Microsoft greatly increases the compute capabilities of its data centers
Share on Twitter.
Get the most reliable SMTP service for your business. You wished you got it sooner!
June 16, 2014
Microsoft said earlier this morning that it has found a new method to greatly increase the
computing capabilities of its data centers, despite the fact that Moore's Law is wheezing towards
its inevitable demise sooner as opposed to later.
In a paper to be presented this week at the International Symposium on Computer Architecture (ISCA) titled
'A Reconfigurable Fabric for Accelerating Large-Scale Datacenter Services' a group of top Microsoft
Research engineers explain how the company has dealt with the slowdown in single-core clock-rate
improvements that has occurred over the past eleven years.
To get around that issue, Microsoft has designed a system it calls Catapult, which automatically
offloads some of the advanced technology that powers its Bing search engine onto clusters of highly
efficient, low-power FPGA chips attached to typical Intel Xeon server processors.
Think of FPGAs (field-programmable gate arrays) as chips whose circuits can be customized and
tweaked as required, allowing specific tasks to be offloaded from the Xeon processors and accelerated
in FPGA hardware.
This approach may save Microsoft from a rarely acknowledged problem that lurks in the technology
industry-- that processors are not getting much faster these days. And that would be a reality.
For the past fifty years or so, almost every aspect of our global economy has been affected
by Moore's Law, which states that the number of transistors on a chip of the same size will
double every 1 1/2 year or so, resulting in faster performance and much improved power efficiency.
But there's one small problem with that-- Moore's Law is not a law. Instead, it was an assertion
by Intel founder Gordon Moore in a 1965 article that the semiconductor industry got rather
carried away with.
In the past ten years, the salubrious effects of Moore's Law have started to wane, because
although companies are packing more and more transistors on their chips, the performance gains that
those transistors bring with them are not as great as they were during the law's early days.
And as you probably know by now, Intel has based its entire business model to the successful
fulfillment of Moore's Law, and proudly announces each new boost in transistor counts.
Those new transistors can help to increase a compute core's all-important instruction set per
cycle (IPC) metric-– IE: improved branch prediction, larger caches, more-efficient scheduling,
beefier buffers, etc.
However, the simple fact is that although chips have gone multi-core and are getting better
at multi-tasking, those individual cores are not getting much faster due to any significant new
discovery. And that's where the core of the problem resides in.
As AMD CTO Joe Macri recently told us, "There's not a whole lot of revolution left in CPUs per se, you know."
He did, however, note that "there's a lot of evolution left." Microsoft's new Catapult system is a
bit of both.
Under new chief executive Satya Nadella, Microsoft is investing billions of dollars at massive
data centers in its attempt to become a cloud-first company.
Understandably, part of that effort is to determine a way to jump-start consistent data-center compute-performance
The solution that Microsoft Research has come up with is to pair field-programmable gate
arrays with typical x86 processors, then let some data-center services such as the Bing search
engine offload certain well-understood operations to the arrays.
To say that the performance improvements in this approach have been noticeable would be a
Microsoft tells us that a test deployment on 1,632 servers was able to increase query
throughput by 95 percent, while only increasing power consumption by about 10 percent.
Though FPGA technology is well understood and used widely in the embedded technology industry,
it's not often that we hear it being paired with standard off-the-shelf CPUs for accelerating
web-facing software – at least until now, that is.
"We're moving into an era of programmable hardware supporting programmable software," Microsoft
Research's Doug Burger said. "We're just starting down that road now."
If Microsoft has indeed determined how to almost double the performance of its computers
while only paying a tenth more in electricity for large-scale data center tasks – and we see
no reason to doubt them – that's not only a huge saving, but also one that saves the company
from the slowdown in run-of-the-mill CPUs chips.
"Based on the actual results, Bing will roll out FPGA-enhanced servers in one data center
to process customer searches starting in early 2015," said Derek Chiouh, the principal architect
"We were looking to make a big jump forward in data center capabilities. It's an important
area," Microsoft Research's Doug Burger said.
Of course, Microsoft isn't doing this just on a hunch either. Burger wrote a paper in 2011
called Dark Silicon and the end of Multicore Scaling, which predicted that "left to the multicore
path, we may hit a 'transistor utility economics' wall in as few as three to five years, at which
point Moore's Law may end, creating massive disruptions in our industry."
In other IT news
Microsoft says it wants to start the development of quantum computers, in a partnership with the
Massachusetts Institute of Technology.
Head of research at Microsoft, Peter Lee has told the MIT Technology Review digital summit
that the company will be supporting research as well as doing its own work at its Station Q
research lab at the University of California's Santa Barbara campus.
In fact, what the software behemoth is after is something that can be exploited as a “fab-friendly”
To be sure, creating qubits has become almost routine in quantum research labs. In 2012, the University
of New South Wales created a single-atom qubit on silicon.
But creating qubits and getting them to behave properly, as a superposition of states rather than
a quantum 1 or 0, is difficult. Doing so reliably is even more difficult and creating qubits in a
manner that could be turned into some kind of microelectronics foundry is a long way off, no matter
how hard you try.
Lee knows this, and he told the MIT Technology Review that as far as Microsoft is aware, the
current approaches to creating qubits don't scale very well.
Microsoft's direction is to work on a “topological qubit”. The description “topological” refers to
how entanglement is created and maintained and the point of the work is that Redmond told MIT
that it sees the approach as more robust than other research into qubits.
Only a cynic would also note that whoever successfully builds a reliable, mass-producable
qubit will have an intellectual property of almost incalculable value, making it unlikely that
Microsoft would follow paths where others have taken the lead.
In other IT and research news
Cloud economics sometimes make a lot of sense, and now the phenomenon is pushing prices lower
for IBM's SoftLayer cloud solution close to those of its rivals Amazon and Microsoft.
IBM is announcing today that it has dropped SoftLayer object storage prices to $0.04 per
gigabyte per month, lightly cutting compute prices, and adding a low-cost networking service
that lets companies rent a secure, high-bandwidth connection into the cloud.
By reducing its storage prices, SoftLayer has brought its costs into parity with prices far
below the $0.085 per gigabyte offered by rivals Amazon ($0.03 per GB per month) and Google ($0.026
per GB per month) and Microsoft ($0.05 per GB per month).
It also added a connection option named Direct Link, which allows SoftLayer customers to create
private, dedicated connections into the cloud from 18 IBM network points of presence around the world.
This simply means that a customer with a rack in a Softlayer-linked data center from Equinix,
Coresite, Terremark, Pacnet, Interxion, TelecityGroup or Telx, can now rent a dedicated connection
between their computer equipment and IBM's cloud.
The scheme is similar to Microsoft's Express Route and Amazon's Direct Connect schemes, with
the main variation being a greater diversity of suppliers for IBM's scheme, plus of course, lower pricing.
Direct Link connections are available now and cost $147 per month for a 1 Gbps connection or $997
per month for a 10 Gbps one.
This also undercuts Amazon and Microsoft, which charge respectively $219 and $600 a month for equivalent
1 Gbps services.
"Direct Link itself is a similar product offering to others in the marketplace, but the use
case is different because it gives access to SoftLayer's cloud platform and network," said IBM's
SoftLayer manager of product innovation, Marc Jones.
"SoftLayer has had this service available to select customers for several months already, and
IBM's large enterprise install base was requesting this service, so this is the official 'productization'
of the service."
Although SoftLayer has some advantages over its larger rivals – such as bare metal servers that
guarantee higher performance at the cost of more ornate manageability requirements – IBM is yet to
build out the set of services on the cloud to gain parity with rivals.
Get the most dependable SMTP server for your company. You will congratulate yourself!
Share on Twitter.