Microservers were expected to carve out a corner of the datacentre and in doing so let low power ARM-based processors into the server room.
But are forecasts that one in 10 of the servers shipped in 2016 will be a microserver on track to be realised?
No, says Dave Hill, senior systems engineer at chipmaker Intel, which has itself bought out a range of low-power processors aimed at the microserver market, based on its Atom System on a Chip (SoC) line.
”Three or four years ago, when we were coming up with our Atom line of SoC products in development, we thought that microservers would be the number one market segment [for Atom SoCs],” he said.
”There are some microserver designs but it's a much smaller market compared to cold storage and networking.”
Hill expects actual shipments of microservers to be one tenth of that predicted by the most optimistic forecasts.
”As of today, it's just really not interesting. Does it go from one to two percent [of shipments] in a year? Maybe. But to 20 percent? There's no way,” he said.
Microservers are low-power servers designed to be packed into dense clusters and spread computationally light workloads between thousands of processor cores. The belief is that dense clusters of low-power servers can handle certain parallelisable tasks more efficiently than smaller numbers of powerful chips. This can deliver better performance per watt and per square foot of datacentre space, important measures for driving down the cost of running a large server estate.
So why hasn't takeup of microservers matched expectations? A key reason is performance, says Intel's Hill. A big market for microservers was expected to be serving static and limited dynamic content for the web. But to date the SoC processors aimed at the microserver market – both x86 and ARM-based – don't offer the performance necessary to run the web stack satisfactorily, he said.
”A lot of the software available on the web today was written between 2005 and 2010. What did they write it on? They wrote it on Xeon [Intel's server processor family] and so the single-threaded performance of Xeon and the capabilities of Xeon is where their software is meant to run. If you try and use a lower unit of compute, it doesn't run very well,” he said.
”Facebook has been very public in saying, "In order for us to even think about deploying my web front end I need two and a half gigahertz parts and somewhere around the 500 SPECint rate.' None of the parts that have come out today have fit in that category.”
The engineering group Linaro is working on improving the performance of the LAMP stack and other crucial web server software software on ARM-based SoCs, however. In summer this year Linaro claimed it was close to having the major enterprise web server applications ready to run on an ARM-based SoC stack.
Another obstacle to microserver takeup that Intel's Hill sees is their inflexibility. HP pitches its Moonshot range of microservers as "application-defined servers", based on SoCs whose circuitry has been pared back to what's necessary to execute well-understood workloads.
The problem with this approach is that workloads aren't static, said Hill.
”What we're hearing from customers is: 'My application changes so fast I can't predict if that microserver is going to meet my needs six months from now, let alone a year from now. If I lock myself to that thing and need more compute than that microserver node handles then where do I go?'.”
These mercurial workloads are better suited to running on virtual machines on a general purpose CPU, said Hill, as the number of VMs can be scaled with demand and the underlying hardware isn't tied to a particular application.
”We see our customers really tearing up entire code bases and rewriting them every couple of years because they've got to handle the scale. There's so much variability in there that locking into a smaller unit of compute is scary to them, because they don't know if it meets their needs in a few years or not.
”It's not that microservers are dead. It just seems very unlikely there will be a tipping point where it is the majority of the market versus the minority,” Hill said.
Despite HP heavily pushing its Moonshot range of microservers, which use both Intel x86 and, more recently ARM-based SoCs, the company has pushed back the point at which it expects the servers will start paying dividends from 2014 to 2015.
Hill said the next generation of its Xeon D SoC, based on its 14nm die shrink of its Haswell architecture, may help stimulate demand for microservers when it becomes available in 2015 as it meets the “minimum [processing power] threshold of a lot of different cloud service providers”.
ARM and its partners claim the first microserver SoCs based on ARMs v8 architecture offer decent ROI compared to Intel Xeon E3 processors. They cite findings that three racks of Intel Xeon E3 1200 v3 web servers with a single rack of the first generation X-Gene processor from Applied Micro, a 64-bit CPU based on ARM's v8 architecture, could result in almost $140,000 savings in running costs over three years. However, this research was commissioned by AppliedMicro and HP, which sells the X-Gene-based ProLiant m400 server cartridge.
Intel in turn claims its microserver-focused Avoton SoC can offer 20 to 30 percent better performance than the X-Gene, while also consuming less power.
While the microserver market may not have taken off, one area where Intel expects its low-power SoCs to thrive is in the networking space.
Intel hopes to serve the low-end of the networking market with its next generation 'Denverton' Atom SoC family, and, maybe at a later point, the medium to high end, served by the likes of the Cisco Catalyst 6000, with the Xeon D and its successors.
Intel's acquisition of the LSI Axxia business will provide the intellectual property for it to add dedicated logic to its processors to accelerate essential networking tasks such as encryption, decryption, and packet-forwarding, said Hill.
”We think the market in terms of silicon is roughly about the size of the server processor market. So it's a large total addressable market and we've got maybe five percent market share there today,” said Hill.
”We definitely want to enter markets where we feel like we could become the number one or number two player over a period of time. The products that we have today get us to the lower end of the networking space,” he said, referencing Intel's Atom family with its sub-20W TDP.
Hill also predicts strong growth for Intel in providing processors to serve the market for cold storage, where data is written once and accessed infrequently. Facebook has a huge demand for cold storage, said Hill, to host the 350 million photos its members upload each day.
Intel and Facebook recently released the board designs for 'Honey Badger' cold storage appliance through the Open Compute project. The appliance, which will be used by Facebook, holds 180TB of data in a 2U form factor and is based on the Intel Atom C2000 processor family.
Intel will face competition in both the networking and cold storage space from ARM and its partners - with network and storage-focused SoCs based on ARM's 64-bit v8 architecture due out from AMD, Applied Micro, and Cavium.
Intel also offers its major customers the option of buying custom chips - standard Intel processors with additional logic to handle their workloads. Hill said Intel is providing more than 35 custom Xeon E5 v3 processors to customers, with the most recent example being a bespoke processor that will underpin a new compute-optimised instance type, known as C4, on Amazon Web Service's EC2 platform.