A new generation of servers has appeared in data centers. These new servers,
fitted with ARM processors, promise extremely high processing scalability while
keeping power and cooling demands in check. In spite of the promise, however,
ARM processors are not quite ready for wide adoption in data centers. This tip
explains the current benefits and limitations of an ARM server and shows you
what to look for as this technology develops.
Emergence of the ARM processor
Traditional processors are general-purpose devices capable of handling hundreds
of different instructions. The problem is that general processors need hundreds
of millions of transistors to handle the countless logical conditions needed to
support so many instructions. This translates directly into expensive
manufacturing costs: The latest Intel Xeon processors routinely cost more than
$1,000 each. Even more substantially, each processor has significant power and
cooling demands, which drives up the ongoing operating costs of data center
servers.
By comparison, ARM processors, based on reduced instruction set computing
(RISC) architectures, overcome many of these obstacles. By reducing the number
of instructions, the processor is simpler, cheaper, uses far less power and
runs with little (if any) substantial heat. The reduction in transistor count
also improves the processor’s performance because there are fewer logical
stages needed to process instructions. ARM processors are not new – they have
been around for decades and can be found in smartphones, printers, digital
cameras and other commercial electronic devices.
ARM processors in servers
Today, ARM processors have slowly begun to appear in some cutting-edge servers.
The key factor driving the adoption of ARM processors is scalability. It’s important
to understand that data centers are changing quickly. Internet-based companies,
such as Google and Facebook, have discovered that it is generally more cost
effective to fill data centers with vast amounts of inexpensive commodity
hardware than to rely on high-performance, overpriced servers. The idea of many
general-purpose servers is sometimes called “hyperscaling,” and many of the
cloud service providers have also adopted the concept of hyperscaling their
data centers.
While hyperscaling may offer a low-cost alternative to large, high-end
servers, it also poses some problems. Perhaps the biggest issue that
organizations may face as a result of hyperscaling their data centers is that
of power consumption. The cost of the electricity required to keep a server
online is usually trivial. However, when you multiply that cost by the hundreds
of thousands of servers in a large data center, the cumulative server power
cost can become staggering. And all of those servers produce heat, so cooling
the data center can easily increase the cost of electricity by another 50
percent.
Organizations are discovering that ARM servers can be ideal for handling the
power and cooling challenges of hyperscaling. To put the power savings into
perspective, Hewlett-Packard Co. estimates that servers equipped with ARM
processors could potentially consume up to 90% less power than their Intel
counterparts.
While each ARM processor core typically provides less raw processing power
than a traditional Intel or AMD chip, the low power and cooling requirements
allow a level of server scalability that would be impossible with traditional
processors. For example, the Tilara TILE-Gx family of processors can provide
from 16 to 100 cores on a single chip.
In addition, a large array of processors can be assembled in a single server
chassis. For example, the SeaMicro SM10000-64-HD includes dozens of Intel Atom
chips, providing 768 cores in a single chassis. Compare this to a common 1U
rack server with two 8-core (or even 12-core) processors. Ideally, systems like
the SeaMicro SM10000-64-HD can replace 60 traditional servers with a quarter of
the power and weight and one-sixth of the space.
ARM processors not quite ready
But hold onto those purchase orders. ARM technology is improving quickly, but
the technology isn’t quite ready yet. Before you add an ARM server to your next
technology refresh plan, it’s important to understand some of the current
limitations.
First, there are currently competing ARM standards under development. ARM
processors are developed differently than mainstream processors. If you buy a
server with an Intel or AMD processor, then you can rest assured that the
server’s processor was manufactured by Intel or AMD. If you were to purchase an
ARM server or other device with an ARM processor, that processor was not
manufactured by ARM Holdings. Instead, the ARM architecture is licensed to
manufacturers who develop their own ARM-based CPUs.
There are several manufacturers that are developing ARM processors, and
these processors will vary in capability because each manufacturer puts its own
spin on the ARM standard that it has licensed.
Performance is another stumbling block that ARM processors must overcome.
Current ARM processors are 32-bit, which is a poor choice for busy data center
servers. However, ARM Holdings announced in October that it is developing a
64-bit core. While ARM processor licensees will invariably adopt the 64-bit
license, it’s impossible to know exactly how the licensees will develop the new
ARM architecture.
As one example, Applied Micro is developing an ARM-based chip known as the
X-Gene that will function as a “system on a chip.” This single chip will
feature multiple ARM cores (although the number of cores has not yet been
released), twin 10-gigabit Ethernet ports, SATA storage control and even
virtualization support. Although the X-Gene chip isn’t ready for purchase just
yet, Applied Micro expects to be producing X-Gene chips by late 2012.
And finally, it’s important to consider the impact of operating systems (OSes)
and workloads. ARM processors and ARM-based servers will need a suitable OS.
Consumer devices may use an OS like Android or Apple’s iOS, but these are not
suited for servers. Linux will support ARM, but Windows Server versions, such
as 2008 R2, will not support ARM processors until (at least) the release of
Windows 8. Also, current ARM processors tend to favor less-sophisticated
workloads, such as Web servers, so deploying high-performance database
applications on an ARM server may not be possible for some time.
The future of the ARM server is closer than you think
In spite of the challenges, highly scaled ARM servers are appearing in
projects like HP’s Project Moonshot, using 2,800 Calxeda Energycore processors
in a single rack. Future efforts, such as HP’s Redstone project, may look to
use other ARM-type processors like the Intel Atom. When development of the
hardware and software finally utilize the sheer number of ARM processors
available in a single chassis, data center operators may see an extremely
competitive computing platform for at least some workloads.
