and support are delivered.
With a little more granularity, the major components of the Gen8 server technology announcement included:
Onboard Automation – A suite of capabilities and tools that provide improved agentless local intelligence to allow quicker and lower labor cost provisioning, including faster boot cycles, “one click” firmware updates of single or multiple systems, intelligent and greatly improved boot-time diagnostics, and run-time diagnostics. This is apparently implemented by more powerful onboard management controllers and pre-provisioning a lot of software on built-in flash memory, which is used by the onboard controller. HP claims that the combination of these tools can increase operator productivity by up to 65%. One of the eye-catching features is an iPhone app that will scan a code printed on the server and go back through the Insight Management Environment stack and trigger the appropriate script to provision the server.[i]Possibly a bit of a gimmick, but a cool-looking one.
Data centers, like any other aspect of real estate, follow the age-old adage of “location, location, location,” and if you want to build one that is really efficient in terms of energy consumption as well as possessing all the basics of reliability, you have to be really picky about ambient temperatures, power availability and, if your business is hosting for others rather than just needing one for yourself, potential expansion. If you want to achieve a seeming impossibility – a zero carbon footprint to satisfy increasingly draconian regulatory pressures – you need to be even pickier. In the end, what you need is:
Low ambient temperature to reduce your power requirements for cooling.
Someplace where you can get cheap “green” energy, and lots of it.
A location with adequate network connectivity, both in terms of latency as well as bandwidth, for global business.
A cooperative regulatory environment in a politically stable venue.
This week AMD finally released their AMD 6200 and 4200 series CPUs. These are the long-awaited server-oriented Interlagos and Valencia CPUs, based on their new “Bulldozer” core, offering up to 16 x86 cores in a single socket. The announcement was targeted at (drum roll, one guess per customer only) … “The Cloud.” AMD appears to be positioning its new architectures as the platform of choice for cloud-oriented workloads, focusing on highly threaded throughput oriented benchmarks that take full advantage of its high core count and unique floating point architecture, along with what look like excellent throughput per Watt metrics.
At the same time it is pushing the now seemingly mandatory “cloud” message, AMD is not ignoring the meat-and-potatoes enterprise workloads that have been the mainstay of server CPUs sales –virtualization, database, and HPC, where the combination of many cores, excellent memory bandwidth and large memory configurations should yield excellent results. In its competitive comparisons, AMD targets Intel’s 5640 CPU, which it claims represents Intel’s most widely used Xeon CPU, and shows very favorable comparisons in regards to performance, price and power consumption. Among the features that AMD cites as contributing to these results are:
Advanced power and thermal management, including the ability to power off inactive cores contributing to an idle power of less than 4.4W per core. Interlagos offers a unique capability called TDP, which allows I&O groups to set the total power threshold of the CPU in 1W increments to allow fine-grained tailoring of power in the server racks.
Turbo CORE, which allows boosting the clock speed of cores by up to 1 GHz for half the cores or 500 MHz for all the cores, depending on workload.
Emerging ARM server Calxeda has been hinting for some time that they had a significant partnership announcement in the works, and while we didn’t necessarily not believe them, we hear a lot of claims from startups telling us to “stay tuned” for something big. Sometimes they pan out, sometimes they simply go away. But this morning Calxeda surpassed our expectations by unveiling just one major systems partner – but it just happens to be Hewlett Packard, which dominates the WW market for x86 servers.
At its core (unintended but not bad pun), the HP Hyperscale business unit Project Moonshot and Calxeda’s server technology are about improving the efficiency of web and cloud workloads, and promises improvements in excess of 90% in power efficiency and similar improvements in physical density compared with current x86 solutions. As I noted in my first post on ARM servers and other documents, even if these estimates turn out to be exaggerated, there is still a generous window within which to do much, much, better than current technologies. And workloads (such as memcache, Hadoop, static web servers) will be selected for their fit to this new platform, so the workloads that run on these new platforms will potentially come close to the cases quoted by HP and Calxeda.
Well actually I meant mobs of flash, but I couldn’t resist the word play. Although, come to think of it, flash mobs might be the right way to describe the density of flash memory system vendors here at Oracle Open World. Walking around the exhibits it seems as if every other booth is occupied by someone selling flash memory systems to accelerate Oracle’s database, and all of them claiming to be: 1) faster than anything that Oracle, who already integrates flash into its systems, offers, and 2) faster and/or cheaper than the other flash vendor two booths down the aisle.
All joking aside, the proliferation of flash memory suppliers is pretty amazing, although a venue devoted to the world’s most popular database would be exactly where you might expect to find them. In one sense flash is nothing new – RAM disks, arrays of RAM configured to mimic a disk, have been around since the 1970s but were small and really expensive, and never got on a cost and volume curve to drive them into a mass-market product. Flash, benefitting not only from the inherent economies of semiconductor technology but also from the drivers of consumer volumes, has the transition to a cost that makes it a reasonable alternative for some use case, with database acceleration being probably the most compelling. This explains why the flash vendors are gathered here in San Francisco this week to tout their wares – this is the richest collection of potential customers they will ever see in one place.